Abstract

Introduction

Approximately 35 years ago,Hildebrandt (1987) asked the question whether remote sensing in forestry should be considered as a ‘toy or a tool’. In the introductory sentences of his study, the reader discovers that this same question had already been asked another 35 years earlier (in the 1950s), when the potential use of black and white aerial photographs for forestry-related information needs in Central Europe was being considered. While Hildebrandt highlights that in the 1950s remote sensing scientists agreed that remotely sensed data were likely to be useful for forestry, he also notes that they faced widespread lack of interest amongst the majority of their colleagues. Thirty-five years later in 1987, the use of photo-interpretation to support forest management was well established and widely used; however, again only a small group of forest experts expressed interest in the new spaceborne remotely sensed datasets available through programs such as Landsat and SPOT (Wulderet al., 2019).

Now, 35 years afterHildebrandt (1987), remote sensing has become a much more common technology in forestry. Global forest cover and change products (Hansenet al., 2013), and forest disturbance products covering large portions of continents and time-periods spanning several decades (Whiteet al., 2017b) are available. Furthermore, maps and data products representing forest types or tree species information that are derived from time series of freely available passive optical satellite data are being generated for regions of varying sizes (Grabskaet al., 2020;Hermosillaet al., 2022). These types of products are in turn used in some jurisdictions to enhance estimates of official national sampling programs, such as national forest inventories, and to provide spatially explicit estimates of forest attributes (Tomppoet al., 2008;McRobertset al., 2010).

In some regions, airborne light detection and ranging (LiDAR) as well as digital aerial photogrammetry (DAP) have emerged as indispensable technologies for mapping forest structure and as key auxiliary information to improve estimates of forest attributes provided by forest inventories (Næsset, 2002;Nelson, 2013;Whiteet al., 2013a). Fine spatial resolution optical satellites, some of them operating in constellations of multiple micro-satellites, provide spatially detailed data with a high temporal revisit that can increasingly match the information requirements associated with emergency responses or disturbance events (Dalponteet al., 2020). Besides these traditional applications of remote sensing, new technologies and platforms, including drones (Goodbodyet al., 2017) and proximate sensing such as terrestrial and mobile laser scanning (Bauwenset al., 2016;Lianget al., 2016) and sensors associated with smart devices (Siipilehtoet al., 2016) have evolved rapidly. These technologies foster the development of new applications and innovative approaches suitable for assessing a diverse range of forest attributes in the field.

Rapid technological developments in the scientific field of remote sensing have been mirrored by a marked increase in scientific research and publishing; however, the pace and level of uptake of remote sensing technologies into operational forest inventory and monitoring programs varies notably by geographic region (Whiteet al., 2016;Kangaset al., 2018;Maltamoet al., 2021;Neshaet al., 2021;Coopset al., 2021a). Variability in the uptake of remote sensing technologies relates to several factors including: forest planning and management traditions in the country; complexity of forest structure and composition; availability of remote sensing technologies, expertise, data and derived products; information needs of forest policy; forest land ownership structures; linkages between research and industry; and, public subsidies and regulations, among others (Wulderet al., 2007) (Figure 1).

Figure 1

Overview of factors contributing to user uptake of remote sensing technology in forestry.

Open in new tabDownload slide

In this communication, we highlight some key challenges that remote sensing research will need to address to further increase the acceptance, suitability and integration of remote sensing technologies, products and data into operational forest inventory and monitoring programs. We first address general challenges related to the uptake of remotely sensed data in forestry, then discuss current technological and methodical challenges and opportunities with respect to four specific topics:

1. National forest inventories (with the main aim to improve estimates of forest attributes on a national and regional scale and satisfy reporting obligations).

2. Management-level forest inventories (with the main aim to regularly produce spatially continuous maps with associated estimates of important forest attributes at a fine spatial resolution in order to inform management decisions).

3. Plot-level measurements (with the main aim to either improve the efficiency of field measurements or expand the attributes to be measured on the ground) and

4. Temporally continuous forest monitoring (with the main aim to provide spatially explicit information about the state of, and changes to, forests).

Finally, we provide rationale and guidance regarding the importance of rigorous validation for map products derived from remotely sensed data (Breidenbachet al., 2022), with a special focus on issues related to estimating uncertainties.

We conclude by summarizing key challenges and opportunities that emerged from our synthesis and that are relevant to all of the aforementioned topics. Our aim is not to both pose and offer solutions to address all possible challenges, but to draw attention to those topics, which if addressed or advanced, will serve to enhance the utility and uptake of remote sensing in forestry.

User uptake of remotely sensed data in forestry: challenges and opportunities

As a discipline, remote sensing science is often focused on a particular data source or methodological development rather than on the value or utility of the derived information, the latter of which is more crucial with respect to user uptake. Furthermore, uptake of new and emerging technologies is often restricted by costs and technical capacity (de Gouwet al., 2020). Uptake is often most successful when there is a strong understanding of the information needs associated with a specific application, as well as the remote sensing technology most appropriate to address those needs (Whiteet al., 2016). But even then, uptake will rarely happen organically unless there is a clear (often economic) incentive and/or the value of the information is well understood and communicated to potential users. Realistically, even applied research will often not reach those whom could benefit the most if the results are only disseminated via scientific communication channels. Explicit knowledge transfer activities (Whiteet al., 2013a,2017a;Næsset, 2014) or collaborations are essential to uptake, but often require considerable time and effort and are seldom considered as required outcomes of research proposals.

Research and initiatives, which aim at improving the uptake of existing remote sensing technology into operational use, have to consider the local context and existing work-flows in the forest industry and administration of forest-relevant programs. For example, in Germany, forest management plans are based on detailed field inventories that are conducted every 10 years. Over the last few decades, this information has been perceived as a sufficient database for planning forest operations. However, today foresters are increasingly concerned about changing dynamics in German forests as a consequence of extreme weather and related forest damage (Popkin, 2021). The rapid pace of change may mean that information provided by the 10-year forest plans is quickly outdated and no longer accurately characterizes the reality on the ground. Indeed, accessing timely information to update forest plans and/or shorten the inventory cycle is increasingly of interest across many jurisdictions.

The degree to which remotely sensed data can provide the necessary information to address these information needs has been the subject of decades of research that have evolved with rapidly changing technologies. Suggestions for operational monitoring frameworks to regularly update existing inventory information have recently been made (Coopset al., 2022) and research into methods that may enable this, such as data assimilation approaches, is ongoing (Nyströmet al., 2015;Lindgrenet al., 2021). However, this research must always be considered within the local forest management context. For example, forest management across large parts of Europe is increasingly focusing on uneven-aged forest management strategies in mixed forest stands with an additional focus on single-tree precision forestry (individual high-quality timber trees are the main productive aim). In such a setting, information demands of practitioners may increasingly deviate from information delivered by typical forest inventories and refinements of both traditional field-based inventory approaches as well as the typical area-based remote sensing approaches, which deliver information on a sub-stand resolution, but not at individual-tree level, may become necessary.

In contrast, forest management in countries that follow an extensive forest management approach faces other challenges (Wulderet al., 2007). For example, Canada has a vast forest area, the density of field plot networks is sparse, and the area for which forest information is required is markedly larger and less accessible. Furthermore, given jurisdictional stewardship of forests in Canada, each jurisdiction designs and implements its own forest inventory program using temporary and permanent inventory plots identified and inventoried in the field and with remotely sensed data (mostly aerial photographs) (Leckie and Gillis, 1995). To overcome differences in jurisdictional programs and to enable robust forest statistics at the national scale, a sample-based national forest inventory (NFI) is implemented in addition to jurisdiction-level forest inventory programs (Stinsonet al., 2016).

The question of how or when to transition towards the operational use of new remotely sensed data and technologies in forest inventory programs is a significant challenge faced by most countries and jurisdictions. Regional and national inventory programs have typically evolved and grown over longer time periods and data collection, attribution, management and distribution activities have budgets, staff, infrastructure and a dedicated user base. Existing programs often continue to rely on aerial images, while newer but established remote sensing technologies, such as airborne laser scanning (ALS) or data from earth observation satellites are gaining importance in some parts of the world (Whiteet al., 2013a;Næsset, 2014;Kangaset al., 2018;Breidenbachet al., 2021).

One main reason for the comparably slow speed of this process may be a perception that the use of (other) remotely sensed data will necessitate completely different workflows or revamping of existing programs. However, the reality is that new technologies and data streams may often easily be added to existing procedures and field sampling data streams. The addition of remotely sensed data often provides an immediate and positive effect in terms of more precise and more locally relevant estimates, in addition to enabling the generation of spatially explicit map outputs (Franciniet al., 2021). A phased approach to integrating new data streams into forest inventory programs may be necessary to enable uptake and acceptance within programs and among stakeholders (Tomppoet al., 2008;Næsset, 2014). While such a phased approach reduces risk, simultaneously generating the required information for meeting monitoring and reporting needs for multiple independent data streams is challenging. Depending on the initial conditions of the inventory program, a complete switch to a more costly but more informative data source (typically ALS) is possible, as illustrated in the examples of the Nordic countries Finland, Norway and Sweden. In these countries, the uptake of remotely sensed data has been notably further advanced relative to other jurisdictions in central Europe and North America. In the following, we will explore the reasons for this in more detail.

National case studies

Systematically assessing and understanding user requirements and finding ways to integrate new remote sensing technologies into existing workflows in order to minimize the adaptation needs required of data users has a generally positive influence on uptake. Support via adequate funding schemes and incentives is also essential. A common denominator for the Nordic countries has been the clear demonstration in value to the end user through collaborations between researchers and industry. Presently, success in the uptake of technology and procedures in the Nordic countries is driving investments in repeat ALS data acquisitions as part of national programs. Related success stories and the means to achieve these successes are often not communicated to a wider international audience. While it is unlikely that solutions identified in a national context are directly transferable to other countries (due to differences in forest management procedures and objectives, centralized data purchases, ownership structures, forest types and structure, etc.), communicating such outcomes and lessons learned, along with the corresponding methodological approaches in international journals and other media may spur similar activities in other parts of the world. From this perspective, awareness coupled with the capacity to adapt approaches appropriately to a given forest management context will be key for uptake. To address this uptake aim, we discuss these points in more detail for five countries each with a strong, yet contextually different, forestry sector.

A variety of factors have contributed to the notable uptake of remote sensing in Nordic countries. On the one hand, existing forest management strategies and land ownership contexts, coupled with the structure, composition and size of the forests as well as a traditionally strong forest industry with an appetite for innovation may have benefited rapid uptake of ALS technology. However, the Nordic countries were also pioneers in establishing a collaborative and highly integrative approach to methods development and implementation among remote sensing researchers, practitioners and forest inventory service providers at the time when ALS technology was nascent and just emerging commercially (Næsset, 2014). There are, however, important differences among the approaches pursued in Finland, Norway and Sweden, and an even larger contrast to other nations such as Canada and Germany (Table 1).

Finland has a strong tradition of nationally coordinated efforts with respect to methods development, methods approval and financial support from the government to stimulate and coordinate adoption of new methods and ways of organizing forest management inventories. These efforts are complemented with subsidies directly to forest owners whenever a forest plan is undertaken. This context, coupled with a long history of research in the application of remote sensing in forestry (Hyyppäet al., 2000) has resulted in a positive environment for innovation and uptake of emerging technologies in the forest sector. The shift to the use of ALS as a primary source of data for the management inventories over a period of around five years (2005–2010) was guided and supported by a massive investment in research and development (Maltamo and Packalen, 2014).

Finland is characterized by a blend of large industrial forests and numerous small private forest holdings in the south (the latter form of ownership representing an average forest area on the order of 50 ha or less) and state forests in the north. The Finnish model of promoting adoption of the most recent remote sensing methods and technologies has applied to all types of ownership. Over time, a diversity of private service providers has emerged, offering stronger competition and striving for better and more economically viable ways of using the technology to produce information for forest management purposes (Maltamoet al., 2021). However, the government still plays a substantial role, for example by offering free and publicly available nation-wide ALS datasets for forestry purposes as well as stand-based prediction maps of forest attributes constructed from ALS data and local sample plot surveys. Finland is now conducting a second full ALS survey of the entire national territory with forestry needs as an important motivating factor.

Norway also has a long history of pioneering research on the use of remote sensing in forestry (Næsset, 1997a,1997b). The country has a decentralized structure, which complicates the implementation of incentives for uptake of remote sensing into operational forestry. As a consequence, the government focused on offering subsidies directly to the forest owners for management inventories and management planning under the condition that there is a collaborative effort among the forest owners over large tracts of land. Despite a fragmented ownership structure with almost 80 per cent of the productive forest area being owned by private individuals and with an average property size of around 50 ha, the way the subsidies are distributed has provided a strong incentive to strive for large-area applications to provide for economies of scale, thereby creating an ideal environment for the use of ALS technology. At the same time, an active private sector on the service provider side has stimulated competition and innovation in the development of more effective methods. So, despite different ways of organizing the service sector in Finland and Norway, the strong governmental influence and oversight on the sector in Finland and the public subsidies used to stimulate collaboration and economies of scale in Norway have yielded almost the same outcomes in terms of innovation and the widespread uptake of emerging technologies.

In this context, Sweden represents an interesting third example, wherein mapping for forest management planning traditionally has been done at the individual property level and financed entirely by the forest owners themselves. The governmental role has been mainly limited to the operation of the national air photo program. When data from the first nation-wide ALS acquisition became available around 2010, the data started to be used operationally for large areas by industrial forest companies in the northern Sweden. The forest sector in Sweden has had strong and active support from the research community with similar capacity and competence as in Finland and Norway (Jaakkolaet al., 1988). The Swedish Forest Agency and the Swedish University of Agricultural sciences were tasked by the government to create a detailed spatially explicit forest database for all of Sweden, using area-based models generated from ALS data and national forest inventory plots (Nilssonet al., 2017). The resulting database was made open access, and has become a valuable source of information for keeping the forest management plans for all types of ownerships up to date, including the numerous smaller privately owned forests. In order to keep the forest database up to date and share the cost of the ALS data, a second national ALS acquisition co-financed between the government and the forest sector was initiated in 2018. In Sweden, these national, detailed open-access information products are expected to be maintained in the future, with the ambition of repeating national ALS surveys at regular time intervals. Such national information products are currently also under consideration for use in operational forest management in Norway.

Uptake of ALS technology in forest inventories in Canada has varied regionally across the country. In Canada 89 per cent of the land is publicly owned and each jurisdiction (i.e. province or territory) has defined stewardship responsibilities for natural resources and accordingly maintains its own unique forest inventory program. In Canada, governments own the land in trust of the public with the lands actually being managed by industry (via tenure agreements) or other arrangements such as by communities (Stinsonet al., 2019). Regardless of who is actively managing a given forest, defined practices must be followed and regulations to be met are in place. Governments are typically interested in strategic inventories to assist in overall management and sustainability of the forest resource, including but not limited to allocations of tenure and allowable harvesting levels. These strategic jurisdictional inventories have been primarily based on the interpretation of aerial photography and implemented at a cadence and level of detail intended to be informative at a strategic level (Leckie and Gillis, 1995). While of value for understanding trends in forest resources and supporting timber allocations to the forest industry, these inventories were less informative for supporting more detailed, operational forest management. The uptake of ALS technology to support these information needs has expanded rapidly in the past decade and several jurisdictions in Canada have transformed—or are currently in the process of transforming—their inventory approach to incorporate ALS data. It is worth noting that areas under consideration for management are large. Forest-dominated ecosystems in Canada occupy ~650 MHa, with >350 MHa of forest cover (Wulderet al., 2020) and an overall managed forest land base of 225 MHa reported for 2020 (https://natural-resources.canada.ca/sites/nrcan/files/forest/sof2022/SoF_Annual2022_EN_access.pdf,2022). Individual allocated tenures are often >1Mha in size, with some jurisdictions (e.g. Quebec, Alberta, Ontario) having collected millions of hectares of ALS data.

This transformation of inventory practices to include ALS data has been enabled in part by concerted government investments in collaborative research efforts among academics, government and industry stakeholders (Coopset al., 2021b), as well as through extensive and coordinated technology transfer activities, both nationally (Whiteet al., 2013a,2017a;D’Eon and MacAfee, 2016) and within jurisdictions. Variability in uptake among jurisdictions is, however, pronounced, with early adopters having advanced and refined their approaches and now considering options for their next inventory cycle, whereas other jurisdictions have pursued a more phased approach or are considering other remote sensing technologies. This patchwork of uptake is likewise present among forest industry stakeholders, with increasing capacity among service providers. Key to understanding this variability across the country is the nature and value of the forest resource, and the legacy of established forest inventory programs. Given, the vast extent of forested areas in Canada, non-Federal management responsibilities over natural resources, and the subsequent disparate, jurisdictional approaches to forest inventory and monitoring across the country, a diversity of solutions can be expected to continue to exist in parallel.

In Germany, the uptake of remote sensing technology into forestry has thus far been limited. One important reason for this is the decentralized structure of the forestry sector in Germany in combination with limited government incentives to promote the use of remote sensing technology in operational forestry. The decentralized structure is particularly problematic in privately owned forests, wherein owners of small to intermediate sized forest properties organize themselves in forest owner associations. These associations then hire a forester or a company who supports them with planning and management activities. There is a traditional preference of these associations to work with local companies or individuals, which often lack the capacity to provide remote sensing-based solutions in a cost-effective manner. In addition, most of the high-quality remotely sensed data (e.g. aerial photographs and ALS) collected by the state are not free and open. Although some federal states have recently made their ALS data and aerial photographs publicly available, there is still a lot of progress to be made. Furthermore, whereas research on forest remote sensing applications is well funded, there are limited financial incentives that promote the use of remote sensing in the forestry sector. Within publicly owned forests, the ownership structure is less problematic; however, concerns about inducing changes to established planning workflows strongly limit innovation, despite an increased awareness of the need for new technologies to update forest inventory information in a timely and spatially explicit manner.

Challenges to adoption arising from technical aspects

While nation-specific examples illustrate some of the factors that influence uptake of remote sensing technologies, there are at least two additional important technical considerations: data quality and spatial resolution. Data quality is often raised as a barrier to uptake of remotely sensed data products. While the term data quality is somewhat vague, it often refers to the accuracy and precision with which remotely sensed data can characterize estimates of the forest attribute of interest. In some cases, remotely sensed data may be perceived as an insufficiently accurate data source because the reported error exceeds a certain threshold or tolerance. At the same time, the error associated with established datasets on which forest planning has been based for several decades may be unknown, yet is accepted. In some cases, these data may be perceived as reliable because they were collected by an expert in the field and not necessarily because their reliability has been quantified.

Indeed, several studies have demonstrated the uncertainty associated with conventional forms of forest measurement (Eid and Næsset, 1998;Luomaet al., 2017;Wanget al., 2019;Jurjevićet al., 2020) and air photo interpretation (Eid and Næsset, 1998;Tompalskiet al., 2021). Other studies have characterized data quality in terms of the value of the information to support decision-making and forest planning (Kangas, 2010;Bergsenget al., 2015). A study byUlvdalet al. (2023) shows that for example in Sweden, many forest planning experts are aware of high uncertainties in their base data and acknowledge the value of wall-to-wall maps provided by ALS data. In this context, more studies that quantitatively and qualitatively benchmark the information content of remote sensing-based information products with the information that is currently used by forest planning experts could be valuable.

Finally, another potential barrier to the uptake of remotely sensed data are differences in the spatial scale or the exact information content of operationally used inventory data by forest planners and the corresponding remote sensing products. A typical example is the difference between the fixed grain size at which many ALS-based forest inventory products are provided (e.g. 20 × 20 m grids) and the typical operational scale of forest management, which is often conducted for forest stands (of varied size). Planning procedures and software may for many applications require the continued use of stand-level information and thereby tools that readily translate grid-based products to stand-wise estimates may be important for uptake. At the same time, there seems to be a trend towards more locally adapted and diversified management of the forests requiring very fine spatial resolution of the data as well. The further development of these tools allowing a dynamic scale of the data should be pursued in collaborative efforts and may require some further developments in the field of spatial statistics.

Technical challenges and potential solutions of remote sensing in forest inventories

The use of remote sensing technologies in forest inventories is highly diverse and can occur at a variety of spatial and temporal scales. In the following, we discuss the technical challenges and potential solutions associated with the use of remote sensing in forest inventories in four contexts related to (1) national forest inventories, (2) management-level forest inventories, (3) plot-level measurements and (4) temporally continuous forest monitoring. This structure should not be seen as specific to spatial scale, but rather a structure that reflects the different information needs that prevail in these different inventory contexts (Table 2) as well as differences in the targeted end users. Due to the plethora of forest inventory programs around the globe and the large diversity of administrative and environmental settings in which they are conducted, some of the main aims as defined inTable 2, may apply for multiple types of inventories. In each of these contexts, we introduce and discuss some key remote sensing technologies and sensor types that are most commonly applied. However, it must be noted that the use of these technologies is often not limited to that particular inventory context.

National forest inventories

The main objective of national forest inventories (NFI) is to deliver accurate and up-to-date estimates of a plethora of forest inventory attributes including forest area, wood volume, biomass, growth, forest health, etc. Typically, these estimates are reported at the national level, as well as for smaller spatial administrative units, and may also be further broken down according to other categories such as ownership classes, forest types or species (McRoberts and Tomppo, 2007). NFIs usually cover extensive areas and are conducted on a regular cycle (i.e. at time-intervals typically up to or less than 10 years;Gschwantneret al., 2022), often with continuous measurements made according to panel systems. Thus, there is great interest in increasing the efficiency of NFIs, with remotely sensed data playing a key role (Kangaset al., 2018). Remote sensing is already being used in many NFIs around the world. In a survey of 45 countries representing 65 per cent of the world's forest area,Barrettet al. (2016) reported that 71 per cent of these countries indicated that their NFI programs were ‘dependent or partially dependent on earth observation data’.

The integration of remotely sensed data into NFI procedures can be associated with at least five tasks: (1) pre-checking of field plot locations (e.g. using aerial photographs to check whether the field plot is forested and how it can be reached); (2) replacing field-plot information through visual interpretation of aerial images (e.g. forest type, density, height, forest condition) for those plots, which may be inaccessible or in the context of a double-sampling procedure; (3) improving field-plot-based estimates by using remotely sensed data as auxiliary information (e.g. via post-stratification) (McRobertset al., 2006); (4) obtaining map products of forest inventory attributes by combining NFI data with remotely sensed data (Tomppoet al., 2008;Breidenbachet al., 2021;Hermosillaet al., 2022) and finally, (5) obtaining more local estimates (via map products), which cannot be obtained by NFI field plots alone due to insufficient sample sizes (McRobertset al., 2006;Masseyet al., 2014;Katila and Heikkinen, 2020).

As mentioned inRätyet al. (2018), over the last two decades many studies have shown that using remotely sensed data as auxiliary information to improve NFI estimates is effective (tasks 2 and 3 above) whether it be via double-sampling (Masseyet al., 2014), post-stratification (McRobertset al., 2012;Tiptonet al., 2013;Myllymäkiet al., 2017) or model-assisted estimation (Opsomeret al., 2007;McRobertset al., 2013;Saarelaet al., 2015;Kangaset al., 2016). This has been demonstrated with multispectral satellite data (mostly Landsat and SPOT 1–5) (Barrettet al., 2016) in selected regions of the world (Haakanaet al., 2019) as well as for smaller extents with ALS data (Gregoireet al., 2011;Magnussenet al., 2018).

Similarly, wall-to-wall maps have been successfully produced by combining NFI and remotely sensed data (task 4) in numerous studies (Nilssonet al., 2017;Waseret al., 2017;McInerneyet al., 2018;Urbazaevet al., 2018;Chiriciet al., 2020;Hawryłoet al., 2020;Langet al., 2020;Hauglinet al., 2021). While the accuracy of those maps depends on the forest attribute and the geographic region, they have been produced and used routinely, for example, in Finland and Sweden to calibrate forest management plans, for district-level planning purposes, and to support law enforcement (Tomppoet al., 2008).

Using freely available earth observation data as auxiliary information and for mapping forest inventory attributes is of particular interest, as these approaches provide added value without changing any of the main procedures or frame conditions of the NFI (e.g. sampling design, sample size, etc.). Given increasing computational processing power and availability of open-access medium to high resolution satellite data, the main technical limitations related to NFI tasks 3 and 4 have been overcome and there appears to be no clear technical or financial limitation to the operational uptake of these procedures into NFI practices around the world.

According toBarrettet al. (2016), 12 out of the 45 surveyed countries were already using remotely sensed data in double- or multi-phase sampling for stratification, and 30 out of the 45 countries were combining NFI data with remotely sensed data to create maps. One reason why the number of countries applying remotely sensed data for post-stratification is still comparably low may be that the verification of the approaches in some forest types is still pending; most of the research to date has focused on boreal and temperate forests of the Nordic countries and North America. Hence, it may be worthwhile to continue earlier efforts to raise awareness and convince the responsible professionals in charge of NFIs around the world to verify and then integrate the available solutions in their established procedures. Providing the requisite tools via open-source software solutions along with the necessary technical support and knowledge exchange may be one step towards accomplishing this. As shown in very recent discussions about a new pan-European forest monitoring system, the demand for more precise NFIs clearly exists and an improved integration of NFI field plots and remotely sensed data to improve precision and reduce costs is seen as a promising way forward.

Furthermore, several studies demonstrated benefits of integrating remotely sensed data into NFI programs on top of the already mentioned approaches. For example, integration of multi-temporal or time series of satellite data may improve mapping of forest extent (Wulderet al., 2020), discrimination of forest types (Wulderet al., 2018) and contain information related to the age of forest stands (Tomppoet al., 2008;Helmeret al., 2009). Moreover, the use of maps constructed from previous NFI surveys as auxiliary information for post-stratification (Haakanaet al., 2019) may be worthwhile to pursue to achieve further improvements. An additional field of interest for future studies is the application of remotely sensed data as auxiliary information for modelling changes between repeated NFI surveys. For example,Pulitiet al. (2021) demonstrated that Norwegian NFI-based biomass change estimates notably benefited from integrating multispectral satellite data as auxiliary information. Similar results were reported byBreidenbachet al. (2021) who used multi-temporal NFI data in Norway, Sweden, Denmark and Latvia, Landsat-based forest-loss data, and wall-to-wall ALS data to improve estimates of changes in carbon stock.

Tasks 1 and 2, pre-checking of field plot locations and replacing field-plot information through visual interpretation of aerial images, both refer to the direct extraction of information from remote sensing data at field plot locations and mostly rely on fine spatial resolution imagery and particularly aerial imagery. There is a long tradition of supporting both national and management-level forest inventories with information obtained via visual interpretation of aerial imagery (Barrettet al., 2016). The reasons for this are numerous (seeKoch, 2013) but particularly include the information content and spatial detail offered by the images, and—at least in many European countries and the United States—the easy availability of imagery that is routinely acquired to support on-going base-mapping or other programs.

Besides the already mentioned tasks related to pre-checking NFI plots, another potential application of aerial imagery is the mapping of trees outside forests. The latter have been surveyed in the context of the NFIs of some countries and may be integrated in more NFIs in the future (Malkoçet al., 2021). Recent advances in the field of deep learning and in particularly convolutional neural networks (CNN)-based algorithms may enable automated workflows to reliably detect individual trees from aerial photographs in the near future (Weinsteinet al., 2020). The main obstacle for applying deep learning algorithms is their dependence upon very large training datasets. However, given the fact that vast amounts of digital aerial images have already been interpreted in the context of NFIs in the past, these training datasets may already be available. Hence exploring the suitability of deep learning algorithms for characterizing trees outside forests may be an interesting field for future studies. The also common tasks of detailed stand delineation and photo interpretation from aerial imagery are more commonly associated with management-level inventories and will be discussed below.

Management-level forest inventories

NFIs and management-level forest inventories vary in both their spatial and temporal characteristics (Table 2). Whereas NFIs are typically sample-based and designed to monitor changes in forest attributes over time, management-level forest inventories are typically wall-to-wall, spatially explicit products that can support management-level decision-making. While consistency of sampling design and measurement over time is key for NFIs, management-level inventories characterize current forest conditions in order to support operations and management.

Due to the significant progress made in the development and uptake of remote sensing workflows for management-level inventories, most challenges currently relate either to the question of how to further optimize existing workflows or to identify additional forest attributes beyond those commonly characterized in the literature (e.g. volume, basal area, height;Whiteet al., 2013a). For example, distinguishing vertical layers in the canopy could be accomplished using existing ALS technologies (Valbuenaet al., 2016). Many of the challenges related to further optimization are difficult to address as they often require additional data acquisition. Particularly any optimizations related to data acquisition settings face the problem that repeat acquisitions (for example of ALS or VHSR data) can be costly and may be challenging to finance solely for the purposes of forest inventory alone. Given the value of the data for a range of natural resource management and public safety applications, consortia of stakeholders provide a mechanism for sharing costs. Likewise, progress in the development of realistic data acquisition simulators (Winiwarteret al., 2022;Schaeferet al., 2023) as well as additional efforts with respect to data-sharing initiatives (Pirotti, 2019) may be further ways to overcome these challenges in the future. Given these limitations in data availability, it is important to focus research efforts on remote sensing technologies that have the highest potential to support and inform forest management in the operational planning stage in the near future.

Aerial photography has traditionally been the most commonly used form of remotely sensed data in forest inventories. While manual air photo interpretation is operational in many forest inventories (both NFIs and management-level inventories), it is also a time-consuming, subjective process, for which it is increasingly difficult to find qualified interpreters familiar with the management area of interest. Over the years, considerable research efforts have been targeted at automating air photo interpretation, including both stand delineation (Wulderet al., 2011) and attribution (Falkowskiet al., 2009;Morgan and Gergel, 2013); however, this has proven challenging (Hill and Leckie, 1999).

When attempting to automate the information extraction from aerial images (and other fine spatial resolution remotely sensed data) over the last few decades, most of the examined procedures to extract information were pixel based. These procedures often failed to reach the same perceived performance as a trained photo-interpreter. However, it is important to differentiate between perceived performance and true performance, as forest attributes resulting from manual air photo interpretation were typically not validated but simply assumed to be of sufficient quality. The recent study ofTompalskiet al. (2021) demonstrated that photo-interpreted data may contain notable levels of variation related to individual interpreter and the photo-interpretation procedure applied. This coincides well with findings reported in the literature from the 1970s to the 1990s, albeit based on stereo-interpretation of analogue imagery and not modern digital imagery (Næsset, 1991a).

Clearly, variation among interpreters as an expression of uncertainty of the interpretations can have a substantial effect on uncertainty of attributes estimated from the interpretations, such as land cover and other area attributes (McRobertset al., 2018;Stehmanet al., 2022). Furthermore, in manual photo interpretation, some forest attributes that are challenging to directly interpret from imagery may be estimated with the help of look-up-tables based on attributes that are relatively easier to interpret (i.e. volume from interpretations of species composition, crown closure and age;Næsset, 1991b;Stinsonet al., 2016). Similar look-up table approaches could also be applied to forest attributes derived with automated approaches; however, this has thus far received less attention from the scientific community.

Still, automated approaches have also been suffering from technical challenges in extracting the desired forest attributes since most of the information in the images that are used by photo-interpreters relate to differences in textural patterns (for example related to tree species). Even with the additional use of texture metrics and object-based approaches, successful studies were often limited to comparably small datasets (Tuominen and Pekkarinen, 2005). Furthermore, algorithms were often not readily transferable, due to the often high variability in illumination conditions (Tuominen and Pekkarinen, 2004) and in the case of fine spatial resolution satellite imagery, complexities of sun-surface-sensor geometry (Wulderet al., 2008a). Hence, despite the extensive use of aerial images in forest inventories, automation levels are often low. The advent of new algorithms in the field of deep learning (Kattenbornet al., 2021) may contribute towards increased automation in the future, particularly for relatively simple tasks, such as verifying whether a field plot is located in forest or the assignment of plots to forest types.

Besides traditional aerial photography-based approaches,Whiteet al. (2016) posited that there were four remote sensing technologies that were likely to have the greatest influence on management-level forest inventories in the near term: ALS, DAP, high spatial resolution (HSR)/very high spatial resolution (VHSR) satellite optical imagery and terrestrial laser scanning (TLS). In particular, continued operationalization of ALS data into forest inventory workflows for large areas and the combination of ALS and DAP data as time series observations for characterizing forest growth (and site productivity) were highlighted as significant opportunities for forest practitioners.

The use of ALS data in an area-based approach (ABA) for forest inventories is already considered operational (Whiteet al., 2013a;Næsset, 2014). However, on-going research related to the ABA has sought to keep pace with the evolution and advancement of laser-scanning sensors in order to further refine methodological approaches. One key objective in this context is to quantify the impacts of acquisition conditions and a range of environmental parameters on the derived forest inventory information, particularly over large areas. Such considerations are important for operational forest inventory programs, which are often implemented over a broader range of forest conditions and larger areas than is typical of most research studies. This gap between research-level datasets and operational scale implementation is a key concern that needs to be addressed via targeted research. Otherwise, important issues in workflows developed based on research-level datasets may only be identified during operational implementation, leading to dissatisfaction on the side of forest practitioners and reinforcing barriers to uptake of these remote sensing technologies.

Recent ABA research has focused on the use of multispectral ALS for species-specific forest inventories (Kukkonenet al., 2019;Yanet al., 2020;Kotivuoriet al., 2021) and single photon lidar, which can provide acquisition advantages for large areas (Wästlundet al., 2018;Yuet al., 2020;Queinnecet al., 2021;Whiteet al., 2021). Methodological refinements to the ABA have explored a range of topics, from the impacts of mismatched plot and gridcell sizes and shapes (Packalenet al., 2019;Packalenet al., 2022) and plot location errors (Gobakken and Næsset, 2009), to different strategies for selecting field training plots (Maltamoet al., 2011) or alternative modelling approaches (Næssetet al., 2005;Aulló-Maestroet al., 2021;Cosenzaet al., 2021;Yanget al., 2021a,2021b). Issues relevant for large-area implementations have included exploring the influence of terrain slope and aspect on the empirical relationship between ALS metrics and field-measured forest inventory attributes (Næsset, 2004;Ørkaet al., 2018), the impact of the increasingly large acquisition scan angles with which ALS data are being acquired operationally on ABA outcomes (Keränenet al., 2016;van Lieret al., 2022), and the transferability of ABA models, both spatially (Karjalainenet al., 2018;Tompalskiet al., 2019) and temporally (Feketyet al., 2015;Magnussenet al., 2015;de Lera Garridoet al., 2020). These studies highlight that the need to examine the suitability of the ABA approach under a wide range of acquisition and environmental conditions is a priority within the research community; however, the issue of spatially limited datasets is still valid in in several of the mentioned studies.

DAP, whereby 3D point clouds can be generated from digital imagery acquired with sufficient image overlap, has emerged as a complementary remotely sensed data source to that of ALS data (Goodbodyet al., 2019). However, DAP differs from ALS data in two important ways. First and foremost, DAP only characterizes the outer canopy envelope and cannot provide the same characterization of the vertical distribution of vegetation within the canopy as ALS. Second, the use of these data to derive accurate canopy height measures requires a pre-existing, high-quality digital elevation model, which underneath forest canopy is typically only available from ALS data (Whiteet al., 2013b). In the scientific literature, DAP have most commonly been applied in the context of management-level inventories but may also serve as auxiliary information in NFIs, particularly for smaller countries with nation-wide surveys for digital aerial photographs.

Research on the application of DAP data has expanded markedly and has focused on comparing the performance of DAP data relative to ALS for area-based forest inventories across a range of forest environments (Goodbodyet al., 2019) and different camera systems (Iqbalet al., 2018;Strunket al., 2020). All of the studies examined inGoodbodyet al. (2019), reported that ALS data provided superior results relative to those achieved using DAP data and more recent comparisons have corroborated these results (Noordermeeret al., 2019). While the majority of studies have documented performance at the plot level or at the stand level, specific challenges associated with the application of DAP data over large spatial extents have been examined. For example,Rahlfet al. (2017) examined the influence of varying image illumination conditions on ABA outcomes over large areas using DAP data—a common issue for image datasets acquired over large areas and with several successive flights. The authors found that accounting for varying illumination had a minimal effect on area-based outcomes, reducing relative root mean square error ( RMSE) by <2 per cent and concurring with earlier results reported inWhiteet al. (2015). For large areas,Strunket al. (2019) reported notable efficiency increases by integrating DAP data for estimating forest volume in Washington State, USA, despite the use of field plots with low positional accuracy and comparably coarse digital terrain models (<10 m) for height normalization.

The use of DAP-derived canopy height models as a mechanism to spatially transfer ABA models was also explored (Stepperet al., 2017). Benchmarking of DAP acquisition parameters and quantification of their impacts on area-based outcomes, particularly over large areas, remains an outstanding research challenge across a range of forest environments (Goodbodyet al., 2021). Moreover, most image-matching software currently in use were not designed or optimized for forest environments and applications (cf.Baltsaviaset al., 2008), indicating opportunities for the development of software solutions that specifically target forests. Finally, a lack of standardized workflows may lead to a varied quality in derived outputs, and best practice recommendations, informed by existing scientific literature, are needed to ensure greater consistency in DAP point clouds and derivatives (Iglhautet al., 2019). In summary, as in the case of ALS data, there is still a need for studies that specifically work with spatially extensive DAP datasets that are representative for an operational case (data collected with various camera systems, under differing illumination conditions and varied environmental situations).

A further data source that may be particularly interesting for forest management inventories in areas where neither ALS or DAP data are available are HSR (<5 m spatial resolution) and VHSR (<1 m spatial resolution data) satellite data. HSR and VHSR optical satellite imagery has received comparatively less research interest in management-level forest inventories in recent years. Previous studies have explored the use of VHSR data to estimate a range of forest attributes (Moraet al., 2010a,2010b;Leboeufet al., 2012) and more recently, studies have used stereo VHSR (Immitzeret al., 2016;Fassnachtet al., 2017;Wallneret al., 2022), reporting results for forest attribute estimation that have larger errors than what can be attained using ALS or DAP. However, studies that directly compare ALS, DAP, VHR and VHRS are rare. The capacity to acquire cloud-free observations within the required time frame and according to desired specifications (e.g. view angles <15° from nadir) remains an operational challenge for large-area forest inventory applications with VHSR data (Falkowskiet al., 2009). Further challenges arise from the small image footprints (which increases processing overhead) and inconsistent sun-surface-sensor geometry (Wulderet al., 2008a). Hence, VHSR data may be more suitable for sample-based inventories or for targeted updates of forest inventory information (Hilkeret al., 2008) after, for example, the detection of disturbances in a monitoring system based on moderate resolution data (Boltonet al., 2018;Coopset al., 2022), than for operational forest inventories across large areas.

With repeated ALS and DAP acquisitions over the same area, the potential to derive additional attributes has arisen. For example, site productivity has been identified as an important information need for which remotely sensed data may provide important inputs (Wulderet al., 2010). Site productivity is perhaps the single attribute that has been most sub-optimally determined by traditional methods in economic terms (Eid, 2000). Starting in 2020, site productivity became part of the standard information product in ALS-supported inventories in Norway (Noordermeeret al., 2020). Site productivity can be derived quite reliably based on repeated ALS surveys or combinations of ALS and DAP data (Noordermeeret al., 2021). Deriving site productivity from remote sensing data is of particular interest, as existing productivity indicators (often developed several decades ago and from very limited samples) do not fully represent current growing conditions due to changing temperature regimes, increased CO₂ concentrations and notably increased atmospheric N depositions (Cannellet al., 1998). Furthermore, due to the spatially explicit nature of ALS and DAP data, fine-grain spatial differences in site productivity, which are difficult to capture with field data, can be observed.

Tree species composition is another forest attribute that remains challenging to estimate, regardless of the remote sensing technology or methods applied (Ørkaet al., 2021). Innovations using deep learning may improve outcomes (Kattenbornet al., 2021;Maltamoet al., 2021); however, the main challenge associated with tree species mapping from remotely sensed data is that approaches are often not transferable to larger areas and methods or data sources that work well over small areas do not function or do not function with the same level of accuracy when applied over larger areas (Fassnachtet al., 2016). There is often also a disconnect between the inventory information need (i.e. detailed species composition) and what the remotely sensed data and methods are capable of reliably mapping (i.e. dominant species, genus, forest type). Along with the estimation of stand age, accurate mapping of species is undoubtedly one of the greatest challenges associated with the use of remotely sensed data in forestry.

Plot-level measurements (ground)

Plot-level measurements on the ground at geolocated locations form the basic elements of most traditional forest inventories (Table 2). Remote and proximate sensing technologies can contribute to improvements in the efficiency, accuracy, precision and level of detail of field measurements and are able to capture attributes of the forest stand that are difficult to measure with traditional field devices. Furthermore, photographs and 3D representations of a field plot collected with these techniques represent a permanent 2D or even 3D data record for that plot that allows for retrospective checking of the plot data, and that may support analyses of additional attributes in the future. Improvements of field-plot measurements are crucial in the remote sensing context since any remote sensing-based approach will benefit from an increase in quantity and quality of field plot data for calibration and validation. The ways in which remote and proximate sensing technologies can contribute to field measurements are diverse and depend notably on the applied sensing technology. In the following, we will therefore summarize the advancements made in this field with respect to several sensor types particularly including ground-based laser scanning and smart devices and discuss some key challenges and potential solutions for making further progress.

The most studied ground-based laser-scanning approach is terrestrial laser scanning (TLS) using static scanner systems (Lianget al. 2018). After their emergence in the early 2000s, TLS data have been extensively examined for their suitability to measure key forest attributes such as tree position, diameter at breast height (DBH) and tree height (HT) (Disneyet al., 2019). Many studies demonstrated that TLS is suitable for estimating DBH, and with newer scanning systems also HT with very good accuracies which would be suitable to replace field measurements (Calderset al., 2020). However, a key challenge with TLS measurements is that a single scan is typically insufficient to capture all trees in a field plot, due to occlusion effects (Bauwenset al., 2016). Several studies have demonstrated that occlusion effects can be minimized by using multi-scan set-ups (Lianget al., 2016); however, the time and effort related to collecting multiple scans and to co-register the individual scans makes the procedure often less efficient than standard field inventory methods for DBH and HT. However, improved automated co-registration algorithms and methodical approaches mimicking angle-count sampling (Molina-Valeroet al., 2022), as well as mobile laser-scanning systems (MLS) have shown potential to resolve the aforementioned efficiency problems and account for occlusion effects. In MLS, a portable laser scanner is moved through the forest, either with a vehicle or by a human operator (Hyyppäet al., 2020a) while an inertial measurement unit (IMU) tracks the movement of the scanner with high precision even if no reliable GNSS signal is available. Several studies have demonstrated that these devices can efficiently collect point cloud data that is sufficiently accurate to obtain reliable DBH and HT estimates, with RMSE typically ranging between a few cm (0.7–7.0) for DBH and 0.8–6.5 m for height depending on applied scanner and stand situation (Lianget al., 2016); larger errors are typically observed for denser stands and at greater distances from the scanner. However, verification of this technology in more complex forest structures is still mostly missing, with few exceptions (Vatandaslaret al., 2023).

More recently, laser-scanning sensors are being made available on smart phones (Mokrošet al., 2021), and open-source software applications to support tree measurements with these devices have been presented (Gollobet al., 2021). Initial results indicate that these low-cost devices show potential to capture basic forest inventory attributes including DBH and tree position with reasonable accuracies that may be sufficient for certain tasks (Gollobet al., 2021). Due to the limited scanning range of the built-in scanners, these low-cost sensors are not suitable for estimating HT, and data acquisition is more time-consuming in comparison to mobile laser scanning. However, due to their off-the-shelf availability and ease-of-use, they could still complement and enhance established forest inventory approaches in the field.

So far, most studies examining ground-based laser-scanning approaches have focused on imitating traditional field plot inventories, with considerable efforts to identify and segment individual trees and estimate the DBH of the detected trees.Disneyet al. (2019) provide an overview of open-source software solutions for these tasks; however, an increasing number of studies are now addressing other forest attributes. One widely examined application is the reconstruction of the complete 3D model of stems (taper) as well as the branching structure (Lianget al., 2016;Calderset al., 2020). This is particularly interesting in the context of volume and biomass estimation since the gold-standard for biomass estimation in the field, i.e. the use of allometric models between biomass and DBH and in some cases HT, is known to introduce a notable amount of uncertainty (Lianget al., 2016;Disneyet al., 2019). Detailed 3D-models of stems and branches in combination with wood density information allows for very accurate biomass estimates for individual trees which are close to the values obtained with destructive sampling. This technique could hence also allow for the measurement of notably more trees than is possible with destructive methods and this in turn may also support ongoing efforts to harmonize growing stock volume estimates between regions and countries that often use differing allometries (Stovallet al., 2018;Gschwantneret al., 2022).

An additional field of application which has so far received limited attention is the immense potential of 4D (3D + time) ground-based laser-scanning data to capture forest growth processes on a very fine level of detail (Eitelet al., 2016). Long-term time series of TLS scans of forest stands could substantially increase current silvicultural knowledge on how trees compete and react to silvicultural treatments on fine spatial scales. For example, in highly resolved TLS scans, it should be possible to observe the growth of individual branches and hence also tree crown forms. Such processes are currently not captured in established tree growth simulators, most likely also due to the lack of data. First long-term experiments to collect such time series of TLS data have been initiated in Finland (Camposet al., 2021), but there is large potential for additional research in this field.

Finally, ground-based laser-scanning data may also be suitable to answer a range of ecological questions and applications. For example, TLS data have been used to derive metrics that serve as indicators for the structural complexity of a forest site. These may, for example, relate to micro-habitats and can support the identification of retention sites that support nature conservation (Freyet al., 2020). As summarized byMalhiet al. (2018) such data may also make a significant contribution to an improved understanding of causes and consequences of variation in tree form for example due to phylogeny, biotic and abiotic environmental factors (e.g. competition for resources, wind, water availability, etc.) or, as discussed above, silvicultural treatments. Corresponding analyses may be conducted with both mono-temporal and multi-temporal datasets.

In summary, challenges still exist with respect to integrating ground-based laser-scanning approaches into efficient operational work-flows that can be applied also in complex forest stands. Technical advancements, particularly in the field of mobile laser scanning and further progress with algorithms enabling a robust extraction of tree metrics may contribute to address these challenges. Further opportunities for future research lay in the analyses of multi-temporal ground-based laser-scanning data to analyze growth behaviour of trees at fine detail and in applying these data to address increasingly pressing questions related to biodiversity conservation and corresponding needs to develop metrics able to describe the ecological values of forests.

Beyond laser scanning, data acquisition from a variety of ground-based platforms, consumer grade off-the-shelf smart-devices (Marzulliet al., 2020;Arietta, 2021) and RGB-camera systems (Murrayet al., 2018;Pernget al., 2018;Daiet al., 2021) have recently been suggested as tools to measure forest attributes in the field. Several studies have used RGB-cameras within forest stands to collect 3D photogrammetric point-clouds and then extract DBH (RMSE of approximately 2 cm) and tree positions (0.2–2 m) from the collected data (Tomaštíket al., 2017;Marzulliet al., 2020). However, these studies have primarily been conducted in relatively simple forest stands with low understorey cover. Furthermore, significant time is required for collecting and processing the RGB images to derive the target attribute (Piermatteiet al., 2019); however, this approach may be efficient if diameters at multiple heights have to be recorded (e.g. to approximate stem taper, as perMulverhillet al., 2020) (Marzulliet al., 2020). Given the recent improvements in mobile laser-scanning systems discussed above that are now already available in smart devices, it currently seems unlikely that an efficient and operational approach for a wide variety of forest stand situations will evolve from approaches using only RGB-images.

In contrast, approaches that use several 2D-RGB or 360° spherical images of forest stands to estimate forest stand attributes such as basal area, stems per ha or wood volume seem promising. As discussed inDaiet al. (2021), 360° panoramic images enable an easy implementation of horizontal point sampling also known as angle count sampling (Bitterlich, 1984). The problem of occluded trees can to a certain degree be corrected for using statistical approaches. Since consumer-grade 360° cameras are now available at comparably low prices, such an approach may develop into an alternative to traditional horizontal point sampling using a Bitterlich relascope or related instruments. Combining 360° images with additional DBH and HT measurements of a few tree individuals selected by a sector subsampling approach may serve as a highly efficient biomass inventory (Daiet al., 2021). Comparable approaches to estimate basal area and tree stem number have also been suggested in the form of some smartphone apps (Vastarantaet al., 2015;Siipilehtoet al., 2016;Rybakovet al., 2018;Ficko, 2020;Pitkänenet al., 2021).

Smart devices were also used to derive estimates of forest cover, canopy openness, LAI and health state using upward-looking photographs. These attributes are otherwise typically either assessed by visual assessment in the field, expert-interpretation of aerial images or by using data collected with comparably costly devices. Estimates from visual assessment are known to have comparably high uncertainties (Tichý, 2016). Recent works for example byArietta (2021) show that smartphone spherical panorama images are able to produce reliable estimates of various attributes describing canopy structure, performing similarly or even better than established approaches.Bianchiet al. (2017) andTichý (2016) had already earlier reported a good potential of smartphones equipped with a fish-eye lens to derive canopy structural attributes. Upward looking photographs may also enable information on the health state of certain tree species. For example,Murrayet al. (2018) demonstrated how the fractal dimensions of hemispherical images of sessile oaks in leaf-off state can be used to provide an estimate of the health state of the trees. As these technologies and associated approaches improve, they may provide another source of information for plot-level assessments.

In summary, there seems to be at least two application fields in which smart devices could play an important role in future field-inventories in forests: (1) the derivation of wood volume or basal area estimates using horizontal (360°) images mimicking the horizontal point sampling approach and (2) the derivation of canopy structural attributes from hemispherical images. If, how, and when such approaches find their way into operational forest inventory remains to be seen and will probably also depend on further studies that systematically assess these approaches in a wide range of forest types with varied structural complexities.

Plot-level measurements (drones/RPAS)

Besides ground-based remote and proximate sensing techniques, data collected by drones may also support local forest inventory tasks. With the increasing availability of user-friendly off-the-shelf drone hardware and software that makes RGB- and photogrammetric surveying possible even for unexperienced users and the parallel continuous improvement of passive and active remotely piloted aerial systems (RPAS) sensors (including laser-scanning devices in the high-end section of the drone market) has led to a very high number of studies in a short-time period examining the potential of drones and RPAS in forestry (Coopset al., 2019). Numerous studies have examined the use of drones to collect data to derive common forest attributes including forest height (Liseinet al., 2013), tree species (Schieferet al., 2020), biomass (Dandois and Ellis, 2013), health status (Michezet al., 2016), disturbances (Zhanget al., 2018), etc.

One of the key restrictions of drone applications in forestry is their comparably high operating cost, once larger areas have to be surveyed. A sound analysis of cost-efficiency of RPAS is often omitted in scientific case-studies based on the argument that the investments in the RPAS system itself are low. However, this clearly does not show the full picture and one key challenge in most RPAS-based studies focusing on forests may be the proof of cost-efficiency under real-world conditions. In consequence, drone data may be of limited use for many operational mapping applications. The studies mentioned above have been extensively reviewed in nearly 20 review articles published since 2016 (Goodbodyet al., 2017;Torresanet al., 2017). Herein, we therefore decided to focus on an overarching key question that arises with respect to the use of drones in forestry: in which application fields can RPAS technology provide a true added-value?

From our perspective, RPAS data can be particularly relevant in situations where high-resolution data is required over limited areas. Such situations may include, for example, reference data acquisition (Hyyppäet al., 2020b) and to support forest regeneration assessments over selected sites (Goodbodyet al., 2018).

The application of RPAS to collect reference data for coarser spatial resolution remotely sensed data or to replace field measurements has been presented in several studies. One approach is to conduct under-canopy drone flights to collect information on DBH, tree position, canopy height and stem volume (Hyyppäet al., 2020a). Drone data have also been successfully used as training data for a Sentinel-2-based work-flow to map woody invasive species in forests and shrublands (Kattenbornet al., 2019) and to better understand the relationship between common vegetation index products with forest-related land-cover and land-cover changes induced by wildfires (Fassnachtet al., 2021). In some cases, reference information can be obtained directly from the drone data without need to collect additional reference information in the field. One such example is the aforementioned mapping of canopy-dominating woody invasive species that are directly identifiable in the UAV images (Kattenbornet al., 2019;Gränziget al., 2021). In such cases, drone data provide various benefits including greater spatial coverage than typical field plots, a permanent record that supports retrospective assessments and future applications, a high degree of objectivity, and the bird’s-eye perspective, which can allow for upscaling to coarser satellite or airborne data (Kattenbornet al., 2019) acquired from the same above-canopy perspective.

Another application field for RPAS is characterizing regenerating forest areas and assessing the state of natural tree regeneration below the main canopy. With respect to regenerating forests (young stands) the fine spatial grain of drone data has been demonstrated to be of particular value for estimating key attributes such as mean stand height and tree density. This has been demonstrated in very young boreal forests with an average height below 1 m (Pulitiet al., 2019) but also in 7–9-year-old tropical forests (Zahawiet al., 2015). Examples for mapping individual seedlings with high accuracies of more than 90 per cent have also been presented recently (Finnet al., 2022). Seedling frequencies were mapped byFeducket al. (2018). Characterization of natural regeneration below the main canopy may also be possible using below-canopy drone flights, which become increasingly feasible with modern obstacle avoidance systems (Hyyppäet al., 2020a).

In summary, there are numerous studies examining drones for a diverse set of applications in forestry. However, the crucial limitations of drones in an operational context, namely limited spatial coverage, fulfilment of safety regulations, comparably high workload and hence cost (human hours of pilot, travel to the survey sites), and the large variability of the data quality, are often not sufficiently considered in research. At the same time, these points are and will remain the most important obstacles for their operational use in forestry. Overcoming these limitations or alternatively identifying those applications for which these limitations are not of crucial importance is hence likely to be the most important challenge in drone research in forestry.

Forest monitoring

In general, forest inventories are aimed to support forest management and the information needs of the forest sector, although the utility of these inventories often extends to other application areas including water management or nature conservation. Alternately, forest monitoring is applicable over the entirety of forest ecosystems regardless of management status. Put directly, remote sensing can be focused on forest management and inventory or can be aimed at forests more generally in a monitoring context. Typically, many of the same attributes are desired in monitoring and inventory, with monitoring systems developed to increase the temporal frequency for modelling and mapping of these attributes. In recent years, satellite-based monitoring with year-on-year attribute depictions or characterization of change, have become increasingly common. Indeed, a great deal of recent research and applications in remote sensing of forests are focused on using satellite data (such as from Landsat or Sentinel-2) to map forest attributes and dynamics in a spatially explicit fashion over large areas at an annual time step. While the products emerging from these activities may not be forest inventory dataper se, opportunities for both integration and independent utilization exist.

Forest inventories are a necessity to regularly derive detailed baseline information on the state of forests, and indeed forest monitoring is a key component of many NFI programs. However, although forest inventories are typically designed to have a regular remeasurement cycle (e.g. decadal), the actual time between inventory cycles can be much longer and even if the planned intervals are kept, the temporal representation is infrequent and possibly irregular. A second drawback is that some forest inventories may not produce continuous map products of forest attributes at all (e.g. sample-based NFIs), or may only generate mapped outcomes periodically (e.g. in sync with inventory re-measurement cycle). Information on certain forest dynamics, such as forest disturbances, require spatially explicit information that is available with much greater temporal frequency than most inventory cycles in order to match information needs that are sensitive to both location and timing. For example, the extent, severity and impacts of natural disturbances such as wildfire on future timber supply or wildlife habitat. Forest managers may also benefit from monitoring systems that are able to detect and quantify the impacts of other non-stand replacing disturbances such as insect infestations or windthrow, which may increase options to mitigate impacts and update forest management plans according to current conditions.

Some forest ecosystems may not be subject to any kind of regular forest inventory as a function of ownership or because the forests are of limited economic interest due to low productivity or limited accessibility. For example, the area of unmanaged forest in Canada represents approximately 118 Mha (Ogleet al., 2018) or about 20 per cent of Canada’s forested ecosystems (Wulderet al., 2008b). By way of contrast, this 118 MHa of unmanaged forest area in Canada is more than twice the size of the managed forests of Finland, Norway and Sweden combined. While variable internationally, many other nations similarly have large forest areas that are outside what is considered managed forests. One prominent example are tropical forests around the world. In these forests the demand for forest information is large but due to the complexity of tropical forests and the often limited financial resources of the countries harbouring large areas of tropical forests, the corresponding inventory infrastructure is often sparse (McRobertset al., 2010) and systematic management is rare.

Unmanaged forests are often not subject to management-level inventories and the associated data collection activities that are implied. However, these unmanaged forest areas are included in national forest inventories and are integral to national and international reporting obligations (e.g. State of the Forests, FAO Forest Resource Assessment and reporting to UNFCCC). Additionally, these unmanaged forest ecosystems provide a wide-range of valuable ecosystem services, including animal habitat, harbours of biodiversity, storage and filtration of water, and the exchange of carbon between the atmosphere and terrestrial vegetation (Andrewet al., 2012).

For such areas where regular forest plot installation or inventory mapping is irregular or lacking, remote sensing provides the only source of information regarding the forest land-base. Use of available ground data for model calibration or validation, augmented with sample plot data collected for specific use-cases inform models for mapping of forest attributes over large areas. These spatially explicit maps provide otherwise unavailable information on forest characteristics. Implementation of time series analyses further enable the monitoring of change in state as well as to mapped attributes. Such remote sensing-based approaches outside of the regular forest inventories are particularly suited for providing lacking baseline information on the state of these forests in the form of estimates of selected forest attributes. The systematic implementation of models using remotely sensed and auxiliary spatial data means that mapped attributes are produced for all treed area, not just inventoried area, and provide consistently produced measures that can augment, inform or be integrated with forest inventory data, where available.

One key component to any of such monitoring systems with high temporal resolution are freely accessible data at a high processing level. Open access to remotely sensed data has been shown to broaden the use of a given dataset, but also to improve outcomes, and enable programmatic use of imagery (Zhuet al., 2019). Open access to data that are stored in disparate or arcane locations or in datatypes that require multiple stages of pre-processing prior to use in an analysis are less likely to be used. Costs are also incurred through additional processing needs, errors that can be introduced, and a higher level of expertise required. Reducing needs for pre-processing is a co-requirement for the increased use of satellite imagery, whether by individual users or facilitating inclusion in cloud-based environments (Gorelicket al., 2017).

More than two decades agoTeilletet al. (1997) already indicated that to promote the use and uptake of remotely sensed data, important geophysical and biophysical parameters would be required by users in a ‘ready-to-use’ form. Surface reflectance and derived parameters such as vegetation indices were identified. Presciently, besides the need for atmospheric correction, geometric registration and the correction of bidirectional effects, the authors also identified traceability of processing, flexibility of automated interpretation techniques including scale considerations, as well as data formats and standards as important requirements.

Over the past two decades, numerous advances have been made over these thematic areas making analysis ready data (ARD) possible (Dwyeret al., 2018). This was first demonstrated for Landsat and then other sensors followed. The availability of ARD has led to several regional, continental and global remote-sensing-based geoinformation products that are nowadays widely used. Many of the aforementioned innovations (open data, computational capacity) have likewise enabled large-area mapping and estimation of forest structural attributes using approaches such as non-parametric k-nearest neighbour (k-NN) imputation (Chiriciet al., 2016;Coopset al., 2021a,2021b), which, for example, have been a component of Finland’s NFI since the early 1990s (Tomppo and Halme, 2004).

Amongst such geoinformation products, land cover, land cover change and forest structure are basic attributes required for forest monitoring (Whiteet al., 2014;Table 1). For monitoring programs, the characterization of a targeted set of attributes over a large spatial extent and at a high temporal cadence is required. Identifying the appropriate spatial scale for both data inputs and outputs is critical. Given the interest in large-area coverage, spaceborne data that provides wall-to-wall and repeat coverage are typically required for producing spatially explicit (mapped) outcomes. To date, Landsat has been the exemplar program for obtaining such data, with Sentinel-2 now offering similar utility independently (Buchhornet al., 2020) or in combination (Griffithset al., 2019). Programmatic traits of Landsat were replicated in the Copernicus program with Sentinel-2 (Druschet al., 2012) and include, open access (Woodcocket al., 2008), radiometric calibration (Markham and Helder, 2012), robust ground segment and archive (Wulderet al., 2016).

Consequently, it has been possible to develop a spatially and spectrally harmonized analysis ready data product based upon two Sentinel-2 satellites and Landsat 8 (Claverieet al., 2018). Coined Harmonized Landsat Sentinel-2 (HLS), the integrated data result in an effective revisit time of 2.9 days at the equator and more frequently towards the poles (Li and Roy, 2017). With the recent launch of Landsat 9, this will further improve (Maseket al., 2020). Sentinel-2 and Landsat are effectively a virtual constellation (Wulderet al., 2015), allowing for data interoperability at present and integrated development and launch planning into the future. Commitments by governments and cross-government space agencies to continue to build and launch such satellites to collect data and place these into an open access archive allows for governments and related agencies to make forward going plans with minimized risk.

Confidence with respect to future data availability is a key requirement to stimulate planning and development of satellite-based forest monitoring frameworks. To date, detecting and mapping forest change and capturing the related effects on forest attribute estimates (mostly forest area) has been one of the most significant contributions of remote sensing to forest monitoring efforts (Hansen and Loveland, 2012). Change detection methods have evolved rapidly with the advent of free and open data, resulting in an increasingly sophisticated array of time-series-informed algorithms and approaches (Zhu, 2017). Time series data have likewise transformed land cover mapping (Gómezet al. 2016). Historically, change detection and land cover mapping were considered as independent activities whereas today these two activities are increasingly integrated (Wulderet al., 2018). This integration enables insights on where and when changes have happened to inform the allocation of land cover classes, and how successional processes result in class transitions over time (Hermosillaet al., 2018). There has been an increase in the extent and variety of land cover maps being produced, with mapping undertaken by a range of agencies (Gonget al., 2013;Buchhornet al., 2020;Zanagaet al., 2021). With a proliferation of new land cover products available to end users, tensions between regionally-specific and global products emerge (Tulbureet al., 2021) as do methods to harmonize products (Liet al., 2021). This again stresses the need for robust and transparent assessment of accuracy.

Three main trends have characterized research on this topic in the last decade. First, there has been a shift from scene-based analyses towards pixel-based approaches that rely on pixel compositing algorithms (Griffithset al., 2013;Whiteet al., 2014). Second, change detection algorithms, enabled by open data policies and increases in computational capacity, have enabled forest changes to be captured with a spatial resolution of ≤30 m that is commensurate with anthropogenic impacts on the landscape offering relevant insights for forest management. Finally, attribution of these changes to a disturbance agent (i.e. harvest, fire, non-stand replacing;Hermosillaet al., 2015;Zhanget al., 2022) has been a critical to characterizing forest dynamics and post-disturbance recovery (Whiteet al., 2017b;Senf and Seidl, 2020).

Innovations in change detection have enabled global (Hansenet al., 2013), continental (Senf and Seidl, 2020), national and regional (Souza Jret al., 2013;Hermosillaet al., 2016;Brownet al., 2020) forest change products to be generated. While mentioned products have been widely evaluated over some parts of the world, they may be largely unvalidated in others and their usefulness may vary with geographic region. For example, while the forest-cover product ofHansenet al. (2013) proved to be reliable for boreal and temperate forests, a strong underestimation of forest cover has been reported for semi-arid and arid forests (Bastinet al., 2017;Cunninghamet al., 2019;Shafeianet al., 2021). Such discrepancies point to a key challenge emerging regarding the use of such global map products of forest attributes. Confidence is thus required in the protocols defined and implemented to characterize the quality of the mapped attributes (Tulbureet al., 2021). Circumspection by users will be required to understand the amount and nature of the error present and to incorporate this understanding when making management or policy related decisions (Palahíet al., 2021).

Given the availability of data and increasing computational capacity, a proliferation of forest change products can be expected in the coming years. Multiple similar product descriptions will confront users with various questions related to these products: What sources are authoritative? Which are supported by scientific transparency and peer-reviewed methods? What is the reported accuracy and what were the data and methods applied to determine and report accuracy? The definition of traits and best practices will aid in building confidence in the outcomes (Olofssonet al., 2013,2014;Mac Dicken, 2015), with additional detail provided below. Of concern is that with increased ease of access and application, inexperienced users are in a position to produce regional and even global maps using shared algorithms and cloud-computing environments such as the Google Earth Engine. Good knowledge of the training data used and the methods applied, and the application of a sound accuracy assessment protocol is critical to having confidence in the maps generated and for correctly interpreting the information contained in the maps (Palahíet al., 2021;Breidenbachet al., 2022). In this context, acknowledging and quantifying uncertainty is increasingly critical when using remotely sensed data to characterize forest status and changes over large areas (Franciniet al., 2021).

We expect that time series of forest attributes will increasingly be used to provide value-added information in support of science and policy. The capacity to generate annual information products enables the accounting for gains and losses in certain forest attributes such as biomass in a spatially explicit manner. When combined with the aforementioned forest change attribution, the biomass consequences associated with different disturbance types can be quantified over time (Wulderet al., 2020). As mentioned earlier, such information can also be integrated with field-plot information of NFIs to increase precision of forest attribute change estimates (Breidenbachet al., 2021).

However, to reach such a system delivering time-series of forest attributes, some methodical challenges still have to be overcome. One key challenge is that the quality of these products has to fulfil user needs and in order to confirm that this quality is reached, clear and transparent validation approaches are required, following best practices that have emerged in the remote sensing scientific literature. Commonly accepted best practices for assessing the accuracy of forest change products are now widely applied (Olofssonet al., 2014), and community-driven initiatives have sought to build consensus on validation protocols for forest structure products such as biomass (Duncansonet al., 2021). Accuracy assessments of remotely sensed land cover products in particular have been the focus of extensive research (Foody, 2002;Hansen and Loveland, 2012). Studies have demonstrated the importance of using appropriate measures of accuracy (Stehman, 1997) and provided guidance on design and analysis (Stehman and Czaplewski, 1998). Accuracy assessment protocols have been synthesized in the context of large-area land cover products (Strahleret al., 2006;Wulderet al., 2007), yet as detailed inFoody (2021), despite decades of research on this topic, many studies continue to use inappropriate designs, data sources and statistics to assess and report accuracy of their land cover products.

Uncertainties in map-assisted and map-based estimates

Traditionally, maps or other sources of spatially continuous auxiliary information have been used in large-area forest inventory, such as NFI, in the design-phase to stratify the population under study prior to allocating ground plots (Tomppoet al., 2014), or in the estimation phase to enhance the precision of estimates by post-stratification (McRobertset al., 2012). At local scales, maps of various kinds have been used to help define homogenous treatment units in the form of forest stands, for which biophysical stand attributes have been estimated from field measurements, manual interpretation of aerial photography (Næsset, 1997a,1997b), or more automatically by combining field plot data and various types of remotely sensed data (Maltamo and Packalen, 2014). Over the past 20–30 years there has been a tremendous increase in the application of remotely sensed data from airborne sensors, satellites and drones, to produce pixel-based prediction maps of various forest attributes at a huge range of geographical scales and for a multitude of purposes. The increasing amount of greater-quality remotely sensed data from different sensing techniques, greater computational power and emerging prediction techniques such as machine-learning and other non-parametric approaches, have made computer-based production of prediction maps a relatively trivial task for researchers and practitioners.

However, a critical obstacle for a meaningful use of these derived map products, is the availability of methods for the rigorous statistical assessment of the uncertainty associated with estimates from such maps. Maps can be useful tools for showing the location and spatial context for a certain attribute, for example biomass, and especially so if maps are accompanied by pixel-based uncertainty maps. However, maps are often just an intermediate product enroute to an estimate, for example, the wood volume in a forest stand subject to potential harvest, or the change in carbon stocks in the forests of a country under reporting to international conventions. In the first case, a lack of trustworthy estimates of uncertainty may lead to suboptimal harvest decisions and financial losses (Eidet al., 2004), in the latter case, it may lead to a failure to meet international reporting obligations (IPCC, 2019a), and/or a failure to reach aspirational targets (e.g. for reducing greenhouse gas emissions).

It is often fruitful to distinguish between two primary modes of estimation and inference. Firstly, estimation for which the map is used as auxiliary information to enhance the precision of the estimate while the estimated precision—in the form of a standard error or a confidence interval—is based on the design-properties of the sample (typically consisting of data collected in field plots), i.e. design-based or probability-based inference (Opsomeret al., 2007;Baffettaet al., 2009). Secondly, estimation for which the map itself is used as the primary source of data and no assumptions are made regarding the field data that are used to calibrate the prediction model or technique, i.e. model-based or model-dependent inference (McRoberts, 2006).

There are also combinations of these two primary modes of inference, often denoted as hybrid inference (Coronaet al., 2014), which is particularly relevant when remotely sensed data are collected as a probability sample of ALS transects, for example. Further, field plots used for model construction may or may not have been collected according to probabilistic principles or perhaps in another area than the ALS survey (Ståhlet al., 2011). The GEDI mission has adopted hybrid estimation for the gridded 1 km × 1 km L4B biomass product (Pattersonet al., 2019).

Finally, another mode of inference is used when estimates are calculated from a sequence of prediction maps (i.e. a sequence of prediction models), which is often denoted hierarchical inference (Saarelaet al., 2016). A highly relevant example is the use of ALS data as an intermediate dataset to link, for example, growing stock volume in field plots to spatially continuous data from satellite remote sensing, such as optical or radar satellites (Saarelaet al., 2016). Likewise, ALS is sometimes used to link field observations to spaceborne data even when the spaceborne data come from a sampling instrument, such as spaceborne lasers (i.e. GEDI), an application for which the inference mode may be designated hybrid hierarchical inference (Holmet al., 2017). Below we highlight some of the most critical research gaps regarding inference that must be addressed in order to fully benefit from the huge stream of powerful remotely sensed data in a statistically rigorous manner.

The use of maps as auxiliary information in combination with probabilistic samples of field data is a mature option in forest inventory and practical examples and even uptake of this mode of inference in operational monitoring programs are many in number. Applications in this field have typically adopted model-assisted estimators such as model-assisted generalized regression estimators or model-assisted difference estimators. Sometimes the adopted sampling designs are those commonly used in forest inventory, such as simple random or systematic sampling (McRoberts, 2010); sometimes more complex designs such as various forms of cluster sampling or two-phase sampling are used (Eneet al., 2016). In this mode of inference, a probabilistic field sample collected in the area of interest is often used to construct a parametric regression model which subsequently is used to produce the predictions (or prediction map). When the same sample is used for parametric model construction and then subsequently for estimation, the prediction model is characterized as internal and the generalized regression estimator may be used. When the model is constructed partly or entirely on the basis of field and remotely sensed data external to the area subject to estimation, the model is characterized as external and the difference estimator may be used (McRobertset al., 2022a).

In recent years, other prediction techniques such as machine learning techniques and deep learning techniques have become popular for map production. Design-based and model-assisted estimators may very well be used with auxiliary information in the form of prediction maps produced with such prediction techniques, or any other technique for that matter. It is, however, important that the analyst choose an appropriate model-assisted estimator. That will typically be a difference estimator. This was demonstrated byEstebanet al. (2019) who used random forest prediction models and ALS data in combination with model-assisted difference estimators to estimate volume and biomass stocks and their changes over time. When local probabilistic field samples exist, design-based model-assisted estimation can be a means to assess systematic map errors when the model is external, such as when a global biomass map is used to estimate biomass locally (Næssetet al., 2020). Such estimation can also be a means to assess to what extent such a global map can be used to improve precision of estimates, i.e. to determine the ‘value’ of the map. An associated research gap is a need for more empirical evidence associated with the gain in precision enabled by the use of such map products (i.e. ‘value of the map’) as well as improved insights in systematic map errors of existing global and regional forest attribute maps. This gap could be addressed by using various types of maps, in various ecozones of the world, and for different target areas. Such information can help in designing new inventories, such as new national forest inventories (Grafströmet al., 2017) by quantifying the required sample size of field observations to reach a certain precision target for a given attribute.

Model-based inference with simple parametric regression models for which the prediction map serves as the primary source of data, has gained substantial attention in the scientific literature in recent years. However, there remain many examples of less formal, partly incorrect or even flawed and misleading inference in the scientific literature (see discussion inGregoireet al., 2016). Examples of uptake in practical applications are few or non-existing, with an exception for a nation-wide and spatially explicit map of various forest attributes publicly available online where model-based estimates of uncertainty for user-defined small spatial areas (e.g. segments resembling forest stands) are produced from ALS or DAP data (Hauglinet al., 2021). Model-based inference assumes additivity of two primary sources of variability, namely the prediction variability and the residual variability (McRobertset al., 2022b). The prediction variability is caused by the fact that a certain sample of training data was used to calibrate the model. Had a different sample been used, the calibrated model parameters would have been different, and thus also the resulting output map. In other words, a myriad of different samples could potentially be drawn from a population, at least if the population is large, and they would all result in slightly different model parameter estimates. So the prediction variability is manifested in the variance and covariance of the model parameters. We have closed-form analytical estimators for this variance component (McRobertset al., 2014), but simulation-based techniques such as bootstrap resampling may also be used (McRobertset al., 2022a).

When considering model-based inference and residual variability, it is useful to distinguish between local, so-called small area estimation, and estimation for larger domains or territories. For local estimation problems, all the primary sources of variability mentioned above usually need to be accounted for. However, in practice, data will often be lacking for estimation of spatial residual covariance and also for residual variance. There are few examples in the current literature trying to determine the minimum size of an area for which it is necessary to account for the residual variance and covariance. A couple of exceptions are the study byBreidenbachet al. (2016) who analyzed the proportion of total variance contributed by residual variance and covariance in estimates of timber volume in forest stands, and the study byNæssetet al. (2019) who analyzed the similar variance and covariance components in estimates of change in tree height of small pioneer trees in a boreal-alpine treeline in 1-ha spatial monitoring units.Breidenbachet al. (2016) reported that stands needed to be as great as 10 ha for the residual components of the estimator to be negligible whereas ignoring the residual components would result in an underestimation of variance for very small stands (<1 ha).Næssetet al. (2019) on the other hand, reported that the residual components contributed less than 3 per cent of the total variance for units as small as 1 ha. This illustrates that the spatial scale for which an estimation problem must be considered a ‘small area’ problem, is context dependent and that general recommendations across biophysical attributes and types of populations can hardly be provided. That calls for caution when model-based inference is to be considered for small domains in a particular situation.

Inference for a number of biophysical attributes in stands in operational forest inventories under the ABA method, such as volume, mean tree height, mean diameter and site productivity, typically is a small-area estimation problem. Hence, this lack of empirical evidence of the magnitude of residual variance and covariance for different stand sizes, in different forest types, and for different biophysical attributes is a huge knowledge gap. Nevertheless, such lack of empirical knowledge has likely not hampered the uptake of remotely sensed data and map products in practical forestry because practical applications have to a great extent been justified by extensive accuracy assessments based on comparisons against validation data in the field. However, general experiences from accuracy assessments are no guarantee for accurate estimates in a particular stand subject to forest management. Missing or misleading information on uncertainties in forest management-level inventories may lead to incorrect and sub-optimal decision-making in forest management and economic losses in the sector (cf.Eidet al., 2004). Thus, there is a great potential for even greater revenues if additional information on uncertainty in individual cases can be provided, rather than general empirical evidence from accuracy assessments which may or may not be relevant in a particular case.

For larger areas, such as entire forest properties, counties, provinces, nations or even continents, the residual variance and covariance components of the model-based variance estimator are usually negligible (Ståhlet al., 2011) and therefore can be ignored. Yet, we see that such large area estimates from maps, e.g. a global biomass maps, quite frequently only aim to account for the residual components of the variance and disregard the dominant component; i.e. the prediction variability, that represents, say, >95 per cent of the overall variability. There is therefore a great need to establish sound practices for rigorous statistical inference for large-area estimates from maps, both in the scientific community and in practice. For example, generic guidance on the use of biomass density maps for greenhouse gas inventory has recently been established by theIPCC (2019b).

Whenever a user of a map wishes to produce an estimate for a certain area, inference cannot be made by the user unless the map producer provides the necessary metadata required for estimation of the prediction variability (IPCC, 2019b). For parametric regression models, such metadata include the variance–covariance matrix of the parameter estimates (McRobertset al., 2019). Scientific publications rarely include this information and the scientific community should be encouraged to publish such information and otherwise be informed that published maps have limited utility unless such metadata are supplied. One current exception is the published regression models adopted for prediction of biomass for individual GEDI laser waveforms. The documented models come with necessary metadata that allow the users to estimate uncertainties of estimates of biomass for any arbitrarily selected target area (Duncansonet al., 2022).

As use of non-parametric prediction techniques are becoming ever more popular among map producers, there is a need for estimators and guidance for inference based on such prediction maps. Methods for estimating variance for e.g. kNN-based map estimates have existed in the forest remote sensing literature for a while (see review inMagnussen, 2013). For large-area estimates in particular, for which there is a need to account for prediction variability, bootstrap resampling methods may be used for any map prediction or map production technique (Estebanet al., 2019;McRobertset al., 2022b). However, whenever a map user wishes to produce a map-based estimate for a certain area, there is a need for metadata to enable rigorous estimation of the prediction variance. Unlike parametric prediction techniques, the need for communication of necessary metadata cannot be solved by a simple parameter variance–covariance matrix, but rather in the form of a complete variance–covariance matrix across all pixels in the area subject to estimation. It is practically infeasible or even impossible to communicate such a matrix for a large area. As indicated above, simply ignoring this source of variability because it is complicated to handle, is hardly a viable solution as it may account for, say, >95 per cent of the overall variability in certain large-area estimation problems. Although recent research has addressed some practical solutions to this problem (McRobertset al., 2022b), it remains an open question how such metadata can be estimated and communicated in a feasible and precise manner. It should indeed be a research priority to solve this problem because the continued use of these popular non-parametric prediction techniques means that the resulting maps will be useless for multiple purposes unless this scientific and practical problem is resolved.

Summary and conclusions

As outlined herein, remote sensing has developed into an omnipresent technology in the scientific field of forestry and its usefulness to support inventory, monitoring and mapping tasks at levels reaching from the field-plot scale to national or even global levels has been demonstrated. However, the uptake of remote sensing into operational forest management and monitoring still varies widely, both geographically and with respect to the particular application. In this communication, we identified some key challenges and opportunities of remote sensing in the context of forestry and outlined some potential solutions and pathways for how to address and exploit these in the future. Below and inFigure 2 we provide a brief synthesis of the most critical challenges and opportunities.

1. User-uptake of remote sensing approaches may be slow or absent even though suitable technological solutions exist.

Figure 2

Future directions for remote sensing applications in forestry (icons fromFlaticon.com).

Open in new tabDownload slide

This challenge has been resolved in a few countries by a close communication and collaboration between researchers (including between remote sensing and forest scientists), forest industry stakeholders and regulators. Experiences have shown that the key factors to increasing user-uptake of new technologies include (1) a good understanding of user information needs; (2) clear and transparent communication of the potential and limitations of (remote sensing) technology; (3) technology transfer, knowledge exchange and synthesis of best practices to ensure scientific knowledge and experiences are translated into operational practice; (4) consideration of existing work-flows and local circumstances; (5) direct or indirect financial incentives or resources to ease a shift towards the new technology in operational use; (6) provision of data infrastructure that enables long-term planning security (e.g. government supported, freely available repeated ALS surveys, operational satellite missions), (7) communication of success stories and collaborations on an international scale to initiate a learning process in geographic regions where user uptake lags behind. Finally, end users often have limited knowledge or understanding regarding the level of error or uncertainty associated with the data they currently use in their forest inventories (e.g. air photo interpretation). As a result, quantifying the value proposition of new and emerging data sources and technologies is challenging. Studies that can quantify the value of information and or trade-offs associated with different information sources are often a necessary prerequisite to demonstrating the value of alternatives. Inertia due to existing programs, such as for air photo acquisition and archiving, can also limit the enthusiasm for new data streams. New approaches may need to operate in tandem with legacy programs, while workflows and capacity evolve to enable further uptake.

• (2) Discrepancies exist between the potential of remotely sensed data and methods demonstrated in local, spatially limited, and often simplified, case-studies in contrast to the application of these methods under ‘real world’ conditions.

This challenge is increasingly being addressed by large-area investigations across a range of forest conditions that demonstrate the capacity of the technology or data source to address "real world" situations, which often only arise when scaling up an approach to large, heterogeneous, areas. Explorations of new technologies or data sources should consider a range of forest conditions, to the extent possible. Likewise, results should be interpreted appropriately within the context of the forest environment in which the study was done, a practice that is often overlooked. A focus on single-layer forest stands, with relatively simple structure and little to no understorey may not provide a real test of a technologies’ performance under a broader range of forest conditions.

A related challenge is that spatial outputs derived from remotely sensed data may not have been assessed or validated at a spatial unit that is relevant to the end user. For example, validation results are often reported at the field-plot level, but forest managers often make decisions at the stand level, and therefore may be more interested in the level of error and bias associated with predictions made at the stand level and not the plot level.

Lastly, benchmarking studies that systematically assess data, methods, parameters and approaches across a range of forest conditions and environments can be very valuable, both for the scientific community and the end users. Moreover, the publication and sharing of benchmarking datasets covering wide geographic regions and forest conditions as well as the use of realistic synthetic data (as is often done in related scientific fields such as in the image processing community) may also contribute to addressing this challenge. One key requisite for such studies is often the ability to access and share data openly without restrictions, particularly with respect to datasets that are useful for reference purposes (e.g. field plot data, ALS, VHRS satellite data).

• (3) Map products derived from remotely sensed data are increasingly available but may lack reliable quality indicators and uncertainty estimates.

With the increasing availability of easily accessible remotely sensed data in combination with (cloud-based) turn-key solutions to derive map products from these data, even unexperienced users can nowadays create remote-sensing-based maps. As such, it is of utmost importance for the scientific and user community to further develop and disseminate standards of map quality that objectively communicate the uncertainties associated with these map products. Lack of such standards can lead to a risk of negatively affecting decision-making and in turn undermining the perceived value of remotely sensed information. The development of such standards is not straightforward and may in some cases require methodological advancements (e.g. adequately accounting for prediction variability in large area estimates of forest attributes and even residual variability when the target areas are small).

Furthermore, an educative effort is required to acquaint current and future professionals working in the field of remote sensing and forestry with the importance of appropriate sampling designs and the necessity of statistical knowledge to correctly quantify and understand uncertainties in map products. Users will make use of products that are available. It is hence in the best interest of jurisdictions and national agencies to provide high-quality data products in an open access fashion with appropriate meta-data on uncertainties for public consumption. As identified herein, there are a myriad of research issues related to characterizing map uncertainty that merit both further investigation and improved communication to end users and stakeholders.

• (4) Increasing opportunities to use remote sensing to characterize new forest attributes or to characterize existing attributes using new approaches and data sources.

With forest managers acknowledging non-timber related products and services of forests for some time now, the need for different inventory approaches, indicators and metrics arises. This is an opportunity for the remote sensing community, since in some cases, innovations in measurement approaches arise as a function of the use of remote and proximate sensing technologies rather than in the case for attributes where established (likely plot-based) workflows exist. For example, ground-based sensing instruments that capture in-stand 3D information may play an important role in developing structural indicators related to the ecological function of forests (e.g. habitat functions, dead-wood fractions, etc.). Furthermore, site productivity, as one of the most important forest attributes, has received more attention by the remote sensing community recently. With increasing availability of repeated high-quality forest height data, site productivity can be derived on fine spatial scales. This may not only be useful to update established local forest growth models which are often not adequate due to climate change effects but also to improve the spatial details of yield prediction models and contribute to improved forest management decisions.

Acknowledgements

We acknowledge Håkan Olsson and Matti Maltamo for valuable comments on forest management inventory in the Nordic countries and Sebastian Schnell for corresponding comments on the situation in Germany. We also thank the two anonymous reviewers and the two editors for their constructive and helpful comments.

Conflict of interest statement

None declared.

Funding

This study received no specific funding.

Data availability

No new data were generated or analyzed in support of this research.

References