Disclosure of Invention
The invention provides a ground elevation inversion method for photon counting laser radar data, which is used for solving the problem of low inversion precision in the prior art.
In order to solve the technical problems, the technical scheme and the corresponding beneficial effects of the technical scheme are as follows:
the invention relates to a ground elevation inversion method for photon counting laser radar data, which comprises the following steps:
1) Acquiring original photon data of a photon counting radar, extracting noise photons in the original photon data, and removing the extracted noise photons to obtain signal photons;
2) The ground photons are extracted from the signal photons by the following method:
flipping signal photons upside down;
Setting a cloth, and placing all particles included in the cloth above all signal photons;
The method comprises the steps of carrying out cloth simulation, namely simulating each particle in the cloth to fall down due to external force applied to the particle, and marking the particle collided with a signal photon as an immovable particle when the particle collides with the signal photon;
repeating the cloth simulation process until a set end condition is reached, ending the cloth simulation process, calculating the distances between all signal photons and corresponding particles, and marking the distances smaller than a distance threshold as ground photons so as to realize ground photon extraction;
3) The ground elevation is determined from the final positions of the particles corresponding to all signal photons.
The ground elevation inversion method has the beneficial effects that firstly, the extracted noise photons are removed to obtain signal photons, then, the ground photons are extracted from the signal photons by adopting a cloth simulation method, and finally, the ground elevation is determined according to the final positions of particles corresponding to all the signal photons. Experiments show that the method can effectively extract signal photons, further obtain high-precision ground elevation curves under different land coverage and topography conditions, invert the ground elevation with low average error and root mean square error, invert the precision angle and provide an effective solution for estimating the ground elevation by utilizing photon counting laser radar data.
Further, in step 2), under the action of external force and internal force, the positions of the particles are:
Wherein X (t) represents the position of the particle at time t, Fext (X, t) represents the external force exerted on the particle, including the gravity force exerted on the particle and the supporting force generated by the contact of the particle with the corresponding signal photon, Fint (X, t) represents the internal force exerted on the particle, generated by the spring action between the particles, and m represents the mass of the particle.
Further, the displacement of the particles under gravity is:
Wherein X (t+Deltat) represents the displacement of the particles under the gravity at the time of t+Deltat, deltat is a time interval, X (t) represents the displacement of the particles under the gravity at the time of t, X (t-Deltat) represents the displacement of the particles under the gravity at the time of t-Deltat, and G represents constant gravity.
Further, in step 2), the distance that the particles that are not marked adjust is:
wherein d represents the adjusted distance, VD represents the height difference between adjacent particles, and RI represents the probability weighted cloth hardness, and the calculation formula is:
RI=1·Rst+2·Run+3·Rfl
Where Rst、Run、Rfl corresponds to the proportion of steep slopes, undulating surfaces and flat terrain in long road sections, respectively.
Further, in the cloth simulation process of step 2), the method further includes a step of eliminating the influence of residual noise:
Traversing the immovable particles, and calculating the average height mu and variance sigma of three adjacent particles;
if the elevation angle of the current immovable particle is larger than mu 3 sigma, the cloth will be broken and the position of the current immovable particle will be adjusted so that the current immovable particle can move with the new positioning.
Further, in step 1), a quadtree isolation method is adopted to extract noise photons in the original photon data, and the method specifically comprises the following steps:
Carrying out four-way space division on the space where the original photon data are located to obtain four identical subspaces, judging the photon number contained in each subspace, if the photon number contained in each subspace is greater than 1, continuing to carry out four-way space division on the subspace until the photon number contained in each subspace obtained after final division is less than or equal to 1, and stopping carrying out four-way space division on the subspace;
Stopping performing the four-fork space division to obtain an isolated four-fork tree;
And calculating the isolation depth of each original photon in the original photon data, and marking the original photons with the isolation depth smaller than a depth threshold as noise photons, wherein the isolation depth is the length from a root node to a leaf node, the root node is the space where the original photon data is located, and the leaf node is the photon in the subspace after the four-fork space segmentation is stopped.
The technical scheme has the beneficial effects that noise photons in the original photon data are extracted by adopting the quadtree isolation method, so that not only are the negative effects of land coverage and gradient reduced, but also the change of the data acquisition environment can be adapted.
Further, a depth threshold is determined using the oxford algorithm.
Further, a depth threshold is calculated based on the following equation:
σ2＝ω1(μ1-μ)2+ω2(μ2-μ)2
Where μ1 represents the average isolation depth of the potential signal photons, ω1 represents the ratio of the potential signal photons to the photon data, μ2 represents the average isolation depth of the potential noise photons, ω2 represents the ratio of the potential noise photons to the photon data, μ represents the average isolation depth of all photons, σ2 represents the variance between the categories, and the isolation depth at maximum σ2 is the depth threshold.
Further, in step 3), a cubic spline interpolation method is used to determine the ground elevation.
Detailed Description
The basic concept of the invention is that firstly, noise photons in original photon data are extracted by adopting a quadtree isolation method and removed to obtain signal photons, then ground photons are extracted from the signal photons by adopting a cloth simulation method, and finally, ground elevation is searched by interpolation. The method can effectively extract signal photons and obtain high-precision ground elevation curves under different land coverage and topography conditions. The noise removal algorithm based on the quadtree isolation method not only reduces the negative effects of land coverage and gradient, but also can adapt to the change of the data acquisition environment, and experiments show that the average error and the root mean square error of ground elevation inversion are lower by carrying out ground photon extraction based on cloth simulation, and the inversion precision is higher.
The following describes a method for inverting the ground elevation of photon counting laser radar data according to the present invention in detail with reference to the drawings and examples.
Method embodiment:
In this example, an area of investigation was selected as shown in FIG. 1, which was located in Alaska (latitude: 60 deg. North latitude 30' -63 deg. 30'; longitude: 141 deg. 30' -152 deg. 30W), with an area of >191000 square kilometers in the North-west of the United states. The research area is in the cold temperature zone, belonging to the marine climate of the temperature zone. The average temperature in summer is 4-16 ℃, the average temperature in winter is-7-4 ℃, and annual precipitation is more than 2000 mm. Cool and humid climates allow needle-leaved vegetation growth in the western and eastern valleys of the study area, such as hemlock and spruce. The mountain range is widely distributed, and the altitude varies from-257 meters to 6105 meters. Urban areas only appear on the plain close to the ocean in southwest, and most of buildings are characterized by low-rise houses.
ICESat-2 ATLAS are equipped with three pairs of rail-bound laser beams, each consisting of two lasers, a strong beam and a weak beam. The footprints formed by adjacent pairs are 3.3 km apart in the cross-track direction, while the footprints formed by the lasers in each pair are about 90 meters apart. After format conversion and error correction, the latitude, longitude and elevation of each photon were calculated based on WGS84 ground reference and recorded in ICESat-2 att 03 product. Since the spatial positions of photons are recorded in the ATL03 data, the present embodiment uses them as experimental data to verify the validity of the proposed method of the present invention. As shown in fig. 1, from 1 month in 2020 to 12 months in 2020, 78 pieces of data of the download study area ATL03 were collected in total.
In the investigation region, in-situ data was obtained from an on-board measurement system designed by the Godard space flight center-Godard lidar, hyperspectral and thermal imaging (G-LiHT). The G-LiHT is composed of a high-precision laser radar, a hyperspectral imaging system and a thermal infrared imaging system, and can be used for measuring the ground height, the vegetation structure, the plant leaf spectrum and the surface temperature at the same time. Point cloud data of gravity gradient echo is obtained by using a VQ-480 laser radar of the Liguer company, wherein the density of the point cloud is more than 10 points/m2. And when the point cloud data is processed, adopting a progressive morphological filtering algorithm to filter the points. For ground points, a Delaunay irregular triangle network is established, and a Digital Terrain Model (DTM) is generated through interpolation. DTM has a resolution of 1 meter. Since the data trace obtained in alaska by G-LiHT in 2018, 7, was consistent with the atlas data, DTM was used as reference data. 78 data were downloaded and used for the experiment.
Land cover data and terrain data are used to analyze the performance of the algorithm in different scenarios. The national land cover database of alaska (NLCD) of 2016 is the latest high-resolution land cover classification data of alaska. These data are calculated from the appropriate earth satellite images, the earth coverage is divided into 20 categories with a resolution of 30m. In this example, NLCD2016 data is summarized as vegetation, construction, barren land, and water, as shown in fig. 1. In 2006, the japanese aerospace research and development agency (JAXA) transmitted advanced terrestrial survey satellites to acquire global terrain data. Stereoscopic images with a spatial resolution of 2.5 meters were acquired using a three-line camera on ALOS, which was used to generate global digital elevation models with 5 meters and 30 meters spacing. A 30m pitch digital elevation model is widely used for its stability. The data quality of the newly released ALOS digital elevation model is further improved compared to the old version, especially providing supplemental data for cloud and snow-based data blanks in high-latitude areas, including the alaska research area. In this embodiment, ALOS digital elevation models are imported into ArcGIS 10.8. The investigation region is adaptively divided into flat, hilly and mountain areas according to calculated topographical factors such as altitude, gradient and surface roughness.
The ground elevation inversion method of the photon counting laser radar data is realized based on the selected region and the data, and the whole flow is shown in figure 2.
Step one, acquiring original photon data of a photon counting radar, extracting noise photons in the original photon data by adopting a quadtree isolation method, and removing the extracted noise photons to obtain signal photons. Wherein "isolation" is used to measure the degree of discrimination of photons from other photons, and the process of extracting noise photons is as follows:
1. the noise distribution is assumed to be sparse and the signal is assumed to be dense. In order to isolate photons from each other, the space is recursively divided into four identical subspaces by four-space division, and the number of photons contained in each subspace is determined:
If the photon number contained in the subspace is more than 1, continuing to divide the subspace into four-way space until the photon number contained in each subspace obtained after the final division is less than or equal to 1, and stopping dividing the subspace into four-way space;
if the number of photons contained in the subspace is less than or equal to 1, the four-way space division of the subspace is stopped.
2. After the four-fork space division is stopped, the number of photons contained in each subspace is less than or equal to 1, and the obtained tree structure is an isolated four-fork tree. The root node of the isolated quadtree represents the entire space, while the non-leaf nodes represent subspaces of photons, each non-empty leaf node corresponding to each photon. As the noise photons are more dispersed, the noise photons are rapidly segmented into leaf nodes during segmentation. In contrast, signal photons are more closely distributed such that the signal photons are typically divided into leaf nodes by a plurality of divisions.
3. Fig. 3 (a) and 3 (b) show the isolation of signal photons and noise photons, respectively, and fig. 3 (c) shows a corresponding quadtree, where LL, LR, UL and UR represent the corresponding subspaces bottom left, bottom right, top left and top right, respectively. In order to separate the signal photons four divisions are required, whereas the noise photons only have to be two times. In a quadtree, isolation depth is defined as the length of a photon from a root node to a leaf node. Since the isolated quadtree is generated by recursive segmentation, the number of times photons need to be isolated corresponds to the isolation depth. The isolation depth of each original photon in the original photon data is calculated, and the original photons with the isolation depth smaller than the depth threshold are marked as noise photons. Wherein the depth threshold is determined using an oxford algorithm. For each bin of ATL03 data (along a 20m track distance), the variance between categories is calculated as follows:
σ2＝ω1(μ1-μ)2+ω2(μ2-μ)2
Where μ1 represents the average isolation depth of the potential signal photons, ω1 represents the ratio of the potential signal photons to the photon data, μ2 represents the average isolation depth of the potential noise photons, ω2 represents the ratio of the potential noise photons to the photon data, μ represents the average isolation depth of all photons, σ2 represents the variance between the classes, the greater the variance between the signal and noise, and therefore the isolation depth at maximum σ2 is the depth threshold.
And secondly, extracting ground photons from the signal photons by adopting a cloth simulation method.
As shown in fig. 4, if a large enough cloth is tiled over photons, the force will change the shape of the cloth, the final shape being a Digital Surface Model (DSM). Also, overlaying the inverted photons may create DTM. When a wrinkled cloth is simulated, the cloth is represented abstractly as a grid structure consisting of three-dimensional particles and springs between the particles (as shown in fig. 5). These particles are stable in quality and have no shape or size. The movement of these particles is limited in the elevation direction, and following newton's second law, the movement process can be decomposed into external and internal forces, and their positions are determined by both the external and internal forces. At a particular time t, the position of the particle may be calculated as follows:
Wherein X (t) represents the position of the particle at time t, Fext (X, t) represents the external force exerted on the particle, including the gravity force exerted on the particle and the supporting force generated by the contact of the particle with the collided signal photons, Fint (X, t) represents the internal force exerted on the particle, generated by the spring action between the particles, and m represents the mass of the particle.
The process of extracting ground photons by using the cloth simulation method is as follows:
1. the signal photons are flipped upside down.
2. A cloth is set, and all particles included in the cloth are placed over all signal photons, as shown in fig. 6 (a).
3. The cloth simulation process is performed by simulating that each particle in the cloth falls due to an external force applied to the particle, as shown in fig. 6 (b), and when the particle collides with a signal photon, the particle stops moving, and the particle colliding with the signal photon is marked as an immovable particle, as shown in fig. 6 (c), and for the particle that is not marked, the internal force between the particles that it receives is calculated, and the position of the particle that is not marked is adjusted according to the internal force between the particles.
The displacement generated by the movement of the particles under the action of gravity can be calculated by the following formula:
Wherein X (t+. DELTA.t) represents the displacement of the particles under gravity at the moment (t+. DELTA.t), DELTA.t is the time interval, the parameter is set to 0.5m in consideration of the sub-meter level height accuracy of the photon data, X (t) represents the displacement of the particles under gravity at the moment (t+), X (t-DELTA.t) represents the displacement of the particles under gravity at the moment (t-. DELTA.t), and G represents the constant gravity.
The internal force of a particle can be expressed as the tension between adjacent particles. Under internal forces, particles of different heights will attempt to move to the same horizontal plane under tension. When two adjacent particles can move as shown in fig. 7 (a), they move in opposite directions at the same height, and when one of the adjacent particles cannot move as shown in fig. 7 (b), the other particles also move. The distance of movement of the particles is determined by the difference in height between them and the hardness of the cloth. The distance of movement of the particles determines how close the cloth is to the ground. Thus, the hardness RI of the cloth represents the control of the distance travelled by the cloth particles by the relief of the topography. As shown in fig. 4, the harder fabric avoids the inconsistency of ground photons caused by topographical features, while the softer fabric is more suitable for undulating terrain. To obtain a topography-adapted fabric, the photon data is divided into segments of 1000 meters and 100 meters in length. In each small segment, the terrain is determined by the maximum difference in elevation between the signals. When the height difference is more than or equal to 40m, the gradient is steeper, and when the height difference is less than 20m, the terrain is flat. Otherwise, this is a undulating surface. In the long segment, the probability weighted cloth hardness is calculated as follows:
RI=1·Rst+2·Run+3·Rfl
where Rst、Run、Rfl corresponds to the proportion of steep slopes, undulating surfaces and flat terrain in long road sections, respectively. Further, the distance that the particles move (or are called tuning) is:
where d represents the adjusted distance and VD represents the height difference between adjacent particles.
Moreover, residual noise can affect overall performance. As shown in fig. 8, during the simulation, cloth particles are first exposed to noise, so they are marked as immovable particles, resulting in abrupt changes in the local topography. Thus, when a mutation is detected, the cloth is set to break at the changing position. The breaking process traverses the immovable particles and calculates the average height mu and its variance sigma of three adjacent particles. If the elevation angle of the current particle exceeds μ3σ, the cloth will be broken and the current particle will be adjusted so that it can move with the new repositioning.
4. And (3) repeating the cloth simulation process introduced in the step (3) until a set end condition is reached, ending the cloth simulation process, calculating the distances between all signal photons and corresponding particles, and marking the distance smaller than a distance threshold as ground photons so as to realize ground photon extraction.
And thirdly, determining the ground elevation by using a cubic spline curve interpolation method according to the final positions of the particles corresponding to all the signal photons.
Thus, the ground elevation inversion of the photon counting laser radar data is completed. The effectiveness of the method of the present invention will be illustrated in various aspects by applying the method of the present invention to specific examples.
1. Noise photon cancellation.
Fig. 9 (a) -9 (c) show graphs of the results of processing to remove most of the noise photons and extract the signal using the quadtree isolation method under different scenarios. Although the signal-to-noise ratio difference caused by the data acquisition time is the most important factor affecting the denoising effect, the land coverage and the terrain also affect the performance of quadtree isolation. As shown in fig. 9 (a) and 9 (b), some noise photons near buildings and vegetation are misclassified, while some vegetation photons are ignored by the algorithm. In contrast to fig. 9 (a) -9 (c), in undulating terrain, some noise photons near the ground are incorrectly classified as signal photons.
Previous studies have obtained true signal photons through visual inspection due to lack of reference data. Therefore, the experiment also utilizes PhotonLabeler man-machine interaction software to label real signal photons, and quantitatively evaluates the influence of land coverage and topography on algorithm performance. In addition, to make the evaluation process more objective, the original DBSCAN, modified OPERATIONS and DRAGAN algorithms were used in this experiment as comparison algorithms. Algorithm performance is evaluated by precision, probability, recall, probability score, probability parameters:
PPRE＝NTP/(NTP+NFP)
PRE＝NTP/(NTP+NFN)
F=2PPREPRE/(PPRE+PRE)
Where NTP represents the number of correctly extracted signal photons, NFP represents the number of erroneously extracted noise photons, and NFN represents the number of omitted signal photons. Thus, PPRE represents the reliability of the algorithm, PRE represents the completeness of the signal photons extracted by the algorithm, and F represents the overall performance of the algorithm. The comparison results are given in table 1.
Table 1 quantification of five noise cancellation algorithms used by the alaska study area
Table 1 reveals several important findings. By comparing the performance of the algorithm, quadtree isolation achieves the best accuracy under different scenarios and recalls many signal photons (next to DRAGANN). Therefore, quadtree isolation is the best-performing algorithm. The improved DBSCAN and the improved OPTICS achieve similar high precision, but the former performs better than the latter due to its high recall. The original DBSCAN and DRAGANN algorithms perform the worst. Although the accuracy of the original DBSCAN is inferior to quadtree isolation, its recall drops dramatically when the surface is submerged. DRAGANN achieved the highest recall, but this also indicated that many noise photons were falsely marked as signals. The accuracy defect or recall problem of the original DBSCANs and DRAGANN when simulating various scenes.
In terms of land coverage, the algorithm has the lowest accuracy and recall in processing vegetation coverage data. An unexpected result is that the building coverage area reaches the highest recall rate. Although the recall of algorithms in processing lean land data is lower than building coverage data, the overall performance of all algorithms is still best.
When the algorithm processes different terrain data, the result is counterintuitive. Quadtree isolation and improved database scanning perform better on mountain data than on land, while the other three algorithms perform better on land. In modeling hills, all algorithms achieved the worst results.
2. Surface photon extraction and elevation inversion.
Fig. 9 (a) -9 (c) show calculated ground curves. The calculated floor surface visually conforms to the actual floor, indicating that the cloth simulation is effective. Considering that the accuracy of ground photon extraction influences the accuracy of ground elevation inversion, the elevation of the ground photons is compared with the measured elevation of G-LiHT under the corresponding coordinates, and the performance of layout simulation is quantitatively evaluated. In addition, the raw minimum algorithm (OM), the elevation percentile enhanced minimum algorithm (EPEM), and the TIN enhanced minimum algorithm (TEM) are used as the comparison algorithm. The processed signal photons are all isolated and extracted by the quadtree. The Mean Error (ME) and Root Mean Square Error (RMSE) are used to evaluate the performance of the algorithm, the calculation formula is as follows:
Where n is the number of ground photons, ei is the elevation of the ith ground photon, and ri is the corresponding in-situ ground elevation.
As listed in table 2, the maximum likelihood values for the algorithms for different land cover and topography are negative. This is caused by multiphoton scattering of the surface. Multiple scattering returns surface photons wider than the actual surface, so the ground photons extracted by the algorithm are located somewhere below the surface.
Table 2 quantitative results of four ground photon extraction algorithms for alaska study area
For different algorithms, transmission electron microscopy is slightly better than optical microscopy. Although EPEM minimizes error when simulating building coverage and land, ME is greatest when managing data in other scenarios. Of the four algorithms, the RMSE of the transient electromagnetic method is worst. In other scenarios, the cloth simulation achieves the best performance, with significantly higher accuracy than other algorithms when simulating lean land and mountain data.
For different land coverage, the building coverage area is the highest in accuracy, while the barren land is the lowest in accuracy. The algorithm achieves the best performance on flat ground for different terrain data. When processing hilly and mountain data, the accuracy of the algorithm deteriorates as the surface relief increases. The result of the mountain data is the worst, and the result of the mountain data is better than that of the mountain data.
3. Performance of the noise removal algorithm.
Although DRAGANN marks photons with a signal confidence of 1 as signals, resulting in the lowest accuracy, the results also indicate that the other four algorithms are valid.
Consistent with previous studies, the original DBSCAN recalls a small number of signal photons when processing mountain data, because the elliptical filter kernel in the algorithm does not take into account the effects of terrain slope. The improved OPTICS algorithm does not consider the terrain either, but the recall rate of the algorithm is higher than that of the original DBSCAN algorithm when processing hilly and mountain data. In calculating photon density, the improved optical algorithm does not rely on actual photon distances, but rather defines the spatial relationship of photons based on logical reachability distances. Thus, defining logical spatial relationships, rather than measuring specific distances, allows the modified OPTICS algorithm to obtain better results than the original DBSCAN.
The comparison between quadtree isolation and improved database scan also shows the advantage of logical relationships. Unlike the original DBSCAN, the modified DBSCAN adds direction parameters into the filter kernel, so that the kernel can adapt to the terrain gradient, and density statistics closer to actual conditions are obtained. However, due to variations in measurement conditions, such as sun angle, weather, and ground object reflectivity, photon density is inconsistent. Thus, the dependence on density statistics in processing long distance data makes the modified DBSCAN unable to adapt to changes in measurement conditions. For quadtree isolation, the algorithm uses quadtrees to divide photon space, avoiding defining abstract filter kernel coefficients like other algorithms, and the number of quadtree splitting operations required to isolate each photon is only related to the spatial location of the photon, not affected by surface coverage, topography or measurement conditions. Furthermore, the processing object in quadtree isolation is a data bin of 20m along the track length, not long distance data, which avoids the influence of measurement conditions. Although quadtree isolation algorithms have reduced performance in terms of vegetation coverage data, hilly data, and mountain data, this is not because the algorithm cannot adapt to environmental changes, but because the signal-to-noise ratio is reduced when the local surface fluctuates, which also affects the other four algorithms.
4. Anti-intuitive discovery analysis of noise cancellation.
According to two counterintuitive findings described in section 1, "noise photon cancellation," the first unexpected finding is that the recall of the algorithm is higher than that of barren land cover data when processing building cover data. This finding is in contrast to previous studies, which considered that the results of lean land coverage data should be superior to those of land object coverage data due to differences in photon spatial distribution.
Fig. 10 (a) and 10 (b) show the local noise removal results and corresponding images at a track-along distance of 4000-5000m, as shown in fig. 9 (a). The results show that the photons of the building are fully extracted and that noise in the vicinity of the building (especially between buildings) is falsely marked as a signal. The building is a hard reflective target whose photon spatial distribution is similar to the ground photons. Denser distribution is easier to extract building photons, and noise around house photons is more easily falsely marked as vegetation, as low building and vegetation mix distributions are common in alaska.
The data in fig. 11 (a) -11 (e) support this explanation. Fig. 11 (c) -11 (e) show photon data statistics for three land cover types in the elevation direction. There are two peaks on the construction histogram, namely Gao Chengjiao small ground peak and larger elevation construction peak. The photon frequencies of these two peaks are similar, indicating that the spatial distribution of the building photons is as dense as the ground photons. In contrast, there is only one ground peak in the histogram of the vegetation coverage, and the frequency of the vegetation photons is lower than the frequency of the ground photons. Thus, dense distribution will result in higher construction coverage data recall rates.
Furthermore, based on fig. 1, alaska's unhairing land is mainly distributed in grasslands, frozen soil, and snow landscapes in high altitude mountains, where severe surface fluctuations can negatively affect signal photon extraction.
It is unexpected, but reasonable, that the recall of building coverage data is higher than wasteland data given the dense distribution of building photons and the special geographical environment of the experimental zone.
Another unexpected finding is that quadtree isolation and improved database scanning perform better in mountainous areas than in flat areas. In general, signal photons are easier to extract on flat ground than on steep slopes. However, as shown in fig. 1, there is a strong correlation between the land coverage and the terrain of alaska. Flat land typically covers buildings and vegetation, where vegetation is the primary type of coverage and mountainous areas are barren. Land coverage on flat land tends to be complex, while mountain land is often barren. This unexpected finding can be explained by this correlation between land coverage and topography.
As described in section 3 "Performance of noise removal algorithm," quadtree isolation and modified database scanning are highly adaptable to terrain gradients, and therefore they may perform better when dealing with barren mountains. In contrast, vegetation and some building areas are the primary type of surface coverage on flat ground. As shown in fig. 10 (a) and 11 (a), the spatial distribution of vegetation photons is relatively loose, and noise around vegetation is difficult to remove. In contrast to fig. 11 (c) -11 (e), not only are the vegetation photons lower in frequency than land and construction photons, but the vegetation also reduces the transition slope from the ground peak to the air, and furthermore, the boundary between the signal photons and noise photons becomes blurred. Even if the floor covering is a building, as described above, dense distribution of building photons will result in erroneously identifying noise between the buildings as a signal.
Based on table 1, quadtree isolation and improved database scanning still achieve better performance when processing flat ground data. Further analysis showed that the F values for the quadtree isolation method and the modified database scan method were 5.69% and 4.43% lower than the building coverage data, respectively. The effect of complex land coverage is more pronounced for the original DBSCAN, modified OPTICS and DRAGANN, as their F values drop by 8.29, 6.22 and 7.03%, respectively.
In summary, the quadtree isolation and the improved DBSCAN achieve better performance in processing mountain data, mainly because alaska mountain is mostly barren and land coverage is simple, and the two algorithms have good azimuth adaptability. While they are subject to complex land coverage when processing flat land data, this does not indicate quadtree isolation and modified database scanning is not effective. In contrast, quadtree isolation and improved DBSCAN are superior to the other three algorithms in terms of overall performance and relative degradation rate.
5. Performance of the terrestrial photon extraction algorithm.
Similar to previous studies, ground coverage, terrain slope, and residual near-ground noise all affect algorithm performance. For land coverage, the accuracy of the building coverage area is better than that of the vegetation coverage area. As shown in fig. 1, vegetation is distributed on flat lands and hills, while buildings are distributed only on flat lands. One possible explanation is that the terrain slope has a greater impact on the algorithm, based on the "anti-intuitive discovery analysis of noise cancellation" section 4. Although land coverage has a negative impact on algorithm performance, the impact of terrain slope on algorithm performance is more pronounced. The results of the barren land confirm this observation. The barren land of the research area has strong correlation with the mountain area. The results of the barren land are the worst in accuracy, although there are no effects of land objects.
In addition, as shown in the study area used and fig. 10 (a), conifer vegetation and low-rise buildings are mostly distributed in the study area without loss of ground photons. This correlation between surface coverage and terrain does not negate the minimum algorithm. In contrast, the elevation percentile enhanced minimization algorithm (EPEM) achieves better results than cloth modeling when processing flat ground data due to the removal of near ground noise and partial photon multiple scattering.
For different terrain results, flat land results are best, while mountain land results are better than hills. This is appreciated because both the terrain slope and vegetation coverage have a negative impact on the algorithm when processing mountain data, and the algorithm is only affected by the terrain slope when processing mountain data.
Since hilly data is affected by vegetation coverage and gradient, the processing results of the four algorithms are analyzed in detail, taking the 1000-2000m data segment in fig. 9 (b) as an example. Fig. 12 (a) -12 (d) show the ground photon extraction results. The hardness of the cloth can be actively adjusted according to the fluctuation of the terrain by cloth simulation, so that the ground photons on the hillside can be accurately extracted. In addition, the cloth simulation can effectively avoid the influence of near ground noise caused by cloth breakage. OM also successfully extracts ground photons, but it erroneously extracts all near ground noise because it uses local minima as ground photons, without discrimination. EPEM assume that near-ground noise is widely present in the signal photons, so the elevation percentile was designed to cancel it. This assumption is unreasonable and increases the error without near-ground noise. When analyzed similarly to the case at 1300m in fig. 12 (c), EPEM removes the true ground photons of the lower layer, resulting in a ground stairway, in other words, further extracted ground photons are within the same height percentage range. The transmission electron microscope extracts more ground photons on the basis of the optical microscope, so that the adaptability to the terrain is enhanced. However, transient electromagnetic methods are also affected by near-surface noise, and thus this method increases RMSE of the extraction results.
Overall, the cloth simulation achieved optimal performance based on cloth hardness adjustment and cloth breakage. EPEM removes low level photons through the elevation percentile to obtain optimal effect on flat ground, but creates ground stairway phenomenon in hills and mountainous areas. Although OM successfully extracts ground photons, near ground noise cannot be avoided. The transmission electron microscope extracts more ground photons through the triangular mesh encryption, so that the terrain adaptability of the algorithm is enhanced. However, it cannot avoid the influence of near-ground noise, which increases RMSE.