技术领域Technical Field
本发明属于计算机视觉技术领域，尤其涉及一种基于焦点转移图网络的跨场景高光谱分类方法。The present invention belongs to the field of computer vision technology, and in particular relates to a cross-scene hyperspectral classification method based on a focus transfer graph network.
背景技术Background technique
近年来，随着遥感成像技术的不断进步，大量采集了高光谱分辨率的高光谱数据。高光谱数据因其丰富的光谱信息而引起了许多领域(医学、农学等)的关注。高光谱图像(HSI)分类技术的目的是将每个像素分为某一类，广泛应用于遥感图像分析。迄今为止，许多机器学习算法已应用于高光谱图像HSI分类，例如最大似然分类器、支持向量机(SVM)等。这些传统方法是通过手动从高光谱图像HSI中提取特征进行训练，往往分类精度较差，并且很大程度上依赖专业知识和经验。受到深度学习算法在自然图像领域取得的成就的启发，研究人员逐渐引入深度学习，并用深度模型的分布式特征表达能力取代传统的手动特征提取方法。In recent years, with the continuous advancement of remote sensing imaging technology, a large amount of hyperspectral data with high spectral resolution has been collected. Hyperspectral data has attracted the attention of many fields (medicine, agronomy, etc.) due to its rich spectral information. The purpose of hyperspectral image (HSI) classification technology is to classify each pixel into a certain category, which is widely used in remote sensing image analysis. To date, many machine learning algorithms have been applied to hyperspectral image HSI classification, such as maximum likelihood classifier, support vector machine (SVM), etc. These traditional methods are trained by manually extracting features from hyperspectral images HSI, which often have poor classification accuracy and rely heavily on professional knowledge and experience. Inspired by the achievements of deep learning algorithms in the field of natural images, researchers have gradually introduced deep learning and replaced traditional manual feature extraction methods with the distributed feature expression capabilities of deep models.
尽管基于卷积神经网络CNN的方法已经显示出显着的结果，但由于卷积核的形状固定，其处理高光谱图像HSI中不规则区域的能力不足。图卷积网络(GCN)可以同时利用高光谱图像HSI的空间和光谱特征来捕获数据的全局空间信息。对于高光谱图像HSI来说，通常存在成本高、标签获取困难等问题。如果能够利用与目标图像相似但不同的标记高光谱图像HSI来完成目标图像的分类，则可以有效解决这些问题。对此，迁移学习是一个很好的解决方案。Although methods based on convolutional neural networks (CNNs) have shown remarkable results, they are not capable of processing irregular regions in hyperspectral images (HSI) due to the fixed shape of the convolution kernel. Graph convolutional networks (GCNs) can simultaneously utilize the spatial and spectral features of hyperspectral images (HSIs) to capture the global spatial information of the data. For hyperspectral images (HSIs), there are usually problems such as high cost and difficulty in obtaining labels. If the target image classification can be completed using a hyperspectral image (HSI) with similar but different labels, transfer learning can be effectively solved. For this, transfer learning is a good solution.
上述基于迁移学习的高光谱图像HSI分类方法一般考虑两个域的边缘分布，但忽略了对齐两个域条件分布的重要性，这限制了其性能的进一步提高。一方面，仅对齐边缘分布会导致不同类别之间的混淆，使模型难以学习准确的分类边界。另一方面，对于跨场景的高光谱图像HSI分类任务，两个域之间经常存在类别不平衡的现象。仅对齐边际分布会使模型过度关注包含大量样本的类，从而导致模型性能受损。The above-mentioned hyperspectral image HSI classification methods based on transfer learning generally consider the marginal distribution of the two domains, but ignore the importance of aligning the conditional distributions of the two domains, which limits the further improvement of their performance. On the one hand, aligning only the marginal distributions will lead to confusion between different categories, making it difficult for the model to learn accurate classification boundaries. On the other hand, for cross-scene hyperspectral image HSI classification tasks, there is often a class imbalance between the two domains. Aligning only the marginal distributions will cause the model to over-focus on the class containing a large number of samples, resulting in impaired model performance.
发明内容Summary of the invention
发明目的：针对上述背景技术中存在的问题，本发明提供了一种基于焦点转移图网络的跨场景高光谱分类方法，通过抑制邻居干扰、聚焦难以分类的样本、剪枝噪声伪标签、对齐规范子域，实现从源域到目标域的跨域知识迁移。Purpose of the invention: In view of the problems existing in the above-mentioned background technology, the present invention provides a cross-scene hyperspectral classification method based on a focus transfer graph network, which realizes cross-domain knowledge transfer from the source domain to the target domain by suppressing neighbor interference, focusing on difficult-to-classify samples, pruning noisy pseudo-labels, and aligning standard subdomains.
发明内容：为实现上述目的，本发明采用的技术方案为：一种基于焦点转移图网络的跨场景高光谱分类方法，包括如下步骤：Invention content: To achieve the above purpose, the technical solution adopted by the present invention is: a cross-scene hyperspectral classification method based on a focus transfer graph network, comprising the following steps:
步骤1，将源域和目标域的原始高光谱图像输入图采样和聚合GraghSAGE，得到节点的加权嵌入表示，再基于NWS获得空间谱特征。Step 1: The original hyperspectral images of the source and target domains are input into the graph sampling and aggregation GraghSAGE to obtain the weighted embedding representation of the nodes, and then the spatial spectrum features are obtained based on NWS.
步骤2，根据目标域空间谱特征预测伪标签，并使用伪标签修剪策略修剪掉噪声伪标签，从而校正目标域的条件分布。In step 2, pseudo labels are predicted based on the spatial spectral features of the target domain, and the noisy pseudo labels are pruned using a pseudo label pruning strategy to correct the conditional distribution of the target domain.
步骤3，对齐源域和目标域之间的同类子域分布并规范子域对齐；Step 3: Align the distribution of similar subdomains between the source domain and the target domain and standardize the subdomain alignment;
步骤4，基于规范子域对齐后的两域数据，分别计算两域预测概率向量，将两域预测概率向量相加后生成最终输出。Step 4: Based on the data of the two domains after the canonical subdomain alignment, the prediction probability vectors of the two domains are calculated respectively, and the final output is generated by adding the prediction probability vectors of the two domains.
进一步的，步骤1中将源域和目标域的原始高光谱图像输入图采样和聚合GraghSAGE，得到节点的加权嵌入表示，具体包括如下步骤：Furthermore, in step 1, the original hyperspectral images of the source domain and the target domain are input to the graph sampling and aggregation GraghSAGE to obtain the weighted embedding representation of the nodes, which specifically includes the following steps:
首先，将源域和目标域的原始高光谱图像输入图采样和聚合GraghSAGE，训练一组聚合器函数，学习聚合来自节点本地邻居的特征信息，从而结合空间光谱信息来充分表征样本。然后，将聚合后的本地邻居特征信息与其自身特征连接后，进行非线性映射，最终可以得到节点的加权嵌入表示。First, the original hyperspectral images of the source and target domains are input into the graph sampling and aggregation GraghSAGE, and a set of aggregator functions are trained to learn to aggregate feature information from the local neighbors of the node, thereby combining spatial spectral information to fully characterize the sample. Then, the aggregated local neighbor feature information is connected with its own features, and nonlinear mapping is performed to finally obtain the weighted embedding representation of the node.
进一步的，步骤1中所述基于NWS获得空间谱特征，具体包括如下步骤：设计了基于空间光谱信息的NWS，考虑到空间先验，通过计算节点之间的空间距离来分配聚合邻居节点的重要性权重，考虑到高阶邻居节点一般距离中心节点较远，降低高阶邻居节点的注意力以减轻异构节点的干扰。根据空间距离，为邻居节点分配不同的重要性权重，以提取有区别的空间频谱信息，抑制异构节点信息的潜在干扰。Furthermore, the spatial spectrum features obtained based on NWS described in step 1 specifically include the following steps: NWS based on spatial spectrum information is designed, and the importance weights of aggregated neighbor nodes are assigned by calculating the spatial distance between nodes, considering that high-order neighbor nodes are generally far away from the central node, and the attention of high-order neighbor nodes is reduced to reduce the interference of heterogeneous nodes. Different importance weights are assigned to neighbor nodes according to the spatial distance to extract differentiated spatial spectrum information and suppress the potential interference of heterogeneous node information.
进一步的，步骤2提出了一种伪标签修剪策略，从而修剪目标域中远离源域中同类样本的伪标签噪声。具体包括以下步骤：Furthermore, step 2 proposes a pseudo-label pruning strategy to prune the pseudo-label noise in the target domain that is far away from similar samples in the source domain. Specifically, it includes the following steps:
首先，计算源域每个类的原型：First, calculate the prototype of each class in the source domain:
其中，表示目标域第i类第j个样本，/>表示源域第i类的类别表示，n表示第i类的样本个数。in, represents the jth sample of the i-th category in the target domain,/> represents the category representation of the i-th category in the source domain, and n represents the number of samples in the i-th category.
然后，分别计算出目标样本到源域对应类别表示的距离dts，以及到所有类别表示距离的均值并认为/>为样本到源域距离：Then, the distancedts from the target sample to the corresponding category representation in the source domain and the mean distance to all category representations are calculated respectively. And think/> The distance from the sample to the source domain:
最后，修剪目标域中dts大于样本到源域距离的噪声伪标签，得到可靠的样本进行子域适应。Finally, prune the target domain so thatdts is larger than the distance between the sample and the source domain. The noisy pseudo labels are used to obtain reliable samples for sub-domain adaptation.
进一步的，步骤3中设计了规范子域适应，通过对齐源域和目标域之间的同类子域分布，从而能够更精细地对齐其条件分布。此外，规范子域适应(SSA)对子域施加了特征规范约束，为域适应提供了区分性特征。一方面，鼓励同一类的样本具有相似的特征表示，抑制与类无关的噪声。另一方面，促进了特征空间中不同类别样本的分离，突出了有助于区分类别的关键信息。具体包括以下步骤：Furthermore, in step 3, canonical subdomain adaptation is designed to align the distribution of similar subdomains between the source domain and the target domain, so that their conditional distributions can be more finely aligned. In addition, canonical subdomain adaptation (SSA) imposes feature norm constraints on subdomains, providing discriminative features for domain adaptation. On the one hand, it encourages samples of the same class to have similar feature representations and suppresses class-independent noise. On the other hand, it promotes the separation of samples of different categories in the feature space and highlights the key information that helps to distinguish categories. Specifically, it includes the following steps:
首先，为了使子域的簇在类内更紧凑，在类间更远，提出规范因子为：First, in order to make the clusters of subdomains more compact within a class and farther between classes, the normalization factor is proposed as:
其中，β1,β2用于调整这个规范因子的权重。分别表示源域和目标域同类特征之间的距离，/>分别表示源域和目标域不同类别特征之间的距离。Among them, β1 and β2 are used to adjust the weight of this normative factor. Respectively represent the distance between similar features in the source domain and the target domain,/> They represent the distances between features of different categories in the source domain and the target domain respectively.
然后，为了进一步捕获两个域的细粒度信息并充分利用带有伪标签的源域数据和目标域数据，使用SSA来对齐两个域。基于规范子域适应模块对两域进行领域对齐，则规范后的子域适应损失可表示为：Then, in order to further capture the fine-grained information of the two domains and make full use of the source domain data and target domain data with pseudo labels, SSA is used to align the two domains. The two domains are aligned based on the canonical subdomain adaptation module, and the subdomain adaptation loss after standardization can be expressed as:
其中，Ps和Pt分别表示源域和目标域的条件概率分布。Among them,Ps andPt represent the conditional probability distribution of the source domain and the target domain respectively.
进一步的，步骤4中提出了分类器自适应模块来学习两个域分类器的差异信息，以进一步提高分类性能。分类器适应模块由全连接层组成，用于拟合残差函数。分类器的最终输出是通过多步骤过程获得的。首先，计算目标域预测概率向量，然后，通过将源域预测概率向量和目标域预测概率向量相加来生成最终输出。Furthermore, a classifier adaptation module is proposed in step 4 to learn the difference information of the two domain classifiers to further improve the classification performance. The classifier adaptation module consists of a fully connected layer for fitting the residual function. The final output of the classifier is obtained through a multi-step process. First, the target domain prediction probability vector is calculated, and then the final output is generated by adding the source domain prediction probability vector and the target domain prediction probability vector.
进一步的，步骤5中使用了focal loss，这有利于解决数据不平衡带来的模型性能问题，使网络更加专注于难以分类的样本。Furthermore, focal loss is used in step 5, which helps to solve the model performance problem caused by data imbalance and makes the network more focused on samples that are difficult to classify.
有益效果：提出了一种基于焦点转移图网络的跨场景高光谱分类方法，通过抑制邻居干扰、聚焦难以分类的样本、剪枝噪声伪标签、对齐规范子域，实现从源域到目标域的跨域知识迁移。提出了一种基于空间谱信息的邻居加权策略，用于减少节点信息聚合过程中的邻居信息干扰，捕获目标域的条件分布。提出了伪标签噪声修剪策略。基于类度量减少噪声伪标签对传输过程的影响，从而实现目标域条件分布校正。提出了规范子域自适应模块，在子域自适应过程中通过减小类内样本的特征距离并拉宽类间样本的特征距离来完成有效的条件分布对齐。Beneficial effects: A cross-scene hyperspectral classification method based on a focus transfer graph network is proposed. By suppressing neighbor interference, focusing on difficult-to-classify samples, pruning noisy pseudo-labels, and aligning canonical subdomains, cross-domain knowledge transfer from the source domain to the target domain is achieved. A neighbor weighting strategy based on spatial spectrum information is proposed to reduce neighbor information interference in the process of node information aggregation and capture the conditional distribution of the target domain. A pseudo-label noise pruning strategy is proposed. Based on class metrics, the influence of noisy pseudo-labels on the transmission process is reduced, thereby achieving target domain conditional distribution correction. A canonical subdomain adaptation module is proposed. In the subdomain adaptation process, effective conditional distribution alignment is completed by reducing the feature distance of samples within a class and widening the feature distance of samples between classes.
附图说明BRIEF DESCRIPTION OF THE DRAWINGS
图1是本发明方法原理框图。FIG. 1 is a block diagram showing the principle of the method of the present invention.
具体实施方式Detailed ways
下面结合附图对本发明作更进一步的说明。显然，所描述的实施例是本发明一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。The present invention is further described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
本发明提供的基于焦点转移图网络的跨场景高光谱分类方法，具体原理如图1所示，将两个域的原始高光谱图像输入图采样和聚合(GraghSAGE)，基于NWS获得包含空间光谱信息的特征。接着，根据目标域空间谱特征预测伪标签，并使用伪标签修剪策略修剪掉噪声伪标签，从而校正目标域的条件分布。之后，两个域的条件分布与规范子域对齐完全对齐，获得域不变特征。然后，通过分类器自适应来适应两个域分类器之间的差异，并基于域不变特征输出样本的类别预测概率。最后，根据类别预测概率和标签计算分类损失，通过关注难以分类的样本来提高模型的分类能力。The cross-scene hyperspectral classification method based on the focus transfer graph network provided by the present invention has a specific principle as shown in Figure 1. The original hyperspectral image input graphs of the two domains are sampled and aggregated (GraghSAGE), and features containing spatial spectral information are obtained based on NWS. Next, pseudo labels are predicted based on the spatial spectrum features of the target domain, and the noisy pseudo labels are pruned using a pseudo label pruning strategy to correct the conditional distribution of the target domain. After that, the conditional distributions of the two domains are completely aligned with the canonical subdomain alignment to obtain domain-invariant features. Then, the differences between the two domain classifiers are adapted through classifier adaptation, and the class prediction probability of the sample is output based on the domain-invariant features. Finally, the classification loss is calculated based on the class prediction probability and the label, and the classification ability of the model is improved by focusing on samples that are difficult to classify.
步骤1，将源域和目标域的原始高光谱图像输入图采样和聚合GraghSAGE，进行特征学习，得到节点的加权嵌入表示，基于邻接权重聚合模块NWS获得空间谱特征；Step 1: Input the original hyperspectral images of the source domain and the target domain into the graph sampling and aggregation GraghSAGE to perform feature learning, obtain the weighted embedding representation of the nodes, and obtain the spatial spectrum features based on the adjacent weight aggregation module NWS;
首先，将图采样和聚合GraghSAGE用作本发明的转移图网络FTGN的特征嵌入，训练一组聚合器函数，学习聚合来自节点本地邻居的特征信息，从而结合空间光谱信息来充分表征样本。然后，将聚合后的本地邻居特征信息与其自身特征连接后，进行非线性映射，最终得到节点的加权嵌入表示。First, graph sampling and aggregation GraghSAGE is used as feature embedding for the transfer graph network FTGN of the present invention, and a set of aggregator functions are trained to learn to aggregate feature information from the local neighbors of the node, thereby combining spatial spectral information to fully characterize the sample. Then, the aggregated local neighbor feature information is connected with its own features, and nonlinear mapping is performed to finally obtain a weighted embedding representation of the node.
然后，设计了基于空间光谱信息的NWS，考虑到空间先验，通过计算节点之间的空间距离来分配聚合邻居节点的重要性权重，考虑到高阶邻居节点一般距离中心节点较远，降低高阶邻居节点的注意力以减轻异构节点的干扰。根据空间距离，为邻居节点分配不同的重要性权重，以提取有区别的空间频谱信息，抑制异构节点信息的潜在干扰。Then, a NWS based on spatial spectral information is designed. Taking into account the spatial prior, the importance weights of aggregated neighbor nodes are assigned by calculating the spatial distance between nodes. Considering that high-order neighbor nodes are generally far away from the central node, the attention of high-order neighbor nodes is reduced to alleviate the interference of heterogeneous nodes. Different importance weights are assigned to neighbor nodes according to the spatial distance to extract differentiated spatial spectral information and suppress the potential interference of heterogeneous node information.
步骤2，根据目标域空间谱特征预测伪标签，并使用伪标签修剪策略修剪掉噪声伪标签，从而校正目标域的条件分布；Step 2: Predict pseudo labels based on the spatial spectrum features of the target domain, and use the pseudo label pruning strategy to prune away the noisy pseudo labels to correct the conditional distribution of the target domain;
所述伪标签修剪策略，包括如下步骤：The pseudo-label pruning strategy includes the following steps:
首先，计算源域每个类的原型：First, calculate the prototype of each class in the source domain:
其中，表示目标域第i类第j个样本，/>表示源域第i类的类别表示，n表示第i类的样本个数。in, represents the jth sample of the i-th category in the target domain,/> represents the category representation of the i-th category in the source domain, and n represents the number of samples in the i-th category.
然后，分别计算出目标样本到源域对应类别表示的距离dts，以及到所有类别表示距离的均值并认为/>为样本到源域距离：Then, the distancedts from the target sample to the corresponding category representation in the source domain and the mean distance to all category representations are calculated respectively. And think/> The distance from the sample to the source domain:
最后，修剪目标域中dts大于样本到源域距离的噪声伪标签，得到可靠的样本进行子域适应。Finally, prune the target domain so thatdts is larger than the distance between the sample and the source domain. The noisy pseudo labels are used to obtain reliable samples for sub-domain adaptation.
步骤3，规范子域对齐；Step 3, standardize subdomain alignment;
首先，为了使子域的簇在类内更紧凑，在类间更远，提出规范因子为：First, in order to make the clusters of subdomains more compact within a class and farther between classes, the normalization factor is proposed as:
其中，β1,β2用于调整规范因子的权重。分别表示源域和目标域同类特征之间的距离，/>分别表示源域和目标域不同类别特征之间的距离。然后，使用源域样本训练源域分类器，基于训练好的源域分类器对目标域样本进行分类得到目标域伪标签，为了进一步捕获两个域的细粒度信息并充分利用带有标签的源域数据和带有伪标签目标域数据，使用规范子域适应SSA来对齐两个域。Among them, β1 and β2 are used to adjust the weights of the normative factors. Respectively represent the distance between similar features in the source domain and the target domain,/> Respectively represent the distance between different categories of features in the source domain and the target domain. Then, the source domain classifier is trained using source domain samples, and the target domain samples are classified based on the trained source domain classifier to obtain the target domain pseudo-label. In order to further capture the fine-grained information of the two domains and make full use of the labeled source domain data and the pseudo-labeled target domain data, the canonical subdomain adaptation SSA is used to align the two domains.
基于规范子域适应模块对两域进行领域对齐，则规范后的子域适应损失可表示为：Based on the canonical subdomain adaptation module, the two domains are aligned, and the subdomain adaptation loss after standardization can be expressed as:
其中，Ps和Pt分别表示源域和目标域的条件概率分布，表示最大均值化差异，El表示数学期望，/>表示模型参数；Among them,Ps andPt represent the conditional probability distribution of the source domain and the target domain respectively. represents the maximum mean difference,El represents the mathematical expectation, /> represents the model parameters;
步骤4，通过分类器自适应来适应两个域分类器之间的差异，并基于域不变特征输出样本的类别预测概率；具体为：Step 4: Adapt the difference between the two domain classifiers through classifier adaptation, and output the class prediction probability of the sample based on the domain invariant features; specifically:
首先，计算目标域预测概率向量，然后，通过将源域预测概率向量和目标域预测概率向量向量相加来生成最终输出。First, the target domain prediction probability vector is calculated, and then the final output is generated by adding the source domain prediction probability vector and the target domain prediction probability vector.
分类器适应模块由全连接层组成，用于拟合残差函数。分类器适应模块的最终输出是通过多步骤过程获得的。The classifier adaptation module consists of fully connected layers to fit the residual function. The final output of the classifier adaptation module is obtained through a multi-step process.
步骤5，焦点类别失衡。Step 5: Focus category imbalance.
使用损失函数focal loss，根据类别预测概率和标签计算分类损失，通过关注难以分类的样本来提高分类能力，这有利于解决数据不平衡带来的模型性能问题，使网络更加专注于难以分类的样本。Using the loss function focal loss, the classification loss is calculated according to the category prediction probability and label. The classification ability is improved by focusing on samples that are difficult to classify. This is conducive to solving the model performance problems caused by data imbalance and making the network more focused on samples that are difficult to classify.
下面结合仿真实验对本发明的效果做进一步的说明：The effect of the present invention is further described below in conjunction with simulation experiments:
所有实验均在硬件配置为CPU i7-12700k、GPU Nvidia GeForce GTX 1080Ti和32GB内存的工作站上进行。All experiments were conducted on a workstation with a CPU i7-12700k, a GPU Nvidia GeForce GTX 1080Ti, and 32GB of memory.
选择了五个真实的高光谱图像HSI数据集，即Botswana、KSC、B-S、Pavia和Houston。其中，Botswana由BOT5子数据集、BOT6子数据集和BOT7子数据集组成，均为1476×256×242数据类型的高光谱图像HSI。KSC数据集由来自不同地区的KSC1子数据集和KSC3子数据集组成，它们是512×614×176数据类型的现实高光谱图像HSI。B-S来源于休斯顿地区，由两个子数据集Non-shadow和Shadow组成，数据类型分别为349×1318×144和349×587×144。Pavia数据集由Pavia University(PU)和Pavia Center(PC)组成，数据类型分别为610×340×115和1096×492×115。Houston数据集包括Houston 2013和Houston 2018，数据类型分别为349×1905×144和2384×601×48。Five real hyperspectral image HSI datasets were selected, namely Botswana, KSC, B-S, Pavia and Houston. Among them, Botswana consists of BOT5 sub-dataset, BOT6 sub-dataset and BOT7 sub-dataset, all of which are hyperspectral image HSI with 1476×256×242 data type. The KSC dataset consists of KSC1 sub-dataset and KSC3 sub-dataset from different regions, which are real hyperspectral image HSI with 512×614×176 data type. B-S comes from the Houston area and consists of two sub-datasets, Non-shadow and Shadow, with data types of 349×1318×144 and 349×587×144 respectively. The Pavia dataset consists of Pavia University (PU) and Pavia Center (PC), with data types of 610×340×115 and 1096×492×115 respectively. The Houston dataset includes Houston 2013 and Houston 2018, with data types of 349×1905×144 and 2384×601×48 respectively.
为了说明发明的有效性和优越性，选择了11个分类器进行比较：传统分类方法支持向量机SVM；深度学习分类方法CNN；图神经网络GraphSAGE；深度传输方法DANN、DAN、DCORAL、MSTN、M RECON、ADAN、AGAN、CWDAN。In order to illustrate the effectiveness and superiority of the invention, 11 classifiers were selected for comparison: traditional classification method support vector machine SVM; deep learning classification method CNN; graph neural network GraphSAGE; deep transfer methods DANN, DAN, DCORAL, MSTN, M RECON, ADAN, AGAN, and CWDAN.
表1：七个数据对不同方法的OATable 1: Seven data on OA of different methods
从表1中可得：From Table 1, we can get:
1)与传统分类方法相比，深度学习分类方法和图神经网络具有更高的准确率。这是因为深度学习分类方法可以提取数据的深层、高层特征，而图神经网络可以充分利用高光谱图像HSI的空间位置关系。1) Compared with traditional classification methods, deep learning classification methods and graph neural networks have higher accuracy. This is because deep learning classification methods can extract deep and high-level features of data, while graph neural networks can make full use of the spatial position relationship of hyperspectral images HSI.
2)深度迁移方法的精度超过了卷积神经网络CNN，因为它除了提取数据的深层特征之外，还从各个角度减少了两个域之间的分布差异。2) The accuracy of the deep transfer method exceeds that of the convolutional neural network (CNN) because in addition to extracting deep features of the data, it also reduces the distribution difference between the two domains from all angles.
3)FTGN实现了所有数据集中最高的分类精度，并且比深度学习方法更快。主要原因包括以下几点：首先，FTGN充分利用了高光谱图像HSI的上下文信息，无需所有节点参与组成，节省了大量的计算能力。其次，基于空间频谱信息的NWS可以减少聚合周围节点信息时过度的节点信息混合带来的干扰。第三，子域自适应可以对齐两个域的每一类，充分减少两个域之间的域差异。最后，基于类度量的伪标签修剪策略可以减少噪声伪标签的不利影响。3) FTGN achieves the highest classification accuracy among all datasets and is faster than deep learning methods. The main reasons include the following: First, FTGN makes full use of the contextual information of the hyperspectral image HSI, without the participation of all nodes in the composition, saving a lot of computing power. Second, NWS based on spatial spectrum information can reduce the interference caused by excessive node information mixing when aggregating surrounding node information. Third, subdomain adaptation can align each class of the two domains and fully reduce the domain difference between the two domains. Finally, the pseudo-label pruning strategy based on class metrics can reduce the adverse effects of noisy pseudo-labels.
本发明提出了一种迁移学习方法FTGN，并将其用于跨场景高光谱图像HSI分类。一方面，FTGN通过根据空间信息为邻居节点分配不同的权重来减轻空间光谱特征的提取。另一方面，FTGN通过关注难以分类的样本、修剪噪声伪标签、对齐规范子域、适应两个域分类器来实现跨域知识转移。在未来的工作中，我们将进一步探索FTGN在多源情况和源域和目标域类别不一致的异构域场景中适应的可能性。This paper proposes a transfer learning method FTGN and uses it for cross-scene hyperspectral image HSI classification. On the one hand, FTGN alleviates the extraction of spatial spectral features by assigning different weights to neighbor nodes according to spatial information. On the other hand, FTGN achieves cross-domain knowledge transfer by focusing on difficult-to-classify samples, pruning noisy pseudo-labels, aligning normative subdomains, and adapting two domain classifiers. In future work, we will further explore the possibility of FTGN to adapt in multi-source situations and heterogeneous domain scenarios where the source and target domain categories are inconsistent.
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410154382.7ACN117994581A (en) | 2024-02-04 | 2024-02-04 | Cross-scene hyperspectral classification method based on focus transfer graph network |
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410154382.7ACN117994581A (en) | 2024-02-04 | 2024-02-04 | Cross-scene hyperspectral classification method based on focus transfer graph network |
| Publication Number | Publication Date |
|---|---|
| CN117994581Atrue CN117994581A (en) | 2024-05-07 |
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410154382.7APendingCN117994581A (en) | 2024-02-04 | 2024-02-04 | Cross-scene hyperspectral classification method based on focus transfer graph network |
| Country | Link |
|---|---|
| CN (1) | CN117994581A (en) |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170076438A1 (en)* | 2015-08-31 | 2017-03-16 | Cape Analytics, Inc. | Systems and methods for analyzing remote sensing imagery |
| CN115375951A (en)* | 2022-09-20 | 2022-11-22 | 中国矿业大学 | A small-sample hyperspectral image classification method based on primitive transfer network |
| CN117011714A (en)* | 2023-08-10 | 2023-11-07 | 西安电子科技大学 | Hyperspectral image classification method based on pseudo tag assistance |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170076438A1 (en)* | 2015-08-31 | 2017-03-16 | Cape Analytics, Inc. | Systems and methods for analyzing remote sensing imagery |
| CN115375951A (en)* | 2022-09-20 | 2022-11-22 | 中国矿业大学 | A small-sample hyperspectral image classification method based on primitive transfer network |
| CN117011714A (en)* | 2023-08-10 | 2023-11-07 | 西安电子科技大学 | Hyperspectral image classification method based on pseudo tag assistance |
| Title |
|---|
| HAOYU WANG等: "Focal Transfer Graph Network and Its Application in Cross-Scene Hyperspectral Image Classification", IEEE TRANSACTIONS ON ARTIFICIAL INTELLIGENCE（EARLY ACCESS）, 23 January 2024 (2024-01-23), pages 1 - 13* |
| Publication | Publication Date | Title |
|---|---|---|
| CN113408605B (en) | Hyperspectral image semi-supervised classification method based on small sample learning | |
| CN109949317A (en) | A Stepwise Adversarial Learning-Based Approach for Semi-Supervised Image Instance Segmentation | |
| CN105069774B (en) | The Target Segmentation method of optimization is cut based on multi-instance learning and figure | |
| CN101315663A (en) | A Natural Scene Image Classification Method Based on Regional Latent Semantic Features | |
| CN107045640A (en) | A kind of method kept based on neighborhood with kernel space alignment for image recognition | |
| CN115272688B (en) | A small sample learning image classification method based on meta-features | |
| CN116448019B (en) | Intelligent detection device and method for building energy-saving project quality flatness | |
| CN114510594A (en) | Traditional pattern subgraph retrieval method based on self-attention mechanism | |
| CN116486251A (en) | Hyperspectral image classification method based on multi-mode fusion | |
| CN114677333B (en) | A histogram-based image contrast enhancement detection method | |
| CN108664968B (en) | An Unsupervised Text Localization Method Based on Text Selection Model | |
| CN116703932A (en) | CBAM-HRNet model wheat spike grain segmentation and counting method based on convolution attention mechanism | |
| CN106991049A (en) | A kind of Software Defects Predict Methods and forecasting system | |
| Cui et al. | Real-time detection of wood defects based on SPP-improved YOLO algorithm | |
| CN106650820A (en) | Matching recognition method of handwritten electrical component symbols and standard electrical component symbols | |
| Wang et al. | Image target recognition based on improved convolutional neural network | |
| CN110008996A (en) | A kind of isomery domain adaptive method calculated based on divergence | |
| CN104933410B (en) | A kind of high spectrum image spectral domain and spatial domain joint classification method | |
| CN109165679B (en) | A data processing method and device | |
| CN111797795A (en) | A pedestrian detection algorithm based on YOLOv3 and SSR | |
| CN117994581A (en) | Cross-scene hyperspectral classification method based on focus transfer graph network | |
| CN111444762A (en) | Cancer cell identification method based on adaptive resolution | |
| Zhang et al. | An efficient and real-time steel surface defect detection method based on single-stage detection algorithm | |
| CN102496027A (en) | Semi-supervised image classification method based on constrained adaptive transmission | |
| CN117115675A (en) | Cross-time-phase light-weight spatial spectrum feature fusion hyperspectral change detection method, system, equipment and medium |
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | Application publication date:20240507 |