Optical Remote Sensing Image Change Detection Method Based on Adaptive Fusion of NestedUNet

Abstract

This invention discloses a change detection method for optical remote sensing images based on adaptive fusion NestedUNet, comprising the following steps: First, the optical remote sensing images are preprocessed to obtain remote sensing images with consistent data distribution; then, the updated parts of the preprocessed dual-temporal optical remote sensing images are labeled to obtain corresponding change label maps; then, the remote sensing images and corresponding label maps are segmented to obtain training samples; then, the dual-temporal images are concatenated and input into the AFNUNet network to enhance the extraction of overall features of the changed areas and the identification of edge details of the changed targets; after training, all parameter information of the model is saved; finally, the preprocessed test samples are input into the change detection model, and the detection result map is output after calculation, resulting in a more complete overall change map with more refined details.

Description

Technical Field

[0001] This invention relates to the field of optical remote sensing image change detection, and in particular to an optical remote sensing image change detection method based on adaptive fusion of NestedUNet. Background Technology

[0002] Remote sensing image change detection involves observing different regions of a specific area at different times to determine whether the target objects have changed, and to obtain information on changes in their location, shape, quantity, and other attributes. This has significant scientific and applied value and has attracted widespread attention in related fields. Traditional change detection methods can be divided into two categories based on the research object: pixel-based change detection methods and object-based change detection methods. Pixel-based change detection methods generally generate difference maps by directly comparing the spectral or texture information of pixels, and obtain the final change map through threshold segmentation or clustering; examples include change vector analysis based on image algebra, image classification methods, principal component analysis based on image transformation, multivariate change detection methods, and their improved versions such as iterative weighted multivariate change detection methods. Although pixel-based change detection methods are relatively easy to implement, they ignore spatial context information, which can lead to a large amount of salt-and-pepper noise during processing. Object-based change detection methods divide remote sensing images into disjoint objects and analyze the differences between images using the rich spectral, texture, structural, and geometric information in the images. Although this method utilizes the spatial feature information of remote sensing images, the extraction method is relatively complex and has poor robustness.

[0003] In recent years, deep learning-based change detection methods have been continuously developing and have achieved remarkable results. Deep learning-based change detection methods blur the lines between traditional pixel-based and object-based methods because they predict pixel classification maps and highly semantically abstract spatial context from raw images. Compared to traditional pixel- and object-based change detection methods, deep learning-based change detection methods do not require image preprocessing, which not only reduces human intervention and avoids errors caused by preprocessing but also reduces post-processing workload. Due to their non-linear characteristics and excellent feature extraction capabilities, deep learning-based methods can better understand complex scenes, and their performance far surpasses that of traditional methods. Furthermore, satellite remote sensing technology has developed rapidly in recent years, with an increasing number of optical sensors designed for Earth observation. The growing abundance of optical remote sensing images provides strong support for various remote sensing applications. Because of these advantages, the use of deep learning-based optical remote sensing image change detection methods in solving remote sensing image change detection problems has grown exponentially.

[0004] Deep learning-based change detection algorithms for optical remote sensing images have achieved good performance. However, most algorithms directly utilize feature information extracted from deep networks for change detection without considering integration with shallow networks. Deep networks have strong semantic information representation capabilities, but edge details and small objects are easily lost with repeated upsampling and downsampling. Shallow networks have strong detail representation capabilities, but weak semantic information representation capabilities. Therefore, these algorithms often fail to extract complete overall features of changed areas and lack attention to edge details of changed targets. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive fusion method for detecting changes in optical remote sensing images using NestedUNet, which can accurately and effectively identify the subtle differences between two-phase optical remote sensing images.

[0006] The technical solution adopted in this invention is as follows:

[0007] A. Perform orthorectification, image registration, image stretching, and image numerical normalization preprocessing on the dual-temporal optical remote sensing images in sequence to obtain remote sensing images with consistent data distribution.

[0008] B. Label the updated parts of the preprocessed dual-temporal optical remote sensing image obtained in the previous step to obtain the corresponding change label map.

[0009] C. Cut the label image obtained from B and the preprocessed dual-temporal optical remote sensing image from A into the same size to obtain training samples.

[0010] D. Concatenate the two-time optical remote sensing images in the training samples and input them into AFNUNet;

[0011] E. Extract more feature maps related to changing features through a four-layer residual block encoder based on effective channel attention;

[0012] F. Combine the feature maps obtained from the previous step to obtain the final feature maps F1, F2, and F3 containing different semantic information;

[0013] G. Input the final feature maps F1, F2, and F3 obtained in F into the adaptive fusion module to obtain more detailed information about the changed regions and obtain the predicted change map;

[0014] H. Introduce a loss based on Brectis distance and combine it with the binary cross-entropy loss to form a hybrid loss function to calculate the loss between the predicted change map and the corresponding label map of the bi-temporal image obtained in G.

[0015] I. After training is complete, save all parameter information of the trained model;

[0016] J. After preprocessing the preceding and following time-phase remote sensing images to be detected by orthorectification, image registration, image stretching, and image numerical normalization, the images are cut into samples of the same size to obtain the samples to be detected.

[0017] K. Input the sample to be tested into the change detection model obtained in I, and calculate and output the detection result map.

[0018] This invention focuses on change detection in optical remote sensing images. Addressing the shortcomings of existing change detection methods, such as incomplete extraction of changed regions and irregular edges of changed targets, this invention proposes a novel change detection method that improves the extraction of overall features of changed regions and the recognition of detailed features. Specifically, this invention combines NestedUNet with an adaptive fusion module. Through dense feature transfer between the encoder and decoder, and between decoders, it enhances the extraction of overall features, resulting in more complete extracted changed regions. Finally, a hybrid loss method based on Blectis distance loss and binary cross-entropy loss is used to more accurately identify differences between two temporal images, thereby capturing more edge details. Compared with traditional algorithms, this invention enhances the extraction of overall features of changed regions and the recognition of detailed edges, effectively solving the problem of insufficient edge recognition capability of other change detection methods. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the process of the present invention.

[0021] Figure 2 This is the AFNUNet backbone network diagram of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] like Figure 1 As shown, the present invention includes the following steps:

[0024] A. Perform orthorectification, image registration, image stretching, and image numerical normalization preprocessing on the dual-temporal optical remote sensing images in sequence to obtain remote sensing images with consistent data distribution.

[0025] B. Label the updated parts of the preprocessed dual-temporal optical remote sensing images obtained in the previous step (mainly including vegetation changes, new urban buildings, suburban expansion, foundations before construction, and road expansion, etc.) to obtain the corresponding change label map.

[0026] C. Cut the label image obtained from B and the preprocessed dual-temporal optical remote sensing image from A into the same size to obtain training samples.

[0027] D. Most change detection algorithms do not fully extract the overall features of the changed region and lack attention to the edge details of the changed target. To address this issue, this patent establishes an end-to-end network architecture based on NestedUNet. Through dense skip connections between the encoder and decoder, and between decoders, it can combine feature information at different levels and extract overall features. Since the change detection task can be viewed as segmenting the changed region in a two-time image, the pre-change and post-change remote sensing images from the training samples can be concatenated and input into AFNUNet.

[0028] E. The concatenated image from the previous step is decoded using a residual block based on effective channel attention. First, the input feature map is enlarged by a first 3×3 convolutional layer. Then, the number of channels is doubled and halved by a second and third 3×3 convolutional layer, respectively. This is then fed into the improved effective channel attention layer and added to the feature map obtained from the first 3×3 convolutional layer to form a residual, followed by a ReLU layer. Finally, the feature map containing more information about the changing regions is downsampled using a max-pooling layer. This encoding process is repeated four times to obtain multi-size feature maps X with different receptive fields, equivalent to 1 / 2, 1 / 4, 1 / 8, and 1 / 16 of the original image's width and height. 1,0 X 2,0 X 3,0 X 4,0 .

[0029] F. Obtain the multi-size feature map X from E. 1,0 X 2,0 X 3,0 X 4,0 Perform decoding operations, such as Figure 2As shown: First, the encoding layer X, with a size of 1 / 16 of the original image's width and height... 4,0 Upsampling, with a coding layer X that is 1 / 8 the width and height of the original image. 3,0 The streams are fused to obtain the decoded stream X. 3,1 Next, the coding layer X, which is 1 / 8 the width and height of the original image, is... 3,0 Upsampling, with a coding layer X that is 1 / 4 the width and height of the original image. 2,0 The streams are fused to obtain the decoded stream X. 2,1 Next, the decoded stream X... 3,1 Upsampling, with a coding layer X that is 1 / 4 the width and height of the original image. 2,0 and decoded stream X 2,1 The streams are fused to obtain the decoded stream X. 2,2 Then, an encoding layer X, with a size of 1 / 4 the width and height of the original image, is added. 2,0 Upsampling, with a coding layer X of size 1 / 2 of the original image. 1,0 The streams are fused to obtain the decoded stream X. 1,1 (i.e., the final feature map F1). Next, the decoded stream X will be... 2,1 Upsampling, with a coding layer X of size 1 / 2 of the original image. 1,0 and decoded stream X 1,1 The streams are fused to obtain the decoded stream X. 1,2 (i.e., the final feature map F2). Finally, the decoded stream X... 2,2 Upsampling, with a coding layer X of size 1 / 2 of the original image. 1,0 and decoded stream X 1,1 Decoding stream X 1,2 The streams are fused to obtain the decoded stream X. 1,3 (i.e., the final feature map F3).

[0030] G. Input the final feature maps F1, F2, and F3 obtained in F into AFM. AFM adaptively selects features from different feature maps in both the channel and spatial directions to emphasize more feature information relevant to the changes while suppressing irrelevant features. First, the three feature maps F1, F2, and F3 extracted from the AFNUNet backbone network are fused by element-wise summation:

[0031] F = F1 + F2 + F3

[0032] Next, the fused feature F is fed into the spatial attention submodule. Max pooling and average pooling layers are used on the fused feature F to generate two aggregate vectors of size C×1×1 (C is the number of channels). Then, the multilayer perception (MLP) module is applied to each vector. After passing through the MLP, two aggregate vectors of size 3C×1×1 are obtained, and the two aggregate vectors are added together. Soft attention (softmax layer) is used to adaptively select appropriate features at the appropriate semantic level. Let a, b, and c represent the soft attention obtained after passing through the softmax layer [1], where the size of a, b, and c is C×1×1 (a i b i and c i Let a, b, and c represent the i-th elements respectively. Utilizing the property of the softmax layer that the sum of specified dimensions equals 1, we obtain a. i +b i +c i =1. Finally, the feature map F is obtained through the attention weights on different channels. c :

[0033] F c = a·F1 + b·F2 + c·F3

[0034] The spatial attention submodule also employs max pooling and average pooling layers in the first step to generate two matrices of size 1×H×W. A convolutional layer with a kernel size of 7 is applied to each matrix. After the convolutional layer, two matrices of size 3×H×W are obtained. The two matrices are then summed, and attention is assigned to the summed matrix using softmax. Let a, b, and c represent the soft attention obtained after the softmax layer, where a, b, and c are all of size 1×H×W. i,j b i,j and c i,j Let a, b, and c represent the i-th row and j-th element, respectively, where a i,j +b i,j +c i,j =1). Finally, the feature map F is obtained through attention weights in different spaces. s :

[0035] F s = a·F1 + b·F2 + c·F3

[0036] Finally, the feature maps obtained from channel attention and spatial attention are added together to obtain the final fused features:

[0037] F f =F c +F s

[0038] F f Upsampling and 1×1 convolution are used to obtain the predicted change map.

[0039] H. Introduce a loss based on Blaketis distance. and binary cross-entropy loss Combined to form a hybrid loss function L = L bce +L bcd The loss between the predicted change map of the bi-temporal image obtained in G and the corresponding label map is calculated, where y represents the probability that a certain pixel in the predicted change map is a changed pixel. i,j This represents the probability that a certain pixel in the corresponding label image is a variable pixel, and n represents the width and height of the image at the pixel level.

[0040] I. After training is complete, save all parameter information of the trained model;

[0041] J. After preprocessing the preceding and following time-phase remote sensing images to be detected by orthorectification, image registration, image stretching, and image numerical normalization, the images are cut into samples of the same size to obtain the samples to be detected.

[0042] K. Input the sample to be tested into the change detection model saved in I, and output the detection result map after calculation.

[0043] To address the issues of incomplete extraction of change regions and irregular edges of changed targets by existing change detection models, this invention employs an improved NestedUNet and an adaptive fusion module. The improved NestedUNet enhances overall feature extraction through dense feature transfer between the encoder and decoder, and between decoders themselves, resulting in more complete extracted change regions. The adaptive fusion module effectively combines deep networks with strong semantic information representation capabilities with shallow networks with strong detail information representation capabilities, integrating feature information from different network levels. Furthermore, it adaptively selects information related to change features from feature maps at different semantic levels, thereby more accurately and effectively enhancing the extraction of detailed edge information of changed targets.

[0044] This patent designs three sets of experiments: the first set of experiments is conducted on the LEVIR-CD dataset [2] for building change detection; the second set of experiments is conducted on the WHU-CD dataset [3], which mainly contains changes in building types; and the third set of experiments is conducted on the SYSU-CD dataset [4], which contains multiple types of changes. In order to verify the effectiveness of the proposed AFNUNet, the following eight state-of-the-art remote sensing image change detection methods are selected for comparison with the method of this invention, and they are briefly introduced.

[0045] Fully Convolutional-Early Fusion (FC-EF): FC-EF[5] is based on the U-Net architecture, in which bitemporal images are concatenated into multi-band images for input, and skip connections are used to progressively transfer multi-scale features from the encoder to the decoder to recover spatial information. Fully Convolutional-Siamese-Concatenation (FC-Siam-conc): As a variant of the FC-EF model, FC-Siam-conc[5] uses a Siamese encoder to extract features from bitemporal images, and then connects features of the same level from the encoder to the decoder. Fully Convolutional-Siamese-Difference (FC-Siam-diff) is different from FC-Siam-conc. The skip connections in FC-Siam-diff[5] are another type of FC-EF model that transfer the absolute difference between bitemporal features. Dual Task Constrained DeepSiamese Convolutional Network (DTCDSCN): DTCDSCN[6] introduces a dual attention module (DAM) in the feature extraction stage, which can improve feature representation from the correlation between channels and spatial locations to further improve feature recognition. DAM effectively overcomes the heterogeneity problem by fusing features. Image Fusion Network (IFN): IFN[7] uses channel attention and spatial attention to cross-utilize the feature maps obtained by the VGG16 pre-trained model multiple times at multiple scales to perform effective fusion to obtain change maps more accurately. Dual Attentive Fully Convolutional Siamese Networks (DASNet): DASNet[8] applies the attention mechanism to Siamese networks. UNet++ with multiple side output fusion (UNet++_MSOF): UNet++_MSOF[9] inputs concatenated dual-temporal images into the UNet++ backbone to obtain feature maps with multiple sides output, and captures more detailed information by fusing feature maps with multiple sides.Siamese Nested UNet-Change Detection (SNUNet-CD): SNUNet-CD

[10] combines the Siamese network with the UNet++ network. First, the dual-temporal images are input into the two branches of the Siamese network. Then, the feature maps obtained from the two branches are spliced ​​together and input into UNet++. Finally, the feature maps obtained from the backbone network are fused at multiple semantic levels using the Ensemble Channel Attention Module (ECAM), thereby suppressing localization errors and semantic gaps.

[0046] Table I presents comparative experiments on the LEVIR-CD dataset, Table II on the WHU-CD dataset, and Table III on the SYSU-CD dataset. Precision (P), recall (R), F1 score (F1), and intersection over union (IoU) are used to quantitatively evaluate the performance of the methods involved. These metrics are calculated as follows:

[0047]

[0048]

[0049]

[0050]

[0051] In this calculation, true positives (TP) represent the number of correctly detected non-changed pixels, false positives (FP) represent the number of unpredicted non-changed pixels, and false negatives (FN) represent the number of unpredicted changed pixels. Precision represents the probability that all detected pixels have changed. Recall represents the probability that all changed pixels have been correctly detected. F1 is the harmonic mean of recall and accuracy, which balances conflicts by considering both recall and accuracy simultaneously. IoU is the area of ​​overlap between predicted changed pixels and the area of ​​union between them.

[0052] Table I shows the comparative experiments conducted on the LEVIR-CD dataset.

[0053]

[0054] Table II shows the comparative experiments conducted on the WHU-CD dataset.

[0055]

[0056] Table III shows the comparative experiments conducted on the SYSU-CD dataset.

[0057]

[0058] As can be seen from the data in the three tables above, compared with other existing remote sensing image change detection methods, the proposed method improves F1 and IoU by 1.3% and 2.1%, 1.6% and 1.8%, and 1.8% and 1.6% respectively on the LEVIR-CD, WHU-CD, and SYSU-CD datasets. The proposed method achieves best performance on all three publicly available datasets, enabling more complete identification of changed regions.

[0059] To address the problems of existing technologies, this patent constructs an end-to-end network architecture called Adaptive Fusion NestedUNet (AFNUNet). It achieves the combination of feature information at different levels and the extraction of overall features through dense skip connections between the encoder and decoder, as well as between decoders. To enable the network to adaptively select changing features and improve its ability to extract edge details of changing targets, this patent constructs an Adaptive Fusion Module (AFM) based on channel attention and spatial attention. This patent also introduces a loss based on Blectis distance and combines it with binary cross-entropy loss to form a hybrid loss function, thereby more accurately perceiving differences between two-temporal optical remote sensing images.

[0060] The improved NestedUNet proposed in this patent enhances overall feature extraction through dense feature transfer between the encoder and decoder, as well as between decoders, resulting in more complete extracted regions of change. The adaptive fusion module proposed in this patent can adaptively select information related to the changing features from feature maps at different semantic levels, thereby more accurately and effectively enhancing the extraction of edge details of changing targets.

[0061] The references in the invention patent are as follows:

[0062] [1]Li X,Wang W,Hu

[0063] [2]Chen H,Shi Z.A Spatial-Temporal Attention-Based Method and a NewDataset for Remote Sensing Image Change Detection[J].Remote Sensing,2020,12(10):1662.

[0064] [3]Shunping,Ji,Shiqing,et al.Fully Convolutional Networks forMultisource Building Extraction From an Open Aerial and Satellite ImageryData Set[J].IEEE Transactions on Geoscience and Remote Sensing,2019,57(1):574-586.

[0065] [4]Shi Q,Liu M,Li S,et al.A Deeply Supervised Attention Metric-BasedNetwork and an Open Aerial Image Dataset for Remote Sensing Change Detection[J].IEEE Transactions on Geoscience and Remote Sensing,2022,60.

[0066] [5]Daudt R C,Saux B L,Boulch A.Fully Convolutional Siamese Networksfor Change Detection[J].2018 25th IEEE International Conference on ImageProcessing(ICIP),2018.

[0067] [6]Liu Y,Pang C,Zhan Z,et al.Building Change Detection for RemoteSensing Images Using a Dual Task Constrained Deep Siamese ConvolutionalNetwork Model[J].2019.

[0068] [7]Cz A,Peng Y,Dt E,et al.A deeply supervised image fusion networkfor change detection in high resolution bi-temporal remote sensing images[J].ISPRS Journal of Photogrammetry and Remote Sensing,2020,166:183-200.

[0069] [8]Chen J,Yuan Z,Peng J,et al.DASNet:Dual attentive fullyconvolutional siamese networks for change detection of high resolutionsatellite images[J].IEEE Journal of Selected Topics in Applied EarthObservations and Remote Sensing,2020,PP(99).

[0070] [9]Guan H.End-to-End Change Detection for High Resolution SatelliteImages Using Improved UNet++[J].Remote Sensing,2019,11.

[0071]

[10] Fang S,Li K,Shao J,et al.SNUNet-CD:A Densely Connected SiameseNetwork for Change Detection of VHR Images[J].IEEE Geoscience and RemoteSensing Letters,2021,PP(99):1-5.

[0072] In the description of this invention, it should be noted that the directional terms such as "center", "lateral", "longitudinal", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", and "counterclockwise" indicate the orientation and positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. They should not be construed as limiting the specific protection scope of this invention.

[0073] It should be noted that the terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be used interchangeably where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0074] Note that the above description is merely a preferred embodiment and application of the technical principles of the present invention. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the specific embodiments described herein, and may include many other effective embodiments without departing from the concept of the present invention. The scope of the present invention is determined by the scope of the appended claims.

Claims

1. A method for detecting changes in optical remote sensing images based on adaptive fusion of NestedUNet, characterized in that: It includes the following steps: A. Perform orthorectification, image registration, image stretching, and image numerical normalization preprocessing on the dual-temporal optical remote sensing images in sequence to obtain dual-temporal optical remote sensing images with consistent data distribution. B. Label the updated parts of the preprocessed dual-temporal optical remote sensing image obtained in the previous step to obtain the corresponding change label map. C. Cut the change label map obtained from B and the preprocessed dual-temporal optical remote sensing image obtained from A into the same size to obtain training samples; D. The dual-temporal remote sensing images in the training samples are concatenated and input into AFNUNet; the AFNUNet is based on the end-to-end network architecture of NestedUNet, and its output end has an adaptive fusion module of channel attention and spatial attention. E. The concatenated dual-temporal remote sensing images obtained in the previous step are passed through a four-layer residual block encoder based on effective channel attention to extract multi-size feature maps with different receptive fields. Step E specifically includes the following steps: First, the input feature map is upscaled by the first 3×3 convolutional layer. Then, the number of channels is doubled and halved by the second and third 3×3 convolutional layers, respectively. Then, it is input into the improved effective channel attention layer and added to the feature map obtained by the first 3×3 convolutional layer to form a residual. Then, it is passed through a ReLU layer. Finally, the feature map containing more features related to the changing regions is downsampled by a max pooling layer. The above encoding process is repeated four times to obtain multi-size feature maps with different receptive fields, which are equivalent to 1 / 2, 1 / 4, 1 / 8, and 1 / 16 of the original image width and height. F. Decode the multi-size feature maps with different receptive fields in E; enhance the extraction of overall features through dense feature transfer between the encoder and decoder and between decoders, making the extracted change regions more complete, and obtain the final feature maps F1, F2, and F3. G. Input the final feature maps F1, F2, and F3 obtained in F into the adaptive fusion module to obtain more detailed information about the changed regions and to obtain the predicted change map; H. Introduce a loss based on Brectis distance and combine it with the binary cross-entropy loss to form a hybrid loss function to calculate the loss between the predicted change map of the bi-temporal image obtained in step G and the corresponding label map. I. After training is complete, save all parameter information of the trained model; J. After the preceding and following time-phase remote sensing images to be detected are preprocessed by orthorectification, image registration, image stretching and image numerical normalization, they are then cut into the same size to obtain the sample to be tested. K. Input the sample to be tested into the change detection model obtained in I, and output the detection result map after calculation.

2. The optical remote sensing image change detection method based on adaptive fusion NestedUNet according to claim 1, characterized in that: The fusion feature F in step G f It consists of two modules, including a channel submodule: Spatial submodule: , , , These represent the soft attention and fused features obtained after the softmax layer, respectively. .

3. The optical remote sensing image change detection method based on adaptive fusion NestedUNet according to claim 1, characterized in that: The hybrid loss function mentioned in step H is Among them, the binary cross-entropy loss Loss based on Brectis distance ,in This represents the probability that a certain pixel in the predicted change map is a changed pixel. This represents the probability that a certain pixel in the corresponding label image is a variable pixel, and n represents the width and height of the image in pixels.

Images