A remote sensing image change detection method

Abstract

The present application relates to a kind of remote sensing image change detection method, belong to remote sensing image identification technical field.The present application selects the input mode of twin network capable of sharing weight in encoding part, respectively input same network branch to double time phase image and carry out feature extraction, reduce overfitting phenomenon by sharing weight, wherein two network branches select residual connection module as convolution unit and extract the feature of image respectively, accelerate the convergence speed of network;Simultaneously using a kind of dense connection method, the information lost in the process of continuous down sampling of image is aggregated layer by layer by up sampling, so as to improve the extraction ability of small target to network;And introduce attention module in the transverse connection of encoding part and decoding part to enhance the acquisition ability to change information, improve the distinguishing ability of boundary under complex background;Finally, through the up sampling of decoding part, feature map is restored into original image size, and input classification layer, obtain accurate change detection result.

Description

Technical Field

[0001] This invention relates to a method for detecting changes in remote sensing images, belonging to the field of remote sensing image recognition technology. Background Technology

[0002] Remote sensing image change detection has been widely applied in various fields such as topographic map updating, urban change analysis, disaster assessment, and landslide monitoring. Traditional change detection methods, such as image algebra, image transformation, and post-classification comparison, have shown good results on low- and medium-resolution remote sensing images. However, as image resolution increases and image backgrounds become more complex, it becomes not only more difficult to achieve automatic change detection, but the accuracy of the change detection results often fails to meet the needs of practical applications.

[0003] Deep learning, with its powerful self-learning capabilities, is widely used in fields such as computer vision and remote sensing. Models based on the Fully Convolutional Network (FCN) structure have been applied to semantic segmentation tasks, enabling end-to-end feature extraction from images. Typical networks include UNet, SegNet, and DenseNet. The continuous development of semantic segmentation networks has led to the rapid advancement of remote sensing image change detection tasks. Siamese networks, with their dual-branch shared weights, have gradually become a popular input method in remote sensing image change detection. Daudt et al. discussed three types of fully convolutional neural networks: FC-EF, FC-Siam-conc, and FC-Siam-diff, verifying the effectiveness of Siamese networks in change detection tasks. In FC-Siam-conc and FC-Siam-diff, the encoding layer is divided into two branches with the same structure and shared weights. The difference is that the decoding structure of FC-Siam-conc integrates features obtained from the encoding structure, while FC-Siam-diff calculates the absolute value of the feature map difference of the decoding structure and then inputs it into the corresponding layer of the decoding structure. Some researchers have improved the performance of remote sensing image change detection by introducing modules such as attention and depth metrics into the network, which can better distinguish between changed and unchanged features. For example, SUNet-CD adds an integrated channel attention module to UNet++ to refine the most representative features at different semantic levels, thereby achieving the change detection task; STANet adds a Basic Spatial-Temporal Attention module (BAM) and a Pyramid Spatial-Temporal Attention module (PAM) to the Siamese network, enabling the network to better adapt to targets at different scales and improve the accuracy of change detection; DASNet uses a dual attention mechanism to capture long-term dependencies and obtain more discriminative features, thereby improving the model's recognition ability; DSAMNet uses a depth metric module to enhance the learning ability of the feature extractor and generate more refined features.

[0004] Currently, a widely used method in remote sensing image change detection is to directly input remote sensing images from two different time phases into a convolutional neural network (CNN) to output the final change result. There are two ways to directly input remote sensing images into a CNN for feature extraction. One method is to overlay the two time-phase images along the channel dimension and use them together as network input. This leverages the excellent feature extraction capabilities of deep CNNs to extract change information from the overlaid images. However, this method struggles to maintain the high-dimensional features of the original images. The other method is to input the two images separately into a Siamese network. The sub-networks share weights, requiring fewer parameters for training and reducing overfitting. The Siamese network can simultaneously extract features from the two time-phase images, obtaining image feature maps through weight sharing. These feature maps are then compared to obtain the features of the changed areas, achieving accurate pixel-level classification.

[0005] Although the above methods are constantly improving the accuracy of change detection, there are still problems such as coarse boundary segmentation and missed detection of small targets in the change detection of high-resolution remote sensing images. Summary of the Invention

[0006] The purpose of this invention is to provide a method for detecting changes in remote sensing images, in order to solve the problems of coarse boundary segmentation and missed detection of small targets in current remote sensing image change detection methods.

[0007] To solve the above-mentioned technical problems, the present invention provides a method for detecting changes in remote sensing images, the method comprising the following steps:

[0008] 1) Obtain the remote sensing image pair to be detected, namely the first remote sensing image and the second remote sensing image, which are of the same size;

[0009] 2) Construct a change detection model, which includes an encoding part and a decoding part. The encoding part uses a Siamese network, the decoding part uses an upsampling network, and the encoding part and the decoding part are connected by an attention module.

[0010] The twin network includes a first network branch and a second network branch. Both the first and second network branches use several layers of residual connection modules for feature extraction. The first network branch is used to extract features from the first remote sensing image to obtain first remote sensing image features at different scales. The second network branch is used to extract features from the second remote sensing image to obtain second remote sensing image features at different scales. The first and second network branches are connected in a dense manner to superimpose and fuse the first and second remote sensing image features at the same scale to obtain fused features at different scales.

[0011] The attention module is used to perform feature enhancement processing on the fused features at different scales respectively; the decoding part is used to perform feature recovery and fusion on the feature-enhanced features at different scales through an upsampling network to obtain the change detection result.

[0012] This invention employs a Siamese network input method in the encoding section, allowing for weight sharing. Dual-temporal images are input into the same network branch for feature extraction, reducing overfitting through weight sharing. The two network branches use residual connection modules as convolutional units to extract image features, accelerating network convergence. Simultaneously, a dense connection method is used to aggregate information lost during image downsampling through upsampling layer by layer, improving the network's ability to extract small targets. Furthermore, attention modules are introduced at the lateral connections between the encoding and decoding sections to enhance the ability to acquire change information, improving boundary discrimination in complex backgrounds. Finally, upsampling in the decoding section restores the feature map to the original image size, which is then input into the classification layer to obtain accurate change detection results.

[0013] Furthermore, the residual connection module includes a first convolutional layer, a second convolutional layer, a first regularization layer, a second regularization layer, an activation layer, and a stacking module. The first convolutional layer is used to perform convolution operations on the input features. The output of the first convolutional layer is processed by the first regularization layer and the activation layer and then sequentially input into the second convolutional layer and the second regularization layer. After being processed by convolution and regularization again, it is input into the stacking module, which is used to stack the features output by the second regularization layer and the features output by the first convolutional layer. The stacked features are the output of the residual connection module.

[0014] The residual connection module of this invention adopts the concept of identity shortcut connection, which can accelerate the convergence speed of the network and improve the detection efficiency.

[0015] Furthermore, the dense connection includes an upsampling layer and a stacking layer. The upsampling layer is used to upsample the output of a residual connection module in a certain layer of the second network branch, and the stacking layer is used to stack the upsampled result with the output of the residual connection module of the previous layer in the second network branch and the output of the corresponding residual connection module in the first network branch.

[0016] This invention learns shallow spatial information by densely connecting feature maps of different scales obtained in the encoding part, and also transmits the obtained deep semantic information to the shallow layer, thereby improving the ability to acquire information about changes in small targets and solving the problem of information loss during the continuous downsampling of images.

[0017] Furthermore, the attention module is a convolutional attention module.

[0018] The convolutional attention module used in this invention enhances the attention to necessary features while suppressing unnecessary features, thereby enhancing the ability to acquire changing information.

[0019] Furthermore, the first network branch includes N layers of residual connection modules, and the second network branch includes N+1 layers of residual connection modules, where N is greater than or equal to 3. The input to the first layer of the first network branch's residual connection module is the first remote sensing image, and the output is a convolutional feature with the same scale as the first remote sensing image. The input to the second layer of the first network branch's residual connection module is the output of the first layer of the residual connection module, and the output is a convolutional feature with a scale half that of the first remote sensing image. The input to the Nth layer of the first network branch's residual connection module is the output of the (N-1)th layer of the residual connection module, and the output is a convolutional feature with a scale half that of the first remote sensing image. N-1 The input to the first residual connection module of the second network branch is the second remote sensing image, and the output is a convolutional feature with the same scale as the second remote sensing image. The input to the second residual connection module of the second network branch is the output of the first residual connection module, and the output is a convolutional feature with a scale of 1 / 2 that of the first remote sensing image. The input to the (N+1)th residual connection module of the second network branch is the output of the Nth residual connection module, and the output is a convolutional feature with a scale of 1 / 2 that of the second remote sensing image. N The convolutional features.

[0020] Furthermore, the dense connections include N, where the input of the upsampling layer of the Nth dense connection is the output of the (N+1)th layer residual connection module of the second network branch, and the superposition layer of the Nth dense connection is used to superimpose the output of the upsampling layer of the dense connection, the output of the Nth layer residual connection module of the first network branch, and the output of the Nth layer residual connection module of the second network branch, and the superposition result is the output of the dense connection; the input of the upsampling layer of the (N-1)th dense connection is the output of the Nth dense connection, and the superposition layer of the (N-1)th dense connection is used to superimpose the output of the upsampling layer of the dense connection, the output of the (N-1)th layer residual connection module of the first network branch, and the output of the (N-1)th layer residual connection module of the second network branch, and the superposition result is the output of the (N-1)th dense connection.

[0021] Furthermore, N is 4.

[0022] Furthermore, the change detection model employs a hybrid loss function of weighted cross-entropy and Dice Loss during training.

[0023] Furthermore, the weighted cross-entropy loss function used is:

[0024]

[0025] Where i represents a pixel; w is a weighting coefficient; y i p i These represent the predicted label of pixel i and the probability that pixel i belongs to the change class, respectively.

[0026] Furthermore, the loss function used in Dice Loss is:

[0027]

[0028] Where Y represents the true value of the sample. This represents the predicted value of the sample. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the network structure of the detection module used in the remote sensing image change detection method of the present invention;

[0030] Figure 2 This is a schematic diagram of the residual connection module used in this invention;

[0031] Figure 3 This is a schematic diagram of the dense connections used in this invention;

[0032] Figure 4 This is a schematic diagram of the network structure of the convolutional attention module used in this invention;

[0033] Figure 5(a) is a remote sensing image of region 1 selected during the experiment of this invention at time T1;

[0034] Figure 5(b) is a remote sensing image of region 1 selected during the experiment of this invention at time T2;

[0035] Figure 5(c) is a true value diagram of the changes in region 1 selected during the experiment of this invention at times T1 and T2;

[0036] Figure 5(d) is a schematic diagram of the change detection results of the selected region 1 remote sensing image using the FC-Siam-conv model in the experiment of this invention;

[0037] Figure 5(e) is a schematic diagram of the change detection results of the selected region 1 remote sensing image using the FC-Siam-diff model during the experiment of this invention;

[0038] Figure 5(f) is a schematic diagram of the change detection results of the selected region 1 remote sensing image using the DSAMNet model during the experiment of this invention;

[0039] Figure 5(g) is a schematic diagram of the change detection results of the selected region 1 remote sensing image using the SUN-CD model during the experiment of this invention;

[0040] Figure 5(h) is a schematic diagram of the change detection results of the selected region 1 remote sensing image using the CDASNet model of the present invention during the experiment of the present invention;

[0041] Figure 6(a) is a remote sensing image of region 2 selected during the experiment of this invention at time T1;

[0042] Figure 6(b) is a remote sensing image of region 2 selected during the experiment of this invention at time T2;

[0043] Figure 6(c) is a true value diagram of the changes in region 2 selected during the experiment of this invention at times T1 and T2;

[0044] Figure 6(d) is a schematic diagram of the change detection results of the selected region 2 remote sensing image using the FC-Siam-conv model in the experiment of this invention;

[0045] Figure 6(e) is a schematic diagram of the change detection results of the selected region 2 remote sensing image using the FC-Siam-diff model in the experiment of this invention;

[0046] Figure 6(f) is a schematic diagram of the change detection results of the selected region 2 remote sensing image using the DSAMNet model during the experiment of this invention;

[0047] Figure 6(g) is a schematic diagram of the change detection results of the selected region 2 remote sensing image using the SUN-CD model during the experiment of this invention;

[0048] Figure 6(h) is a schematic diagram of the change detection results of the selected region 2 remote sensing image using the CDASNet model of the present invention during the experiment of the present invention;

[0049] Figure 7(a) is a remote sensing image of region 3 selected during the experiment of this invention at time T1;

[0050] Figure 7(b) is a remote sensing image of region 3 selected during the experiment of this invention at time T2;

[0051] Figure 7(c) is a true value diagram of the changes in region 3 selected during the experiment of this invention at times T1 and T2;

[0052] Figure 7(d) is a schematic diagram of the change detection results of the selected region 3 remote sensing image using the FC-Siam-conv model in the experiment of this invention;

[0053] Figure 7(e) is a schematic diagram of the change detection results of the selected region 3 remote sensing image using the FC-Siam-diff model during the experiment of this invention;

[0054] Figure 7(f) is a schematic diagram of the change detection results of the selected region 3 remote sensing image using the DSAMNet model during the experiment of this invention;

[0055] Figure 7(g) is a schematic diagram of the change detection results of the selected region 3 remote sensing image using the SUN-CD model during the experiment of this invention;

[0056] Figure 7(h) is a schematic diagram of the change detection results of the selected region 3 remote sensing image using the CDASNet model of the present invention during the experiment of the present invention;

[0057] Figure 8(a) is a remote sensing image of region 4 selected during the experiment of this invention at time T1;

[0058] Figure 8(b) is a remote sensing image of region 4 selected during the experiment of this invention at time T2;

[0059] Figure 8(c) is a true value diagram of the changes in region 4 selected during the experiment of this invention at times T1 and T2;

[0060] Figure 8(d) is a schematic diagram of the change detection results of the selected region 4 remote sensing image using the FC-Siam-conv model in the experiment of this invention;

[0061] Figure 8(e) is a schematic diagram of the change detection results of the selected region 4 remote sensing image using the FC-Siam-diff model during the experiment of this invention;

[0062] Figure 8(f) is a schematic diagram of the change detection results of the selected region 4 remote sensing image using the DSAMNet model during the experiment of this invention;

[0063] Figure 8(g) is a schematic diagram of the change detection results of the selected region 4 remote sensing image using the SUN-CD model during the experiment of this invention;

[0064] Figure 8(h) is a schematic diagram of the change detection results of the selected region 4 remote sensing image using the CDASNet model of the present invention during the experiment of the present invention;

[0065] Figure 9(a) is a remote sensing image selected at time T1 during the ablation experiment of this invention;

[0066] Figure 9(b) is a remote sensing image of the T2 period selected during the ablation experiment of this invention;

[0067] Figure 9(c) is a true value diagram of the changes in T1 and T2 selected during the experiment of this invention;

[0068] Figure 9(d) is a schematic diagram of the changes in the base model during the ablation experiment of this invention.

[0069] Figure 9(e) is a schematic diagram of the change detection results of the model using base+CBAM during the ablation experiment of this invention;

[0070] Figure 9(f) is a schematic diagram of the change detection results of the model using base+DC during the ablation experiment of this invention;

[0071] Figure 9(g) is a schematic diagram of the change detection results of the model using CDASNet during the ablation experiment of this invention;

[0072] Figure 9(h) is a detailed diagram of the changes detected by the model using CDASNet during the ablation experiment of this invention. Detailed Implementation

[0073] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0074] Method Implementation Examples

[0075] This invention proposes the CDASNet network based on the idea that the UNet network can extract multi-scale features from multi-layer paths. The overall structure is as follows: Figure 1 As shown in the diagram, this network, tailored to the characteristics of remote sensing image change detection data, employs a Siamese network input method in the encoding stage. Overfitting is reduced through shared weights. The two sub-networks use residual connection modules as convolutional units to extract image features, accelerating network convergence. Simultaneously, a novel dense connection method is used to aggregate information lost during continuous downsampling through upsampling layer by layer. Attention modules are introduced at the lateral connections in the encoding and decoding stages to enhance the ability to acquire change information. Finally, after upsampling, the feature map is restored to the original image size and input into the classification layer to obtain the change detection results.

[0076] Specifically, such as Figure 1As shown, the network used for remote sensing image change detection in this invention is the CDASNet network. This network includes an encoding part and a decoding part. The encoding part uses a Siamese network, and the decoding part uses an upsampling network. The encoding and decoding parts are connected by an attention module. The Siamese network includes a first network branch and a second network branch. Both the first and second network branches use several layers of residual connection modules for feature extraction. The first network branch is used to extract features from the first remote sensing image to obtain features of the first remote sensing image at different scales. The second network branch is used to extract features from the second remote sensing image to obtain features of the second remote sensing image at different scales. The first and second network branches are densely connected to superimpose and fuse the features of the first and second remote sensing images at the same scale to obtain fused features at different scales. The attention module is used to perform feature enhancement processing on the fused features at different scales respectively. The decoding part is used to recover and fuse the features at different scales after feature enhancement processing through the upsampling network to obtain the change detection result. Here, the first and second remote sensing images refer to images of the same area at different time periods.

[0077] The first network branch includes N residual connection modules, and the second network branch includes N+1 residual connection modules, where N is greater than or equal to 3. In this embodiment, N is 4, meaning the first network branch includes 4 residual connection modules and the second network branch includes 5 residual connection modules. The input to the first residual connection module of the first network branch is the first remote sensing image (image from time T1, abbreviated as T1), and the output is a convolutional feature F with the same scale as the first remote sensing image. 11 The input to the second-layer residual connection module of the first network branch is the output F of the first-layer residual connection module. 11 The output is a convolutional feature F with a scale of half that of the first remote sensing image. 12 Similarly, the input of the fourth-layer residual connection module in the first network branch is the output F of the third-layer residual connection module. 13 The output is at a scale of 1 / 8 of the first remote sensing image. 1 Convolutional features F 14 The input to the first residual connection module of the second network branch is the second remote sensing image (the image from time T2, abbreviated as T2), and the output is the convolutional feature F with the same scale as the second remote sensing image. 21 The input to the second-layer residual connection module of the second network branch is the output F of the first-layer residual connection module. 21 The output is a convolutional feature F with a scale of half that of the first remote sensing image. 22 Similarly, the input to the fifth-layer residual connection module of the second network branch is the output F of the fourth-layer residual connection module. 24The output is a convolutional feature F with a scale of 1 / 16 of the second remote sensing image. 25 Therefore, the features obtained through the first network branch are F in sequence. 11 F 12 F 13 F 14 The features obtained from the second network branch are F in sequence. 21 F 22 F 23 F 24 F 25 .

[0078] like Figure 2 As shown, the residual connection module adopts the Identity Shortcut Connection, which includes a first convolutional layer, a second convolutional layer, a first regularization layer, a second regularization layer, an activation layer, and a stacking module. The first convolutional layer is used to perform convolution operations on the input features. The output of the first convolutional layer is processed by the first regularization layer and the activation layer and then input into the second convolutional layer and the second regularization layer. After being convolved and regularized again, it is input into the stacking module, which is used to stack the features output by the second regularization layer and the features output by the first convolutional layer. The stacked features are the output of the residual connection module. x is the input. The objective function H(x) decomposes the network structure into two different branches. One branch is the residual mapping F(x). In this branch, x first passes through a 3×3 convolutional layer to obtain a feature vector. Then, Batch Normalization (BN) is used to make the feature vector satisfy the distribution law with a mean of 0 and a variance of 1. After that, the ReLU nonlinear activation function is selected to reduce the computational cost of the network and alleviate the overfitting phenomenon. Then, it passes through a second 3×3 convolutional layer and a BN layer to obtain F(x). The other branch directly passes the input information x after the 3×3 convolutional layer to the output and adds it with F(x) to obtain the final output result, which preserves the integrity of the original input information.

[0079] The first and second network branches are connected via dense connections (DCs), such as... Figure 3As shown, the dense connection includes an upsampling layer and a stacking layer. The upsampling module is used to upsample the output of a residual connection module in a certain layer of the second network branch. The stacking layer is used to superimpose the upsampled result with the output of the previous layer of residual connection module in the second network branch and the output of the corresponding layer of residual connection module in the first network branch. Since the first network branch uses four layers of residual connection modules to extract features of different sizes in this embodiment, four dense connections are also used to fuse the features extracted by the first network branch and the features extracted by the second network branch. For the fourth dense connection, the fifth layer of residual connection module in the second network branch extracts the feature F of 1 / 16 of the original image. 25 The feature F will be obtained. 25 The input is fed into the densely connected upsampling layer for upsampling processing to obtain 1 / 8 of the original image's features. These 1 / 8 features are then combined with the residual connection module of the 4th layer of the first network branch to extract the 1 / 8 feature F of the original image. 14 The residual connection module of the fourth layer of the second network branch extracts 1 / 8 of the feature F of the original image. 24 The feature map F4 is obtained by overlaying and fusing, and then F4 is upsampled and then combined with F. 13 F 23 F3 is obtained by overlaying the data on the channels; similarly, fusion features at different scales can be obtained.

[0080] By employing this dense connection method, the present invention can use feature maps F1, F2, F3, and F4 of different scales obtained in the encoding stage as input to the decoding stage through a single skip connection, reducing the number of parameters caused by excessive skip connections. This not only learns shallow spatial information but also transmits the acquired deep semantic information to the shallow layers, improving the ability to acquire information about changes in small targets.

[0081] Due to the complex backgrounds of high-resolution remote sensing images, relying solely on a single model for feature extraction often fails to distinguish mixed pixel regions. Furthermore, not all features have the same effect on the model during decoding. Therefore, a Convolutional Block Attention Module (CBAM) is introduced before feeding the feature maps F1, F2, F3, and F4 from each layer of the encoding part into the decoding part. This enhances the focus on necessary features while suppressing unnecessary ones, resulting in new feature maps F1′, F2′, F3′, and F4′. Then, F4′ is compared with F... 25 In the decoding stage, the features obtained after upsampling are fused to obtain feature D4. Feature D4 is then upsampled and fused with F3′ to obtain feature D3, and so on, until feature D1 is finally obtained. Finally, feature D1 is reduced to 2 dimensions through a 1×1 convolution to generate the final prediction map.

[0082] like Figure 4 As shown, the Convolutional Attention Module (CBAM) used in this invention concatenates the channel attention submodule and the spatial attention submodule, and captures spatial global information by introducing global pooling. The construction of the channel attention submodule is similar to that of the SE module, except that it parallelizes max pooling and average pooling operations to aggregate global information and compresses features in the spatial dimension; while the spatial attention submodule models the spatial relationships of features and compresses features in the channel dimension. Given an input feature map X∈R... C×H×W C, H, and W represent the number of channels, width, and height of the feature, respectively. A one-dimensional channel attention vector M is obtained through the channel attention submodule. c ∈R C×1×1 Then, after multiplying with the input features, they are passed together through the spatial attention submodule to generate a two-dimensional spatial attention vector M. s ∈R 1×H×W Finally, a new feature map F′ is obtained.

[0083] In remote sensing image change detection, there is a significant class imbalance problem, where the number of invariant pixels far exceeds the number of changing pixels. Therefore, during model training, the model tends to focus on learning larger class data while neglecting smaller class data, making it easy to miss changes in small targets. To mitigate the impact of sample imbalance and increase the weight of changing samples in the overall image, this invention uses a hybrid loss function L combining Weighted Cross-Entropy (WCE) and Dice Loss, defined as:

[0084] L = L wce +L dice (1)

[0085] The weighted cross-entropy loss function can be expressed as:

[0086]

[0087] Where i represents a pixel; w is the weight coefficient; y i p i These represent the predicted label of pixel i and the probability that pixel i belongs to the change class, respectively.

[0088] The Dice Loss function is used to calculate the similarity between the ground truth value of the changed sample and the predicted changed sample. The Dice Loss function is expressed as follows:

[0089]

[0090] Where Y represents the true value of the sample, The predicted value represents the sample.

[0091] To further illustrate the detection performance of the remote sensing image change detection method (CDASNet) of this invention, a simulation experiment is conducted below. This experiment selects two remote sensing image datasets: LEVIR-CD and CDD. The LEVIR-CD dataset contains 637 pairs of high-resolution remote sensing images (1024×1024 pixels each) collected from Google Earth, covering 20 different regions in Texas, USA, spanning 5–14 years, with a resolution of 0.5m per pixel. This dataset focuses on changes in buildings. Due to the large size of the remote sensing images, they cannot be directly input into the model. Furthermore, to avoid information loss at tile edges, this experiment divides the LEVIR-CD dataset into 256×256 tiles with a 20% overlap, where 70% are training samples, 10% are validation samples, and the remainder are test samples. The CDD dataset is a publicly available satellite image pair dataset provided by Lebedev. This dataset consists of 11 pairs of multispectral image pairs, obtained from Google Earth, covering the same area and varying seasonally. Seven pairs are 4725×2200 pixels and are used to create manually labeled images. Four pairs are 1900×1000 pixels and are used to manually add other objects. The dataset comprises multi-source remote sensing images with a spatial resolution of 3cm-100cm per pixel. Significant seasonal variations exist between images from different time periods. During label generation, only the appearance and disappearance of objects are considered as image variations, ignoring seasonal changes. This experiment cropped and rotated the original images to generate 10,000 training samples, 3,000 validation samples, and 3,000 test samples, with each sample measuring 256×256 pixels.

[0092] Remote sensing image change detection can essentially be viewed as a pixel-based binary classification problem. Commonly used accuracy evaluation metrics include precision (P), recall (R), F1 score (F1), and intersection-over-union (IOU). Higher precision indicates fewer false positives and better model accuracy; higher recall indicates that positive samples are detected to the greatest extent possible, increasing model value; the F1 score is the harmonic mean of precision and recall, with a higher F1 score indicating better overall prediction results; and IOU is the ratio of the intersection to the union of predicted positive samples and the true values. Its formula is as follows:

[0093]

[0094]

[0095]

[0096]

[0097] Where TP represents the number of pixels that were predicted as positive, FP represents the number of pixels that were predicted as positive, TN represents the number of pixels that were predicted as negative, and FN represents the number of pixels that were predicted as negative. In this experiment, positive samples are the changed pixels, and negative samples are the unchanged pixels.

[0098] To better illustrate the effectiveness of this invention, it is compared with currently mainstream Siamese network change detection methods such as FC-Siam-conc, FC-Siam-diff, DSAMNet, and SUNet-CD. All experiments were conducted in the same environment with consistent training parameters. The experiments used the PyTorch deep learning framework under Linux, with a CPU of Intel(R) Core(TM) i9-9900K, a GPU of GTX2080Ti, and 11GB of VRAM. During training, Adam was used as the optimizer, with a batch size of 4, a total of 100 EPOCH iterations, and an initial learning rate of lr. base The learning rate is set to 0.0001, and a fixed-length decay strategy is used to adjust the learning rate. The decay step size is 8, and the decay factor gamma is set to 0.5.

[0099] To more comprehensively compare the change detection results of various methods in complex backgrounds, two typical regions were selected from the LEVIR-CD and CDD datasets for analysis, denoted as Regions 1 to 4. Region 1, containing images of small, regularly arranged buildings, was used to explore the detection performance of the aforementioned networks for small-sized ground feature details. Region 2, encompassing buildings with significant color feature variations over a large area, was used to discuss the ability of each network to distinguish large-sized ground features. Region 3, containing images of small target ground features including vehicles, was used to explore the detection performance of different networks under class imbalance conditions. Region 4 was mainly used to explore the ability of each network to extract linear ground features.

[0100] For region 1, since there are unchanged ground features with similar spectral characteristics around the changed area, as shown in Figures 5(a) and 5(b), which are remote sensing images of periods T1 and T2 respectively, and Figure 5(c) is the ground truth map of the changes in periods T1 and T2, the detection results using the FC-Siam-conv, FC-Siam-diff, DSAMNet, and SUN-CD models are as follows: Figure 5(d) , 5(e)As shown in Figures 5(f) and 5(g), it can be seen that none of these methods could detect the small changes in the building area in the lower left corner. Some edges are hidden in shadows, making it difficult for FC-Siam-conv to accurately distinguish building boundaries, resulting in incomplete and overlapping edges; FC-Siam-diff exhibits significant false negatives; DSAMNet and SUN-CD, due to their use of depth measurement and channel attention modules respectively, show improved detection performance compared to the previous two networks, but both are easily affected by shadows and trees, leading to minor false positives. The detection results of CDASNet used in this invention are shown in Figure 5(h). By introducing a convolutional attention module, it effectively distinguishes boundary information and utilizes dense connections to fuse high- and low-level features, detecting small-scale changes more completely, with detection results significantly superior to other networks.

[0101] For region 2, Figures 6(a) and 6(b) show the remote sensing images for periods T1 and T2, respectively. Figure 6(c) shows the ground truth map of the changes for periods T1 and T2. The detection results using the FC-Siam-conv, FC-Siam-diff, DSAMNet, and SUN-CD models are shown below. Figure 6(d) , 6(e) As shown in Figures 6(f) and 6(g), it can be seen that FC-Siam-conv exhibits significant false detections; FC-Siam-diff and SUN-CD both misclassify the surrounding grassland as having changed to varying degrees; DSAMNet can extract deep features and detect the range of change relatively accurately, but its ability to extract edge details is weak. The CDASNet used in this invention (as shown in Figure 6(h)) can completely detect areas with significant color changes and has smoother edges.

[0102] For region 3, Figures 7(a) and 7(b) show the remote sensing images for periods T1 and T2, respectively. Figure 7(c) shows the ground truth map of the changes for periods T1 and T2. The detection results using the FC-Siam-conv, FC-Siam-diff, DSAMNet, SUN-CD, and CDASNet models are shown below. Figure 7(d) , 7(e) Figures 7(f), 7(g), and 7(h) are shown. For region 4, Figures 8(a) and 8(b) show the remote sensing images at times T1 and T2, respectively. Figure 8(c) shows the ground truth map of the changes at times T1 and T2. The detection results using the FC-Siam-conv, FC-Siam-diff, DSAMNet, SUN-CD, and CDASNet models are shown below. Figure 8(d) , 8(e)As shown in Figures 8(f), 8(g), and 8(h), regions 3 and 4 reveal that FC-Siam-conv, FC-Siam-diff, and DSAMNet can only detect a few changing regions, resulting in significant missed detections. SUN-CD utilizes dense skip connections and channel integration modules to fuse multi-stage features, improving its ability to detect small objects. However, due to occlusion by shadows and trees, especially when the texture features of the target are similar to invariant regions, missed and false detections still occur. Road detection is prone to breakage and incomplete detection. The CDASNet proposed in this invention effectively improves the detection capability for small targets, even under shadows, and can still completely detect changing regions with more accurate boundary localization. Furthermore, it provides more ideal results for linear features.

[0103] To quantitatively analyze the detection results, Table 1 summarizes the evaluation results of each method on the LEVIR-CD and CDD datasets. In the LEVIR-CD dataset, FC-Siam-conv and FC-Siam-diff achieved relatively low F1 scores, but FC-Siam-conv's higher F1 score was due to its fusion of features from the encoding structure into the decoding structure. DSAMNet, employing a depth metric module to acquire deep features in the images, outperformed FC-Siam-conv and FC-Siam-diff in recall, F1 score, and intersection-over-union (IoU). For SUN-CD, the accuracy, recall, F1 score, and IoU were 90.20%, 87.26%, 88.71%, and 79.71%, respectively, with accuracy reaching the optimal value among all networks, and the other results being suboptimal. The present invention uses CDASNet to achieve the best values ​​in the experimental results for recall, F1 score and crossover ratio. The recall and F1 score are improved by 1.29% and 0.31% respectively compared with the second-best values, while the accuracy is slightly lower than the second-best values.

[0104] In the CDD dataset, FC-Siam-diff and DSAMNet have lower accuracy in F1 score than other networks. SUN-CD achieved suboptimal values ​​in all evaluation metrics of the experimental results. The accuracy, recall, F1 score, and crossover ratio of CDASNet used in this invention are 95.54%, 93.37%, 94.44%, and 89.47%, respectively, which are 0.87%, 1.65%, 1.26%, and 2.25% higher than the suboptimal values.

[0105] Table 1

[0106]

[0107] To investigate the impact of dense connections and CBAM on the network, ablation experiments were conducted on two datasets, and the specific results are shown in Table 2. In the LEVIR-CD dataset, adding CBAM improved the F1 score by 0.61% compared to the basic network, and the F1 score increased to 89.02% after using dense connections. In the CDD dataset, adding CBAM improved the F1 score by 2.49%, and the F1 score increased to 94.44% after using dense connections, representing a 3.89% improvement compared to the basic network. This demonstrates the necessity of each module in the network.

[0108] Table 2

[0109]

[0110] The remote sensing images selected at different times during the ablation experiment of this invention are shown in Figures 9(a) and 9(b), and their corresponding ground truth maps are shown in Figure 9(c). The change detection results using the base, base+CBAM, base+DC, and CDASNet models are respectively shown in Figures 9(a) and 9(b). Figure 9(d) , 9(e) As shown in Figures 9(f) and 9(g), the comparison reveals that CBAM achieves more accurate edge detection but fails to detect changes in small target features. DC, on the other hand, produces more edge misjudgments and is prone to adhesion, but it accurately identifies changes in small target features. Detection details are shown in Figure 9(h). CDASNet combines the advantages of both modules, achieving the best detection results.

[0111] The above method can be run as a computer program on a computer device to implement the change detection method described in this invention.

[0112] This invention proposes a CDASNet network model and performs remote sensing image change detection based on this model. By utilizing the input method of a Siamese network, it better adapts to the change detection task and adopts a novel dense connection method to fuse information between high- and low-level feature layers, thereby improving the network's ability to extract small targets. Simultaneously, an attention mechanism is introduced into the network to enhance attention to changed regions, improving the ability to distinguish boundaries in complex backgrounds. Experiments were conducted on the LEVIR-CD and CDD datasets, and the results show that CDASNet can adapt to remote sensing image change detection tasks in complex backgrounds and outperforms current mainstream change detection networks.

Claims

1. A method for detecting changes in remote sensing images, characterized in that, The method includes the following steps: 1) Obtain the remote sensing image pair to be detected, namely the first remote sensing image and the second remote sensing image, which are of the same size; 2) Construct a change detection model, which includes an encoding part and a decoding part. The encoding part uses a Siamese network, the decoding part uses an upsampling network, and the encoding part and the decoding part are connected by an attention module. The twin network includes a first network branch and a second network branch. The first network branch and the second network branch respectively use N and N+1 layers of residual connection modules for feature extraction, where N is greater than or equal to 3. The first network branch is used to extract features from the first remote sensing image to obtain features of the first remote sensing image at different scales. The second network branch is used to extract features from the second remote sensing image to obtain features of the second remote sensing image at different scales. The first network branch and the second network branch are connected in a dense manner to superimpose and fuse the features of the first remote sensing image and the second remote sensing image at the same scale to obtain fused features at different scales. Dense connections include upsampling layers and stacking layers. The input of the upsampling layer of the Nth dense connection is the output of the (N+1)th residual connection module of the second network branch. The stacking layer of the Nth dense connection is used to stack the outputs of the upsampling layer of this dense connection, the outputs of the Nth residual connection module of the first network branch, and the outputs of the Nth residual connection module of the second network branch. The stacking result is the output of this dense connection. The input of the upsampling layer of the (N-1)th dense connection is the output of the Nth dense connection. The stacking layer of the (N-1)th dense connection is used to stack the outputs of the upsampling layer of this dense connection, the outputs of the (N-1)th residual connection module of the first network branch, and the outputs of the (N-1)th residual connection module of the second network branch. The stacking result is the output of the (N-1)th dense connection. The attention module is used to perform feature enhancement processing on the fused features at different scales respectively; the decoding part is used to perform feature recovery and fusion on the feature-enhanced features at different scales through an upsampling network to obtain the change detection result.

2. The remote sensing image change detection method according to claim 1, characterized in that, The residual connection module includes a first convolutional layer, a second convolutional layer, a first regularization layer, a second regularization layer, an activation layer, and a stacking module. The first convolutional layer is used to perform convolution operations on the input features. The output of the first convolutional layer is processed by the first regularization layer and the activation layer and then sequentially input into the second convolutional layer and the second regularization layer. After being convolved and regularized again, it is input into the stacking module, which is used to stack the features output by the second regularization layer and the features output by the first convolutional layer. The stacked features are the output of the residual connection module.

3. The remote sensing image change detection method according to claim 1, characterized in that, The attention module mentioned above is a convolutional attention module.

4. The remote sensing image change detection method according to claim 1, characterized in that, The first residual connection module of the first network branch takes the first remote sensing image as input and outputs a convolutional feature with the same scale as the first remote sensing image. The second residual connection module of the first network branch takes the output of the first residual connection module as input and outputs a convolutional feature with a scale half that of the first remote sensing image. The Nth residual connection module of the first network branch takes the output of the (N-1)th residual connection module as input and outputs a convolutional feature with a scale half that of the first remote sensing image. N-1 The input to the first residual connection module of the second network branch is the second remote sensing image, and the output is a convolutional feature with the same scale as the second remote sensing image. The input to the second residual connection module of the second network branch is the output of the first residual connection module, and the output is a convolutional feature with a scale of 1 / 2 that of the first remote sensing image. The input to the (N+1)th residual connection module of the second network branch is the output of the Nth residual connection module, and the output is a convolutional feature with a scale of 1 / 2 that of the second remote sensing image. N The convolutional features.

5. The remote sensing image change detection method according to claim 1, characterized in that, The value of N is 4.

6. The remote sensing image change detection method according to any one of claims 3-5, characterized in that, The change detection model is trained using a hybrid loss function of weighted cross-entropy and Dice Loss.

7. The remote sensing image change detection method according to claim 6, characterized in that, The weighted cross-entropy loss function used is: in Represents a specific pixel; These are weighting coefficients; , These represent the pixel points respectively. Predicted labels and pixels The probability of belonging to the change category.

8. The remote sensing image change detection method according to claim 6, characterized in that, The loss function used in Dice Loss is: in Represents the true value of the sample. This represents the predicted value of the sample.

Images