A remote sensing change detection method based on improved spatial pooling pyramid

Abstract

A remote sensing change detection method based on improved spatial pooling pyramid includes the following steps: Step 1: preprocessing two-period remote sensing images, using X1 and X2 to represent the first and second period images respectively, and generating a differential image, recorded as DI; then cascading X1, X2 and DI, and inputting the result into a residual neural network with integrated dilated convolution; Step 2: in Downsample and extract features to obtain high-level features X H and low-level features X L ; Step 3: Use improved spatial pooling pyramid for high-level features X H Extract features and concatenate the extracted features to form a feature pyramid, denoted as FPYR; Step 4: Upsample FPYR; and upsample the low-level features X output in step 2 L Cascade to the upsampled mirror feature layer through short connections, and then solve the change probability map through the SoftMax layer; Step 5: Use the improved cross entropy loss function to train the network parameters through back propagation based on the change probability map and the true change map, and perform change detection through the trained network parameters; Perform remote sensing change detection through the above steps.

Description

Technical Field

[0001] The invention belongs to the field of remote sensing technology, and relates to a remote sensing change detection technology, and in particular to a remote sensing change detection technology based on an improved spatial pooling pyramid. Background Art

[0002] With the continuous development of deep learning technology, it has been widely used in the field of remote sensing. Compared with traditional machine learning technology, remote sensing change detection based on deep learning has higher accuracy, especially for high-resolution remote sensing images.

[0003] However, existing deep learning-based change detection technologies, such as "End-to-End Change Detection for High Resolution Satellite Images Using Improved UNet++", use two-phase images as input and fail to fully consider their differential image space. In addition, the change detection problem is usually a serious class imbalance problem, and the changed class area is generally much smaller than the unchanged class area. In order to solve the above problems and effectively utilize the multi-scale characteristics of high spatial resolution remote sensing images, the present invention proposes a remote sensing change detection technology based on improved spatial pooling pyramid. Summary of the invention

[0004] Aiming at the problems that the existing deep learning change detection technology is easily affected by pseudo-changes and the edge detection of the changed area is incomplete, the present invention proposes a high-resolution remote sensing image change detection method based on an improved spatial pooling pyramid.

[0005] A remote sensing change detection method based on improved spatial pooling pyramid includes the following steps:

[0006] Step 1: Preprocess the two-period remote sensing images, use X1 and X2 to represent the first period image and the second period image respectively, and use the difference method to generate the difference image of X1 and X2, recorded as DI; then cascade X1, X2 and DI, recorded as X in , Indicates cascade operation, and finally X in Input to the residual neural network with integrated dilated convolution;

[0007] Step 2: Use the residual neural network with integrated dilated convolution to train X in Downsample and extract features to obtain high-level features X H and low-level features X L ;

[0008] Step 3: For high-level features X HConvolution, position and channel attention mechanisms, three types of dilated convolutions with different dilated rates, and global maximum pooling are used to extract features of different scales, and the extracted features are cascaded to form a feature pyramid, denoted as FPYR.

[0009] Step 4: Upsample FPYR; and convert the low-level feature X output in step 2 L It is cascaded to the upsampled mirror feature layer through a short connection, and then the change probability map is solved through the SoftMax layer;

[0010] Step 5: Use the improved cross entropy loss function to train the network parameters through back propagation based on the change probability map and the true change map, and perform change detection based on the trained network parameters;

[0011] Through the above steps, remote sensing change detection is performed.

[0012] In step 2, the following steps are specifically included:

[0013] Step 2-1: Input the network into X in Send it to the convolution layer Conv to initially extract features and increase the number of feature channels;

[0014] Step 2-2: The output of the convolutional layer Conv is passed through the batch normalization layer, the activation layer and the global maximum pooling layer in sequence, so that the network has nonlinear expression capabilities;

[0015] Step 2-3: Get the low-level features X by integrating the residual neural network with dilated convolution L and high-level features X H .

[0016] In step 3, the high-level features X H Perform multi-scale feature extraction, specifically:

[0017] Use convolution, position and channel attention mechanisms, three different dilation rates of dilated convolution, and global maximum pooling to extract features of different scales;

[0018] The output of the position attention mechanism is denoted as PAM(X H ), the output of the channel attention mechanism is denoted as CAM(X H ), the output of the convolution (including the convolution layer, batch normalization layer and Relu function activation layer) is recorded as Conv1(X H ), the outputs of the three types of dilated convolutions (each convolution includes a convolution layer, a batch normalization layer, and a Relu function activation layer) are recorded as AConv1(X H ), AConv2(X H ), AConv3(X H ), the output of the global maximum pooling is recorded as Pool(XH ), and cascade the output to obtain a cascade feature pyramid FPYR, namely:

[0019] The above PAM(X H The specific method of obtaining ) is as follows: first, the high-level feature X obtained in step 2 is H Processed by the position attention mechanism, and then using the convolution layer ConvP to adjust the number of feature channels;

[0020] The CAM (X H The specific method of obtaining ) is as follows: first, the high-level feature X obtained in step 2 is H It is processed through the channel attention mechanism, and then the convolution layer ConvC is used to adjust the number of feature channels.

[0021] The above residual neural network with integrated dilated convolution is composed of multiple dilated convolution residual blocks connected in sequence. The output of the nth residual block Blockn is denoted as B n (X), then the low-level feature X L is the output of the first residual block, the high-level feature X H is the output of the last residual block.

[0022] In step 5, the proposed network structure is trained using the improved cross entropy loss function, which is:

[0023]

[0024] Where N represents the total number of pixels, u and c represent the unchanged class and the changed class respectively, and p ik is the probability that the model predicts that pixel i belongs to category k, y ik is a Boolean variable equal to 0 or 1, indicating the category l to which pixel i belongs. i calculate: w k is the weight of class k, k∈{u,c}, through the formula OK. u and a c They represent the proportion of unchanged class u and changed class c respectively, and β is the balance coefficient.

[0025] The network structure based on improved spatial pooling pyramid for remote sensing image processing includes a specially improved spatial pooling pyramid module: composed of convolution units, position and channel attention mechanism units, three kinds of hole convolution units with different hole rates, and global maximum pooling units; the input of the improved spatial pooling pyramid module is connected to the first output of the hole residual module, and the first output of the hole residual module is a high-level feature X H, the second output of the hole residual module is the low-level feature X L ; Improve the connection between the output of the spatial pooling pyramid module and the input of the first upsampling module; low-level feature X L Cascade with the features output by the first upsampling module to obtain a new feature map after cascading.

[0026] The above-mentioned hole residual module is composed of multiple hole convolution residual blocks connected in sequence. The output of the nth layer residual block Blockn is denoted as B n (X), then the low-level feature X L is the output of the first residual block, the high-level feature X H is the output of the last residual block;

[0027] The input of the void residual module includes the remote sensing image X1 of the first period, the remote sensing image X2 of the second period, and the difference image DI between the remote sensing image X1 of the first period and the remote sensing image X2 of the second period. The three are cascaded to form the network input X in , X in =X1⊕X2⊕DI.

[0028] Network Input X in It passes through the convolution layer Conv, batch normalization layer, activation layer, and global maximum pooling layer in sequence, and the output of the global maximum pooling layer is connected to the input of the void residual module.

[0029] The feature map size output by the first upsampling module is similar to the low-level feature X L The sizes are the same; after cascading, the new feature map passes through at least one convolution layer and is input into the second upsampling module, and the size of the feature map output by the second upsampling module is the same as the image size of the original input remote sensing image.

[0030] Compared with the prior art, the present invention has the following technical effects:

[0031] The technical solution proposed by the present invention considers the original image space and the differential image space of two phases of high-resolution remote sensing images at the same time; considers the multi-scale characteristics of high-resolution remote sensing images by improving the spatial pooling pyramid to ensure the detection capability of targets with large changes; enhances the detection capability of detailed change information by using the attention mechanism in parallel; and proposes a new loss function to adapt to the imbalance problem of changed and unchanged classes. Through the above measures, the present invention can obtain better change detection results. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0033] Figure 1 : It is the main framework structure of the network proposed in the present invention.

[0034] Figure 2 : is a schematic diagram of the residual block structure of the residual neural network with integrated hole convolution adopted in the present invention.

[0035] Figure 3 : is a schematic diagram of the residual unit used in the present invention.

[0036] Figure 4 : is a diagram of the improved spatial pooling pyramid structure proposed by the present invention.

[0037] Figure 5 : is a schematic diagram of the position attention mechanism structure adopted in the present invention.

[0038] Figure 6 : This is a schematic diagram of the channel attention mechanism structure used in the present invention.

[0039] Figure 7 : is a schematic diagram of the upsampling technology DUpsampling adopted in the present invention.

[0040] Figure 8 : are example images of experimental data used in an embodiment of the present invention, including images of the T1 and T2 periods and reference images of their changes.

[0041] Fig. 9 : is a graph showing the change detection results of six comparison techniques and the present invention on example images. DETAILED DESCRIPTION

[0042] The present invention includes a remote sensing change detection method based on an improved spatial pooling pyramid, which includes the following steps:

[0043] Step 1: Preprocess the two-period remote sensing images, use X1 and X2 to represent the first period image and the second period image respectively, and use the difference method to generate the difference image of X1 and X2, recorded as DI; then cascade X1, X2 and DI, recorded as X in , Indicates cascade operation, and finally X in Input to the residual neural network with integrated dilated convolution;

[0044] Step 2: Use the residual neural network with integrated dilated convolution to train X in Downsample and extract features to obtain high-level features X H and low-level features X L ;

[0045] Step 3: For high-level features X HConvolution, position and channel attention mechanisms, three types of dilated convolutions with different dilated rates, and global maximum pooling are used to extract features of different scales, and the extracted features are cascaded to form a feature pyramid, denoted as FPYR.

[0046] Step 4: Upsample FPYR; and convert the low-level feature X output in step 2 L It is cascaded to the upsampled mirror feature layer through a short connection, and then the change probability map is solved through the SoftMax layer;

[0047] Step 5: Use the improved cross entropy loss function to train the network parameters through back propagation based on the change probability map and the true change map, and perform change detection based on the trained network parameters;

[0048] Through the above steps, remote sensing change detection is performed.

[0049] In step 2, the following steps are specifically included:

[0050] Step 2-1: Input the network into X in Send it to the convolution layer Conv to initially extract features and increase the number of feature channels;

[0051] Step 2-2: The output of the convolutional layer Conv is passed through the batch normalization layer, the activation layer and the global maximum pooling layer in sequence to make the network have nonlinear expression capabilities;

[0052] Step 2-3: Get the low-level features X by integrating the residual neural network with dilated convolution L and high-level features X H .

[0053] In step 3, the high-level features X H Perform multi-scale feature extraction, specifically:

[0054] Use convolution, position and channel attention mechanisms, three different dilation rates of dilated convolution, and global maximum pooling to extract features of different scales;

[0055] The output of the position attention mechanism is denoted as PAM(X H ), the output of the channel attention mechanism is denoted as CAM(X H ), the output of the convolution (including the convolution layer, batch normalization layer and Relu function activation layer) is recorded as Conv1(X H ), the outputs of the three types of dilated convolutions (each convolution includes a convolution layer, a batch normalization layer, and a Relu function activation layer) are recorded as AConv1(X H ), AConv2(X H ), AConv3(X H ), the output of the global maximum pooling is recorded as Pool(XH ), and cascade the output to obtain a cascade feature pyramid FPYR, namely:

[0056] The PAM (X H The specific method of obtaining ) is as follows: first, the high-level feature X obtained in step 2 is H Processed by the position attention mechanism, and then using the convolution layer ConvP to adjust the number of feature channels;

[0057] The CAM (X H The specific method of obtaining ) is as follows: first, the high-level feature X obtained in step 2 is H It is processed through the channel attention mechanism, and then the convolution layer ConvC is used to adjust the number of feature channels.

[0058] The residual neural network with integrated dilated convolution is composed of multiple dilated convolution residual blocks connected in sequence. The output of the nth layer residual block Blockn is recorded as B n (X), then the low-level feature X L is the output of the first residual block, the high-level feature X H is the output of the last residual block.

[0059] In step 5, the proposed network structure is trained using the improved cross entropy loss function, which is:

[0060]

[0061] Where N represents the total number of pixels, u and c represent the unchanged class and the changed class respectively, and p ik is the probability that the model predicts that pixel i belongs to category k, y ik is a Boolean variable equal to 0 or 1, indicating the category l to which pixel i belongs. i calculate: w k is the weight of class k, k∈{u,c}, through the formula OK. u and a c They represent the proportion of unchanged class u and changed class c respectively, and β is the balance coefficient.

[0062] The present invention also includes a network structure based on an improved spatial pooling pyramid for remote sensing image processing, which includes an improved spatial pooling pyramid module: composed of a convolution unit, a position and channel attention mechanism unit, three types of hole convolution units with different hole rates, and a global maximum pooling unit; the input of the improved spatial pooling pyramid module is connected to the first output of the hole residual module, and the first output of the hole residual module is a high-level feature X H, the second output of the hole residual module is the low-level feature X L ; Improve the connection between the output of the spatial pooling pyramid module and the input of the first upsampling module; low-level feature X L Cascade with the features output by the first upsampling module to obtain a new feature map after cascading.

[0063] The hole residual module consists of multiple hole convolution residual blocks connected in sequence. The output of the nth layer residual block Blockn is recorded as B n (X), then the low-level feature X L is the output of the first residual block, the high-level feature X H is the output of the last residual block;

[0064] The input of the void residual module includes the remote sensing image X1 of the first period, the remote sensing image X2 of the second period, and the difference image DI between the remote sensing image X1 of the first period and the remote sensing image X2 of the second period. The three are cascaded to form the network input X in , X in =X1⊕X2⊕DI.

[0065] Network Input X in It passes through the convolution layer Conv, batch normalization layer, activation layer, and global maximum pooling layer in sequence, and the output of the global maximum pooling layer is connected to the input of the void residual module.

[0066] The feature map size output by the first upsampling module is similar to the low-level feature X L The sizes are the same; after cascading, the new feature map passes through at least one convolution layer and is input into the second upsampling module, and the size of the feature map output by the second upsampling module is the same as the image size of the original input remote sensing image.

[0067] In order to facilitate those skilled in the art to further understand the present invention, further description is given as follows:

[0068] In the embodiment, the change detection dataset disclosed in the document "S. Ji, S. Wei, and M. Lu, Fully Convolutional Networks for Multisource Building Extraction From an Open Aerial and Satellite Imagery Data Set, IEEE Transactions on Geoscience and Remote Sensing, vol. 57, pp. 574-586, 2019" is used for experiments. The dataset contains two groups of high-resolution remote sensing images, each group of images contains remote sensing images of two periods and real change maps of remote sensing images of two periods. The size of the first group of images is 21243*15354 pixels, and the size of the second group of images is 11265*15354 pixels. To facilitate network training, the two groups of large images are divided into small image groups of 256*256 pixels, and the image groups that are completely unchanged and completely changed are removed, leaving a total of 1863 groups of 256*256 small images, of which 1250 are training sets and 613 are test sets;

[0069] like Figure 1 The figure shows the main framework structure of the network proposed in the present invention. A remote sensing change detection method based on improved spatial pooling pyramid in the present invention comprises the following steps:

[0070] Step 1: Preprocess the two-period high-resolution remote sensing images, including registration and relative radiation correction. Use X1 and X2 to represent the first and second period images respectively, and use the difference method to generate the difference image of X1 and X2, recorded as DI. The calculation formula of DI is as follows:

[0071] DI=|X1-X2|

[0072] Then cascade X1, X2 and DI, denoted as X in , Indicates cascade operation, and finally X in input into the proposed network.

[0073] Step 2: Use the residual neural network AtrousResNet50 with integrated dilated convolution to train the network input X in Downsample and extract features to obtain high-level features X H and low-level features X L .

[0074] First, input the network into X inThe convolutional layer Conv is sent to extract features and increase the number of feature channels. Then, the batch normalization layer (BN layer), activation layer (using Relu function) and global maximum pooling layer are used to make the network have nonlinear expression ability and avoid the degradation caused by continuous convolution. Finally, the low-level feature X is obtained by integrating the residual neural network of the dilated convolution. L and high-level features X H .

[0075] The residual block structure of the residual neural network with integrated dilated convolution used in this embodiment is as follows: Figure 2 As shown, it is composed of four dilated convolution residual blocks, each of which is composed of multiple residual units. The specific parameters such as the number of residual units and the structure of residual units can be adjusted according to the specific application. In this embodiment, each residual unit except the middle layer is composed of a normal convolution with a step size of 1 and a dilation rate of 0. Figure 3 Give an example of a residual unit structure.

[0076] The output of the nth layer residual block Blockn is recorded as B n (X), then the low-level feature X L is the output of the first residual block, i.e. X L =B1(X in ), high-level features X H is the output of the fourth residual block, i.e. X H =B4(B3(B2(B1(X in ))))) Through the downsampling step, we can get the low-level features X L and high-level features X H , which can contain information of different scales. In this step, the number of residual blocks, residual unit structure, hole rate, convolution kernel size and number can be adjusted according to the specific application. The detailed parameters of the convolution layer and residual unit and the number of residual blocks used in this embodiment are shown in Table 1:

[0077] Table 1 Detailed parameters of convolutional layers and residual units

[0078]

[0079] Step 3: For high-level features, use 1*1 convolution, spatial and channel attention mechanisms, three types of dilated convolutions with different dilation rates, and global maximum pooling to extract features of different scales, and cascade them to form a feature pyramid. Figure 4 The specific structure of step 3 of this embodiment is given.

[0080] The obtained cascade pyramid features can fully consider the multi-scale characteristics and spatial context information of high-resolution remote sensing images. The output of the 1*1 convolutional layer Conv1 is denoted as Conv1(X H), the output of the spatial attention mechanism is denoted as PAM(X H ), the output of the channel attention mechanism is denoted as CAM(X H ), the outputs of the three atrous convolutional layers AConv1, AConv2, and AConv3 are denoted as AConv1(X H ), AConv2(X H ), AConv3(X H ), the output of the global maximum pooling layer is denoted as Pool(X H ), where X H is the high-level feature obtained in step 2. The above outputs are cascaded to obtain the cascade feature pyramid FPYR, that is,

[0081]

[0082] The present invention enhances the representation ability of detailed targets and object boundaries by cascading spatial attention (PAM) and channel attention features (CAM). The position attention mechanism is as follows: Figure 5 As shown, the channel attention mechanism is as follows Figure 6 PAM(X H The specific method of obtaining ) is as follows: first, the high-level feature X obtained in step 2 is H pass Figure 5 The spatial attention mechanism shown in Figure 1 is then used to adjust the number of feature channels using the convolutional layer ConvP. H The specific method of obtaining ) is as follows: first, the high-level feature X obtained in step 2 is H pass Figure 6 The channel attention mechanism shown is processed, and then the convolution layer ConvC is used to adjust the number of feature channels.

[0083] Table 2 shows the convolution kernel size, number, dilation rate and step size of the convolution layers Conv1, AConv1, AConv2, AConv3, ConvP and ConvC used in step 3 of this embodiment. The above convolution layers are followed by a BN layer and a Relu function activation layer. It should be pointed out that the convolution kernel size, number, dilation rate, step size, maximum pooling method and number of dilated convolution layers used in step 3 can be adjusted according to specific applications.

[0084] Table 2 Specific parameters of convolutional layers Conv1, AConv1, AConv2, AConv3, ConvP and ConvC

[0085]

[0086] Step 4: Use the improved upsampling operation to upsample the cascade feature pyramid in step 3 and convert the low-level features X output in step 2L It is cascaded to the upsampled mirror feature layer through a short connection, and finally the change probability map is solved through the SoftMax layer;

[0087] First, the cascade feature pyramid FPYR in step 3 is sent to the convolution layer Conv2 to reduce the number of feature channels, and then the nonlinear expression ability is increased through the BN layer and the activation layer (using the Relu function). Then, the improved upsampling technology DUpsampling is used for four times upsampling, and the output feature map size is the same as the low-level feature X L Similarly, the two are concatenated as a new feature map.

[0088] The cascaded feature map is passed through two convolutional layers Conv3 and Conv4 (including BN layer and Relu function activation layer) to adjust the number of channels, where the number of convolution kernels in the convolution layer can be adjusted according to the specific application. Then the improved upsampling technology DUpsampling is used for four times upsampling, and the output feature map size is the same as the input image. Finally, the output feature map is passed through the SoftMax layer to solve the change probability map.

[0089] Improved upsampling technology DUpsampling is an improvement on traditional linear interpolation. The structure is as follows Figure 7 As shown. In this embodiment, DUpsampling is used to replace linear upsampling. DUpsampling generates the padding required in the process of restoring feature resolution through convolution, which is conducive to the restoration of feature map detail information, especially for high-resolution remote sensing images with more details. The detailed parameters of the convolution layer for adjusting the number of channels in this step are shown in Table 3:

[0090] Table 3 Specific parameters of convolutional layers Conv2, Conv3 and Conv4

[0091]

[0092] Step 5: Use the improved cross entropy loss function to calculate the loss based on the change probability map and the true change map obtained in step 4 above, and iteratively train the parameters in the network structure through back propagation until the iteration stop condition is met, and save the parameters when the iteration stops to obtain the change detection map.

[0093] In practical applications, the area that has changed is often much smaller than the area that has not changed. The remote sensing change detection problem is usually a problem in which the weights of the changed class and the unchanged class are extremely unbalanced. To address this problem, the present invention proposes a new adaptive weighted cross entropy loss function, and uses the proposed loss function to calculate the loss of the change probability map output by the network, and optimizes the parameters in the network through back propagation.

[0094] Based on the cross entropy loss function, this step designs a weighted adaptive cross entropy loss function suitable for the problem of class imbalance. The weighted adaptive calculation method is based on the proportion of the changed class and the unchanged class. The calculation formula of the proposed adaptive weighted cross entropy function is:

[0095]

[0096] Where N represents the total number of pixels, u and c represent the unchanged class and the changed class respectively, and p ik is the probability that the model predicts that pixel i belongs to category k, y ik is a Boolean variable equal to 0 or 1, indicating the category l to which pixel i belongs. i calculate: w k is the weight of class k, k∈{u,c}, through the formula OK. u and a c They represent the proportion of unchanged class u and changed class c respectively. Generally, a u Much larger than 1 / 2, a c β is a balance coefficient, which can ensure that the two types of weights are not too unbalanced. In the embodiment of the present invention, β=1 / 10.

[0097] In order to verify the change detection effect of the present invention, the present invention is compared with 6 advanced deep learning change detection technologies. The 6 compared technologies are: Fully Convolutional Siamese Neural Network (FC-Siam-Conv), DeepLab v3+, Dual Attention Neural Network (DANet), Improved U-Network (Unet++), Multi-Output Fusion Improved U-Network (Unet++_MSOF), and Improved U-Network Based on Difference Map (DifUNet++). Four widely used quantitative indicators are used to evaluate the performance of different change detection technologies, namely accuracy, precision, recall and F1 value.

[0098] Please see Fig. 9 , and the change detection result diagrams of the present invention and the comparative technology are given. They are respectively the fully convolutional twin neural network (FC-Siam-Conv), DeepLab v3+, dual attention neural network (DANet), improved U-type network (Unet++), multi-output fusion improved U-type network (Unet++_MSOF), improved U-type network based on difference map (DifUNet++) and the change detection result diagram of the present invention. Table 4 gives four quantitative indicators of different change detection results.

[0099] Table 4 Quantitative indicators of change detection results

[0100]

[0101] Comparing the change detection result graph and the quantitative statistical results, the change detection effect of the present invention is significantly better than the change detection results of several other advanced deep learning change detection technologies. In this embodiment, it can be seen that the present invention has achieved better results in terms of both the completeness of the change and the refinement of the edge. As can be seen from Table 4, the change detection results of the present invention are better than other change detection technologies in terms of the three precision indicators of accuracy, recall rate and F1 value. For example, the F1 value of the present invention is 0.9046, which is 0.0521, 0.0329, 0.0433, 0.0407, 0.0605, and 0.0284 higher than other methods, respectively.

Claims

1. A remote sensing change detection method based on improved spatial pooling pyramid, characterized in that: It includes the following steps: Step 1: Preprocess the two-period remote sensing images, use X1 and X2 to represent the first period image and the second period image respectively, and use the difference method to generate the difference image of X1 and X2, recorded as DI; then cascade X1, X2 and DI, recorded as X in , Indicates cascade operation, and finally X in Input to the residual neural network with integrated dilated convolution; Step 2: Use the residual neural network with integrated dilated convolution to train X in Downsample and extract features to obtain high-level features X H and low-level features X L ; Step 3: For high-level features X H Convolution, position and channel attention mechanisms, three types of dilated convolutions with different dilated rates, and global maximum pooling are used to extract features of different scales, and the extracted features are cascaded to form a feature pyramid, denoted as FPYR. Step 4: Upsample FPYR; and convert the low-level feature X output in step 2 L It is cascaded to the upsampled mirror feature layer through a short connection, and then the change probability map is solved through the SoftMax layer; Step 5: Use the improved cross entropy loss function to train the network parameters through back propagation based on the change probability map and the true change map, and perform change detection using the trained network parameters; Remote sensing change detection is performed through the above steps.

2. The method according to claim 1, characterized in that In step 2, the following steps are specifically included: Step 2-1: Input the network into X in Send it to the convolution layer Conv to initially extract features and increase the number of feature channels; Step 2-2: The output of the convolutional layer Conv is passed through the batch normalization layer, the activation layer and the global maximum pooling layer in turn to make the network have nonlinear expression capabilities; Step 2-3: Get the low-level features X by integrating the residual neural network with dilated convolution L and high-level features X H .

3. The method according to claim 1, characterized in that In step 3, the high-level features X H Perform multi-scale feature extraction, specifically: Use convolution, position and channel attention mechanisms, three different dilation rates of dilated convolution, and global maximum pooling to extract features of different scales; The position attention mechanism output is denoted as PAM(X H ), the output of the channel attention mechanism is denoted as CAM(X H ); the output of the convolution is denoted as Conv1(X H ), the convolution includes convolution layer, batch normalization layer and Relu function activation layer; the outputs of the three types of dilated convolution are denoted as AConv1(X H ), AConv2(X H ), AConv3(X H ), the dilated convolution includes the convolution layer, batch normalization layer and Relu function activation layer; the output of the global maximum pooling is denoted as Pool(X H ), and cascade the output to obtain a cascade feature pyramid FPYR, namely:

4. The method according to claim 3, characterized in that The PAM (X H The specific method of obtaining ) is as follows: first, the high-level feature X obtained in step 2 is H Processed by the position attention mechanism, and then using the convolution layer ConvP to adjust the number of feature channels; The CAM (X H The specific method of obtaining ) is as follows: first, the high-level feature X obtained in step 2 is H It is processed through the channel attention mechanism, and then the convolution layer ConvC is used to adjust the number of feature channels.

5. The method according to claim 2, characterized in that: The residual neural network with integrated dilated convolution is composed of multiple dilated convolution residual blocks connected in sequence. The output of the nth layer residual block Blockn is recorded as B n (X), then the low-level feature X L is the output of the first residual block, the high-level feature X H is the output of the last residual block.

6. The method according to claim 1, characterized in that In step 5, the loss is calculated based on the change probability map and the true change map using the improved cross entropy loss function. The improved cross entropy loss function is: Where N represents the total number of pixels, u and c represent the unchanged class and the changed class respectively, and p ik is the probability that the model predicts that pixel i belongs to category k, y ik is a Boolean variable equal to 0 or 1, indicating the category l to which pixel i belongs. i calculate: w k is the weight of class k, k∈{u,c}, through the formula OK, where a u and a c They represent the proportion of unchanged class u and changed class c respectively, and β is the balance coefficient.

Images