Ultra-lightweight remote sensing image change detection method and system

Abstract

The invention belongs to the technical field of remote sensing image change detection, and particularly relates to an ultra-light remote sensing image change detection method and system, and the method comprises the steps: firstly obtaining a dual-time-phase remote sensing image of a target detection region, and carrying out the preprocessing of the dual-time-phase remote sensing image; inputting the preprocessed dual-temporal remote sensing image into an ultra-light twin encoder sharing weight parameters, and extracting multi-scale high-dimensional semantic features of the dual-temporal remote sensing image; performing cross-scale semantic fusion on the multi-scale high-dimensional semantic features of the dual-temporal remote sensing image by using a space-channel attention module to generate multi-scale semantic features of the dual-temporal remote sensing image; and carrying out differencing on the multi-scale semantic features of the dual-temporal remote sensing image, and inputting a generated multi-scale change feature map into an ultra-light decoder for decoding to obtain a binary image for change detection. According to the method, the parameter quantity and the calculation quantity are greatly reduced while the characteristic characterization capability is maintained, so that the detection speed is improved on the basis of ensuring the detection precision.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image change detection, and in particular relates to an ultra-lightweight remote sensing image change detection method and system. Background Art

[0002] Remote sensing image change detection refers to the use of remote sensing technology and image processing methods to analyze remote sensing images acquired at different times to determine the changes in the surface or target at two or more time points. This technology has a wide range of applications in land use / land cover change monitoring, urban expansion analysis, environmental monitoring and other fields.

[0003] Traditional remote sensing image change detection methods mainly rely on pixel-level analysis, that is, by comparing the pixel differences of dual-phase remote sensing images and setting appropriate thresholds to detect changes. However, these traditional methods are sensitive to factors such as image illumination and sensor differences, and are prone to a large number of false detections or missed detections, which limits their practical application effects. In recent years, with the rapid development of deep learning technology, convolutional neural networks (CNNs) and deep learning networks (Transformers) have been widely used in the field of computer vision due to their powerful feature extraction capabilities, resulting in the emergence of many excellent CNN and Transformer-based models in change detection tasks. With the rise of the state space model (SSM), it quickly attracted the favor of researchers due to its linear complexity and efficient feature extraction capabilities. Mamba is a new selective SSM proposed on the basis of SSM, which overcomes the time-invariant defect of the original SSM, improves the static content perception ability of SSM, and selectively propagates or ignores information. Mamba is introduced into the field of computer vision, and the visual Mamba model (VMamba) is obtained. With the same number of parameters, it achieves better detection results than CNN and Transformer models.

[0004] Although the VMamba model achieved the best results with similar parameter amounts, its parameter amount has not been significantly reduced, and its inference speed is slow and the model calculation amount is high, which still makes it difficult to adapt to the low-complexity and fast processing requirements in real scenarios. In summary, the current change detection method based on deep learning is still in the development stage, and its application scenarios are severely restricted by the model parameter amount and calculation complexity, and there is still a certain gap with the actual application and terminal deployment requirements. Summary of the invention

[0005] The purpose of the present invention is to provide a small-sized, high-speed ultra-lightweight remote sensing image change detection method and system in view of the above-mentioned problems existing in the prior art.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] In a first aspect, the present invention provides an ultra-lightweight remote sensing image change detection method, the method comprising:

[0008] S1, obtaining a dual-phase remote sensing image of the target detection area and preprocessing it;

[0009] S2, input the preprocessed dual-temporal remote sensing image into an ultra-lightweight twin encoder with shared weight parameters to extract multi-scale high-dimensional semantic features of the dual-temporal remote sensing image;

[0010] S3, using the spatial-channel attention module to perform cross-scale semantic fusion on the multi-scale high-dimensional semantic features of the dual-temporal remote sensing images to generate multi-scale semantic features of the dual-temporal remote sensing images;

[0011] S4, performing subtraction on the multi-scale semantic features of the dual-temporal remote sensing images to generate a multi-scale change feature map;

[0012] S5. Input the multi-scale change feature map into the ultra-lightweight decoder for decoding to obtain a binary image for change detection.

[0013] The ultra-lightweight twin encoder described in S2 includes two ultra-lightweight encoders that share weight parameters, and the ultra-lightweight encoder is formed by connecting three convolution blocks in front and three grouped visual GVM blocks in series. Each of the convolution blocks includes a convolution layer, a group normalization layer, a maximum pooling layer and a GELU activation function layer, and each of the grouped visual GVM blocks includes a grouped visual Mamba layer, a group normalization layer, a maximum pooling layer and a GELU activation function layer.

[0014] The ultra-lightweight decoder in S5 is formed by connecting three preceding grouped visual GVM blocks in series with three succeeding convolution blocks. The last convolution block contains a convolution layer, and the number of output channels is set to 2, which are responsible for outputting background and change categories respectively. The remaining convolution blocks include a convolution layer, a group normalization layer, a maximum pooling layer and a GELU activation function layer; each of the grouped visual GVM blocks includes a grouped visual Mamba layer, a group normalization layer and a GELU activation function layer.

[0015] Each block in the ultra-lightweight decoder has two inputs, namely, the decoded features output by the previous block and the deep semantic features weighted by the ultra-lightweight twin encoder and the space-channel attention module and then jump-connected; wherein, the space-channel attention module is formed by a preceding space attention submodule and a succeeding channel attention submodule connected in series.

[0016] The data preprocessing described in S1 includes random cropping, random flipping, random phase swapping, and data normalization.

[0017] In a second aspect, the present invention provides an ultra-lightweight remote sensing image change detection system, the system comprising:

[0018] An image preprocessing module is used to obtain dual-phase remote sensing images of the target detection area and preprocess them;

[0019] The encoding module is used to input the preprocessed dual-temporal remote sensing images into an ultra-lightweight twin encoder with shared weight parameters to extract multi-scale high-dimensional semantic features of the dual-temporal remote sensing images;

[0020] A cross-scale semantic fusion module is used to perform cross-scale semantic fusion on the multi-scale high-dimensional semantic features of dual-temporal remote sensing images using a spatial-channel attention module to generate multi-scale semantic features of dual-temporal remote sensing images;

[0021] The change detection module is used to perform subtraction on the multi-scale semantic features of the dual-temporal remote sensing images and generate a multi-scale change feature map;

[0022] The decoding module is used to input the multi-scale change feature map into the ultra-lightweight decoder for decoding to obtain a binary image, and use the binary image for change detection.

[0023] The ultra-lightweight twin encoder includes two ultra-lightweight encoders that share weight parameters. The ultra-lightweight encoder is formed by connecting three convolution blocks in front and three grouped visual GVM blocks in series. Each of the convolution blocks includes a convolution layer, a group normalization layer, a maximum pooling layer and a GELU activation function layer. Each of the grouped visual GVM blocks includes a grouped visual Mamba layer, a group normalization layer, a maximum pooling layer and a GELU activation function layer.

[0024] The ultra-lightweight decoder is formed by connecting three preceding grouped visual GVM blocks in series with three succeeding convolution blocks. The last convolution block contains a convolution layer, and the number of output channels is set to 2, which are responsible for outputting background and change categories respectively. The remaining convolution blocks include a convolution layer, a group normalization layer, a maximum pooling layer and a GELU activation function layer; each of the grouped visual GVM blocks includes a grouped visual Mamba layer, a group normalization layer and a GELU activation function layer.

[0025] Each block in the ultra-lightweight decoder has two inputs, namely, the decoded features output by the previous block and the deep semantic features weighted by the ultra-lightweight twin encoder and the space-channel attention module and then jump-connected; wherein, the space-channel attention module is formed by a preceding space attention submodule and a succeeding channel attention submodule connected in series.

[0026] The data preprocessing includes random cropping, random flipping, random phase exchange, and data normalization.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] 1. An ultra-lightweight remote sensing image change detection method of the present invention first obtains a dual-phase remote sensing image of a target detection area and preprocesses it; the preprocessed dual-phase remote sensing image is input into an ultra-lightweight twin encoder with a shared weight parameter to extract multi-scale high-dimensional semantic features of the dual-phase remote sensing image; a spatial-channel attention module is used to perform cross-scale semantic fusion on the multi-scale high-dimensional semantic features of the dual-phase remote sensing image to generate multi-scale semantic features of the dual-phase remote sensing image; the multi-scale semantic features of the dual-phase remote sensing image are subtracted to generate a multi-scale change feature map; the multi-scale change feature map is input into an ultra-lightweight decoder for decoding to obtain a binary image for change detection; on the one hand, the above method designs a multi-stage grouping view The GVM block effectively reduces the feature dimension, thereby constructing an ultra-lightweight twin encoder and ultra-lightweight decoder, which greatly reduces the amount of parameters and calculations while maintaining the feature representation capability, thereby improving the detection speed on the basis of ensuring detection accuracy; on the other hand, the spatial-channel attention module realizes cross-scale feature information interaction to reduce the feature semantic gap of the model at different scales; in addition, by using jump links between the ultra-lightweight twin encoder and the ultra-lightweight decoder, specifically, the shallow features output by the spatial-channel attention module are combined with the deep features output by the previous stage of the ultra-lightweight decoder and input into the current stage of the ultra-lightweight decoder to achieve cross-layer information aggregation, thereby further improving the detection accuracy. Therefore, the present invention greatly reduces the amount of parameters and calculations while maintaining the feature representation capability, thereby improving the detection speed on the basis of ensuring detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 The figure is a flow chart of the ultra-lightweight remote sensing image change detection method of the present invention.

[0030] Figure 2 It is a detection schematic diagram of the ultra-lightweight remote sensing image change detection method of the present invention.

[0031] Figure 3 Schematic diagram of the structure of the grouped visual GVM block of the ultra-lightweight encoder in the present invention.

[0032] Figure 4 Schematic diagram of the structure of the space-channel attention module in the present invention.

[0033] Figure 5 It is a structural schematic diagram of the ultra-lightweight remote sensing image change detection system of the present invention. DETAILED DESCRIPTION

[0034] In order to enable technicians in this field to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments.

[0035] Embodiment 1:

[0036] See also Figure 1 , Figure 2 , an ultra-lightweight remote sensing image change detection method, which is carried out in the following steps:

[0037] S1. Obtain a dual-phase remote sensing image of the target detection area, and preprocess it to form a data form that can be processed by a deep learning network. After preprocessing, divide the training set, validation set and test set into a ratio of 7:2:1; the preprocessing includes: slicing a large dual-phase remote sensing image and the corresponding label data according to a fixed size, and then performing random cropping, random flipping, random phase exchange, and data normalization in sequence; the random cropping refers to randomly selecting a sub-area in the input image for cropping and outputting; the random flipping refers to flipping the input image horizontally or vertically according to a random probability; the random phase exchange refers to randomly exchanging the remote sensing images of phase 1 and phase 2; the image normalization refers to subtracting the mean of each channel from the remote sensing image and dividing it by the standard deviation to standardize the image;

[0038] S2. Input the preprocessed dual-phase remote sensing images in the training set into an ultra-lightweight twin encoder with shared weight parameters to extract multi-scale high-dimensional semantic features of the dual-phase remote sensing images; the ultra-lightweight twin encoder includes two ultra-lightweight encoders with shared weight parameters, and the ultra-lightweight encoder is formed by connecting three convolution blocks in front and three grouped visual GVM blocks (GVM for short) in series, and each of the convolution blocks includes a convolution layer, a group normalization layer, a maximum pooling layer and a GELU activation function layer, such as Figure 3 As shown, each of the grouped visual GVM blocks includes a grouped visual Mamba layer, a group normalization layer, a maximum pooling layer and a GELU activation function layer, and the grouped visual Mamba layer is composed of four parallel VSS blocks;

[0039] Taking the remote sensing image of phase 1 as an example, each ultra-lightweight encoder performs the following operations on it:

[0040]

[0041] In the above formula, I1 represents the input remote sensing image of phase 1; Respectively represent convolutional blocks and grouped visual GVM blocks; f1 n represents the semantic features of the nth scale of the remote sensing image extracted from the phase 1 by the ultra-lightweight encoder;

[0042] After the above operations of the ultra-lightweight encoder, the remote sensing image of phase 1 will generate multi-scale high-dimensional semantic features, i.e., I1→f1 n , n∈{1, 2, 3, 4, 5, 6}, the remote sensing image of phase 2 also exists

[0043] S3, after each convolution block and grouped visual GVM block of the ultra-light encoder, a Figure 4 The spatial-channel attention module (SCAM) shown in FIG. 1 performs cross-scale semantic fusion on the multi-scale high-dimensional semantic features of the dual-temporal remote sensing image by using the spatial-channel attention module to reduce the semantic gap of features at different scales, and finally generates the multi-scale semantic features of the dual-temporal remote sensing image.

[0044] Taking the remote sensing image of phase 1 as an example, we get f1 n After that, it is first weighted by the previous channel attention submodule, and then weighted by the subsequent spatial attention submodule. The specific operation can be expressed as:

[0045] x SA =x×φ SA x+x;

[0046]

[0047] In the above formula, x = f1 n ; represents the output feature after weighting by the spatial-channel attention module, that is, x SA represents the output of the channel attention submodule; φ SA ,φ CA Respectively represent spatial attention and channel attention;

[0048] After the above operation of the spatial-channel attention module, for f1 n , a new multi-scale semantic feature will be obtained, namely for There are also

[0049] S4, subtract the multi-scale semantic features of the dual-phase remote sensing image to generate a multi-scale change feature map; specifically, and Subtracting them will give you a multi-scale variation feature map, which can be expressed in the following formula:

[0050]

[0051] In the above formula, For the generated multi-scale variation feature map, it should be noted that due to f1 6 and With f1 5 and are the same size, so f1 directly processed by the spatial-channel attention module 6 and generate;

[0052] S5. Input the multi-scale change feature map into the ultra-light decoder for decoding, restore it step by step to the size of the original input remote sensing image, and then use argmax processing to obtain a binary image for change detection; the ultra-light decoder is formed by connecting the three grouped visual GVM blocks in front and the three convolution blocks in the back in series. The last convolution block contains a convolution layer, and the number of output channels is set to 2, which are responsible for outputting the background and change categories respectively. The remaining convolution blocks include a convolution layer, a group normalization layer, a maximum pooling layer and a GELU activation function layer; each of the grouped visual GVM blocks includes a grouped visual Mamba layer, a group normalization layer and a GELU activation function layer; each block in the ultra-light decoder has 2 inputs, which are the decoded features output by the previous block and the deep semantic features of the jump connection after weighting by the ultra-light twin encoder and the space-channel attention module; specifically, the decoding process of the ultra-light decoder is as follows:

[0053] S51. Due to and The sizes are the same, so the two are directly stacked along the channel dimension, and then the stacking result is input into the corresponding grouped visual GVM block to obtain at this time The size is Where N is the batch size of training, that is, the number of samples in each calculation, C represents the number of channels of the corresponding grouped visual GVM block; H and W represent the height and width of the feature respectively;

[0054] S52. Due to and The sizes are different, so you need to After upsampling, the two are stacked along the channel dimension, and then the stacking result is input into the corresponding grouped visual GVM block to obtain Indicates 2x bilinear interpolation upsampling, which generates The size is

[0055] S53. Similarly, by repeating step S52, at this time The size is

[0056] S54, will The convolution layer with a convolution kernel size of 1 is used to compress the image to 2 channels, and then bilinear interpolation upsampling is performed again to obtain a preliminary classification probability map. At this time f final The size of is (N, 2, H, W), which is consistent with the size of the original input remote sensing image;

[0057] Use the training set to train the model, set the model iteration number to 100000, the optimizer to AdamW, the initial learning rate to 0.001, the training batch size to 36, and calculate f final The cross entropy loss function is used to update the model parameters through back propagation; after training, f is directly final Perform argmax operation and output the final binary image. In the binary image, 0 represents the unchanged area and 1 represents the changed area. The trained model is verified using the validation set. The evaluation indicators used include F1 score, mean intersection over union (mIoU), mean accuracy (mAcc), foreground target accuracy (Precision), recall (Recall), computational effort (FLOPs), and total number of parameters (Params). The specific evaluation results are shown in Table 1:

[0058] Table 1 Evaluation results

[0059]

[0060] It can be seen that the present invention reduces the amount of calculation and the total amount of parameters while ensuring the detection effect, and can be used for rapid detection and identification of land cover changes.

[0061] Embodiment 2:

[0062] See also Figure 5, an ultra-lightweight remote sensing image change detection system, comprising an image preprocessing module, an encoding module, a cross-scale semantic fusion module, a change detection module, and a decoding module. The image preprocessing module is used to obtain a dual-phase remote sensing image of a target detection area and preprocess it; the data preprocessing includes random cropping, random flipping, random phase exchange, and data normalization; the encoding module is used to input the preprocessed dual-phase remote sensing image into an ultra-lightweight twin encoder with shared weight parameters to extract multi-scale high-dimensional semantic features of the dual-phase remote sensing image; the ultra-lightweight twin encoder includes two The ultra-lightweight encoder is composed of three convolution blocks in front and three grouped visual GVM blocks in back, each of which includes a convolution layer, a group normalization layer, a maximum pooling layer and a GELU activation function layer, and each of which includes a grouped visual Mamba layer, a group normalization layer, a maximum pooling layer and a GELU activation function layer; the cross-scale semantic fusion module is used to use the space-channel attention module to perform cross-scale semantic fusion on the multi-scale high-dimensional semantic features of the dual-phase remote sensing image. Semantic fusion generates multi-scale semantic features of dual-phase remote sensing images; the space-channel attention module is formed by a previous space attention submodule and a subsequent channel attention submodule in series; the change detection module is used to perform subtraction on the multi-scale semantic features of the dual-phase remote sensing images to generate a multi-scale change feature map; the ultra-lightweight decoder is formed by the previous three grouped visual GVM blocks and the subsequent three convolution blocks in series, the last convolution block contains a convolution layer, and the number of output channels is set to 2, which are responsible for outputting the background and change categories respectively, and the remaining convolution blocks all include a convolution layer. A multi-scale variation feature map is input into the ultra-lightweight decoder to obtain a binary image, and a multi-scale variation feature map is input into the ultra-lightweight decoder to obtain a binary image, and a variation detection is performed using the binary image.

[0063] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0064] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0065] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0066] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0067] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. An ultra-lightweight remote sensing image change detection method, characterized in that: The method comprises: S1, obtaining a dual-phase remote sensing image of the target detection area and preprocessing it; S2, input the preprocessed dual-temporal remote sensing image into an ultra-lightweight twin encoder with shared weight parameters to extract multi-scale high-dimensional semantic features of the dual-temporal remote sensing image; S3, using the spatial-channel attention module to perform cross-scale semantic fusion on the multi-scale high-dimensional semantic features of the dual-temporal remote sensing images to generate multi-scale semantic features of the dual-temporal remote sensing images; S4, performing subtraction on the multi-scale semantic features of the dual-temporal remote sensing images to generate a multi-scale change feature map; S5. Input the multi-scale change feature map into the ultra-lightweight decoder for decoding to obtain a binary image for change detection.

2. The ultra-lightweight remote sensing image change detection method according to claim 1, characterized in that: The ultra-lightweight twin encoder described in S2 includes two ultra-lightweight encoders that share weight parameters, and the ultra-lightweight encoder is formed by connecting three convolution blocks in front and three grouped visual GVM blocks in series. Each of the convolution blocks includes a convolution layer, a group normalization layer, a maximum pooling layer and a GELU activation function layer, and each of the grouped visual GVM blocks includes a grouped visual Mamba layer, a group normalization layer, a maximum pooling layer and a GELU activation function layer.

3. The ultra-lightweight remote sensing image change detection method according to claim 2, characterized in that: The ultra-lightweight decoder in S5 is formed by connecting three preceding grouped visual GVM blocks in series with three succeeding convolution blocks. The last convolution block contains a convolution layer, and the number of output channels is set to 2, which are responsible for outputting background and change categories respectively. The remaining convolution blocks include a convolution layer, a group normalization layer, a maximum pooling layer and a GELU activation function layer; each of the grouped visual GVM blocks includes a grouped visual Mamba layer, a group normalization layer and a GELU activation function layer.

4. An ultra-lightweight remote sensing image change detection method according to any one of claims 1 to 3, characterized in that: Each block in the ultra-lightweight decoder has two inputs, namely, the decoded features output by the previous block and the deep semantic features weighted by the ultra-lightweight twin encoder and the space-channel attention module and then jump-connected; wherein, the space-channel attention module is formed by a preceding space attention submodule and a succeeding channel attention submodule connected in series.

5. An ultra-lightweight remote sensing image change detection method according to any one of claims 1 to 3, characterized in that: The data preprocessing described in S1 includes random cropping, random flipping, random phase swapping, and data normalization.

6. An ultra-lightweight remote sensing image change detection system, characterized in that: The system comprises: An image preprocessing module is used to obtain dual-phase remote sensing images of the target detection area and preprocess them; The encoding module is used to input the preprocessed dual-temporal remote sensing images into an ultra-lightweight twin encoder with shared weight parameters to extract multi-scale high-dimensional semantic features of the dual-temporal remote sensing images; A cross-scale semantic fusion module is used to perform cross-scale semantic fusion on the multi-scale high-dimensional semantic features of dual-temporal remote sensing images using a spatial-channel attention module to generate multi-scale semantic features of dual-temporal remote sensing images; The change detection module is used to perform subtraction on the multi-scale semantic features of the dual-temporal remote sensing images and generate a multi-scale change feature map; The decoding module is used to input the multi-scale change feature map into the ultra-lightweight decoder for decoding to obtain a binary image, and use the binary image for change detection.

7. The ultra-lightweight remote sensing image change detection system according to claim 6, characterized in that: The ultra-lightweight twin encoder includes two ultra-lightweight encoders that share weight parameters. The ultra-lightweight encoder is formed by connecting three convolution blocks in front and three grouped visual GVM blocks in series. Each of the convolution blocks includes a convolution layer, a group normalization layer, a maximum pooling layer and a GELU activation function layer. Each of the grouped visual GVM blocks includes a grouped visual Mamba layer, a group normalization layer, a maximum pooling layer and a GELU activation function layer.

8. The ultra-lightweight remote sensing image change detection system according to claim 7, characterized in that: The ultra-lightweight decoder is formed by connecting three preceding grouped visual GVM blocks in series with three succeeding convolution blocks. The last convolution block contains a convolution layer, and the number of output channels is set to 2, which are responsible for outputting background and change categories respectively. The remaining convolution blocks include a convolution layer, a group normalization layer, a maximum pooling layer and a GELU activation function layer; each of the grouped visual GVM blocks includes a grouped visual Mamba layer, a group normalization layer and a GELU activation function layer.

9. An ultra-lightweight remote sensing image change detection system according to any one of claims 6-7, characterized in that: Each block in the ultra-lightweight decoder has two inputs, namely, the decoded features output by the previous block and the deep semantic features weighted by the ultra-lightweight twin encoder and the space-channel attention module and then jump-connected; wherein, the space-channel attention module is formed by a preceding space attention submodule and a succeeding channel attention submodule connected in series.

10. An ultra-lightweight remote sensing image change detection system according to any one of claims 6-7, characterized in that: The data preprocessing includes random cropping, random flipping, random phase exchange, and data normalization.

Images