Remote sensing image change detection method and device based on RWKV model

Abstract

The invention discloses a remote sensing image change detection method and device based on an RWKV model, and the method comprises the steps: carrying out the preprocessing of a first detection image and a second detection image, and correspondingly obtaining a first standard image and a second standard image; wherein the first detection image and the second detection image are two remote sensing images with different time phases; constructing an improved RWKV model based on a U-net structure; wherein the improved RWKV model comprises an encoding module with two shared weights and a decoding module; and performing detection through the improved RWKV model according to the first standard image and the second standard image to obtain a detection image. The precision and efficiency of remote sensing image change detection can be improved.

Description

TECHNICAL FIELD

[0001] The application relates to the field of remote sensing image change detection, and particularly relates to a remote sensing image change detection method and device based on an RWKV model. BACKGROUND

[0002] In recent years, with the rapid development of remote sensing technology, the amount of remote sensing image data has increased explosively, and the demand for remote sensing image change detection has become increasingly urgent. However, traditional change detection methods often have problems such as insufficient accuracy and low efficiency when facing complex scenes and large-scale data. For example, pixel difference-based methods are difficult to distinguish between real changes and noise interference, and feature extraction-based methods do not sufficiently mine data features, resulting in inaccurate detection results. The reason is that existing methods do not fully utilize the spatiotemporal context information of multi-temporal images, resulting in an increase in false alarm rate of pixel-level change detection, and limiting the detection accuracy and efficiency of the model when processing large-scale and complex scene remote sensing data.

[0003] In the prior art, there is an RWKV model. The model is a new architecture that combines the advantages of RNN (recurrent neural network) and Transformer model, has strong sequence modeling capability and parallel computing advantage, and can efficiently process time series data and spatial feature information of remote sensing images to capture change patterns in remote sensing images. Therefore, the remote sensing change detection method based on the RWKV model is of great significance. SUMMARY

[0004] To overcome the defects of the prior art, the application provides a remote sensing image change detection method and device based on an RWKV model, which can improve the accuracy and efficiency of remote sensing image change detection.

[0005] An embodiment of the application provides a remote sensing image change detection method based on an RWKV model, comprising the following steps: respectively pre-processing a first detection image and a second detection image to obtain a first standard image and a second standard image correspondingly; wherein the first detection image and the second detection image are two different temporal remote sensing images; constructing an improved RWKV model based on a U-net structure; wherein the improved RWKV model comprises two weight-shared encoding modules and one decoding module; detecting, according to the first standard image and the second standard image, by the improved RWKV model to obtain a detection image.

[0006] Further, the respective pre-processing of the first detection image and the second detection image to obtain the first standard image and the second standard image specifically comprises: The first detection image and the second detection image are respectively standardized by histogram matching to obtain the first standard image and the second standard image.

[0007] wherein, represents the gray value of the detection image pixel, represents the gray value of the standard image pixel, and respectively represent the mean and standard deviation of the gray value of the detection image.

[0008] Further, the improved RWKV model based on the U-net structure is constructed, specifically including: Two encoding modules are constructed, each of which includes a first stage encoder, a second stage encoder, a third stage encoder and a fourth stage encoder, and the first stage encoder, the second stage encoder, the third stage encoder and the fourth stage encoder are connected in series; wherein the weights between the same stage encoders are shared; A decoding module is constructed, which includes a first stage decoder, a second stage decoder, a third stage decoder and a fourth stage decoder, and the first stage decoder, the second stage decoder, the third stage decoder and the fourth stage decoder are connected in series; wherein each stage decoder includes a sub-module; Based on the U-net structure, the output of the first stage encoder is taken as the input of the fourth stage decoder, the output of the second stage encoder is taken as the input of the third stage decoder, the output of the third stage encoder is taken as the input of the second stage decoder, and the output of the fourth stage encoder is taken as the input of the first stage decoder, and finally the improved RWKV model is obtained.

[0009] Preferably, each stage encoder includes a convolution down-sampling block and an rwkv block, and the rwkv block includes spatial mix and channel mix modules.

[0010] Further, the detection image is obtained by detecting the first standard image and the second standard image through the improved RWKV model, specifically including: According to the first standard image and the second standard image, a first feature map set and a second feature map set are respectively extracted through the two weight-shared encoding modules; According to the first feature map set and the second feature map set, the detection image is obtained by feature splicing through the decoding module.

[0011] Further, the first feature map set and the second feature map set are respectively extracted from the first standard image and the second standard image according to the first standard image and the second standard image, and the first feature map set and the second feature map set are respectively extracted by the encoding module, and specifically comprising: The first 1st-order feature map and the second 1st-order feature map are obtained by extracting the first standard image and the second standard image through the first stage encoder respectively; The first 2nd-order feature map and the second 2nd-order feature map are obtained by extracting the first 1st-order feature map and the second 1st-order feature map through the second stage encoder respectively; The first 3rd-order feature map and the second 3rd-order feature map are obtained by extracting the first 2nd-order feature map and the second 2nd-order feature map through the third stage encoder respectively; The first 4th-order feature map and the second 4th-order feature map are obtained by extracting the first 3rd-order feature map and the second 3rd-order feature map through the first stage encoder respectively; The first 1st-order feature map, the first 2nd-order feature map, the first 3rd-order feature map and the first 4th-order feature map are combined to obtain the first feature map set, and the second 1st-order feature map, the second 2nd-order feature map, the second 3rd-order feature map and the second 4th-order feature map are combined to obtain the second feature map set.

[0012] Further, the first feature map set and the second feature map set are respectively extracted from the first standard image and the second standard image according to the first standard image and the second standard image, and the first feature map set and the second feature map set are respectively extracted by the encoding module, and specifically comprising: The first 4th-order feature map and the second 4th-order feature map are spliced by the sub-module in the first stage decoder to obtain a 1st-order sub-spliced image, and then the first 1st-order sub-spliced image is up-sampled to obtain a 1st-order spliced image; The first 3rd-order feature map and the second 3rd-order feature map are spliced by the sub-module in the second stage decoder to obtain a 2nd-order sub-spliced image, and then the 1st-order spliced image and the 2nd-order sub-spliced image are subjected to preset fusion processing, and the result after the preset fusion processing is up-sampled to obtain a 2nd-order spliced image; The first 2nd-order feature map and the second 2nd-order feature map are spliced by the sub-module in the third stage decoder to obtain a 3rd-order sub-spliced image, and then the 2nd-order spliced image and the 3rd-order sub-spliced image are subjected to preset fusion processing, and the result after the preset fusion processing is up-sampled to obtain a 3rd-order spliced image; The first 1st-order feature map and the second 1st-order feature map are spliced by the sub-module in the fourth stage decoder to obtain a 4th-order sub-spliced image, and then the 3rd-order spliced image and the 4th-order sub-spliced image are subjected to preset fusion processing, and the result after the preset fusion processing is up-sampled to obtain the detection image.

[0013] Further, the specific process of the splicing processing includes: The first n-order feature map and the second n-order feature map are spliced and combined through sequential modeling, and the spliced and combined result is input into the rwkv block to obtain a sequential splicing map; ; The first n-order feature map and the second n-order feature map are spliced and combined through cross modeling, and the spliced and combined result is input into the rwkv block to obtain a cross splicing map; The first n-order feature map and the second n-order feature map are spliced and combined through parallel modeling, and the spliced and combined result is input into the rwkv block to obtain a parallel splicing map; The sequential splicing map, the cross splicing map and the parallel splicing map are channel spliced to obtain a sub-splicing map.

[0014] Further, the specific process of the preset fusion processing includes: The i-order splicing map and the i+1-order sub-splicing map are added element by element to obtain an added feature; ; The added feature is subjected to learnable convolution processing to obtain a convolution feature; The convolution feature is subjected to batch normalization processing to obtain the result of the fusion processing.

[0015] Another embodiment of the application provides a remote sensing image change detection device based on a RWKV model, comprising a preprocessing module, a construction module and a detection module. The preprocessing module is used for preprocessing a first detection image and a second detection image respectively, and obtaining a first standard image and a second standard image correspondingly; Wherein, the first detection image and the second detection image are two different time phase remote sensing images; The construction module is used for constructing an improved RWKV model based on a U-net structure; Wherein, the improved RWKV model includes two weight-shared encoding modules and one decoding module; The detection module is used for detecting according to the first standard image and the second standard image through the improved RWKV model to obtain a detection image.

[0016] Compared with the prior art, the beneficial effects of the application are as follows: The RWKV model is combined with the U-Net structure for remote sensing change detection, the advantages of the two are combined, and the understanding and recognition ability of the model to complex change patterns is enhanced. In addition, by designing a weight-shared RWKV encoder, the multi-time feature can be adaptively weighted, the non-target change interference such as vegetation growth cycle and cloud cover is suppressed, and the accuracy and efficiency of remote sensing image change detection are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The present invention provides a flow chart of a remote sensing image change detection method based on the RWKV model according to an embodiment of the present invention.

[0018] Figure 2 A schematic structural diagram of an improved RWKV model provided by one embodiment of the present invention.

[0019] Figure 3 A schematic diagram of a framework of an encoding module provided by one embodiment of the present invention.

[0020] Figure 4 A schematic diagram of the structure of an rwkv block provided by one embodiment of the present invention.

[0021] Figure 5 A schematic structural diagram of a decoding module provided by one embodiment of the present invention.

[0022] Figure 6 A schematic structural diagram of a remote sensing image change detection device based on the RWKV model is provided in another embodiment of the present invention.

[0023] Figure 7 This is a test result diagram of an improved RWKV model provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0024] The accompanying drawings are for illustrative purposes only and are not to be construed as limiting this patent; It is understandable to those skilled in the art that some well-known structures and descriptions thereof may be omitted in the drawings.

[0025] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0026] Reference Figure 1 , which is a flow chart of a remote sensing image change detection method based on the RWKV model provided by one embodiment of the present invention, comprising the following steps: S1: Preprocessing the first detection image and the second detection image respectively to obtain a first standard image and a second standard image respectively; wherein the first detection image and the second detection image are two remote sensing images of different phases; S2: Constructing an improved RWKV model based on the U-net structure; wherein the improved RWKV model includes two weight-sharing encoding modules and one decoding module; S3: Perform detection using the improved RWKV model based on the first standard image and the second standard image to obtain a detection image.

[0027] For step S1, specifically, the first detection image and the second detection image are respectively preprocessed to obtain a first standard image and a second standard image, specifically including: The first detection image and the second detection image are respectively standardized by histogram matching to obtain the first standard image and the second standard image; wherein the specific formula of the standardization processing is as follows:

[0028] Wherein, represents the gray value of the detection image pixel, represents the gray value of the standard image pixel, and respectively represent the mean and standard deviation of the gray value of the detection image.

[0029] In a preferred embodiment, assuming that the first detection image is T1 and the second detection image is T2, then in the preprocessing stage, the histogram matching standardization operation needs to be implemented for T1 and T2 two-phase images respectively. The reason for this operation is to eliminate the image gray value deviation caused by different lighting conditions due to the imaging time difference, and to improve the consistency of the two-phase images in the gray distribution, thereby creating good data conditions for subsequent feature extraction and comparison analysis. This standardization operation can standardize the gray values of T1 and T2, making their mean value 0 and standard deviation 1, realizing the normalization of different images in statistical characteristics.

[0030] For step S2, specifically, the improved RWKV model is constructed based on the U-net structure, specifically including: Two encoding modules are constructed, each of which includes a first stage encoder, a second stage encoder, a third stage encoder and a fourth stage encoder, and the first stage encoder, the second stage encoder, the third stage encoder and the fourth stage encoder are connected in series; wherein the weights between the same stage encoders are shared; A decoding module is constructed, which includes a first stage decoder, a second stage decoder, a third stage decoder and a fourth stage decoder, and the first stage decoder, the second stage decoder, the third stage decoder and the fourth stage decoder are connected in series; wherein each stage decoder includes a sub-module; Based on the U-net structure, the output of the first stage encoder is taken as the input of the fourth stage decoder, the output of the second stage encoder is taken as the input of the third stage decoder, the output of the third stage encoder is taken as the input of the second stage decoder, and the output of the fourth stage encoder is taken as the input of the first stage decoder, finally obtaining the improved RWKV model.

[0031] In a preferred embodiment, referring to Figure 2 Figure 1 is a structural schematic diagram of an improved RWKV model according to an embodiment of the present application. As shown in the figure, the improved RWKV model comprises two encoding modules and one decoding module. Figure 2 It can be seen that the improved RWKV model comprises two encoding modules and one decoding module, the encoding modules are used to extract remote sensing image features, gradually reduce feature resolution and increase channel dimension to capture rich semantic information, and the decoding module is used to gradually up-sample to recover feature space resolution and finally generate change detection results.

[0032] In the improved RWKV model, each encoding module and decoding module has four stages, i.e. the first stage encoder, the second stage encoder, the third stage encoder and the fourth stage encoder connected in series, and the first stage decoder, the second stage decoder, the third stage decoder and the fourth stage decoder connected in series.

[0033] In the improved RWKV model, a weight sharing mechanism is adopted between the two encoding modules, i.e. weight sharing between the same stage encoders. This mechanism can enable the two encoding modules to cooperate with each other during the learning process and jointly optimize the model parameters, so as to more effectively mine the internal correlation information between two time images and realize deep learning of multi-band joint features of the images, thereby providing more accurate and rich feature representation for subsequent time change analysis tasks.

[0034] Preferably, each stage encoder comprises a convolution down-sampling block and an rwkv block, and the rwkv block comprises a spatial mix module and a channel mix module.

[0035] Referring to Figure 3 Figure 2 is a framework schematic diagram of an encoding module according to an embodiment of the present application. As shown in the figure, each stage encoder comprises a convolution down-sampling block and an rwkv block. Figure 3 Referring to Figure 4 Figure 3 is a structural schematic diagram of an rwkv block according to an embodiment of the present application. As shown in the figure, the rwkv block comprises a spatial mix module and a channel mix module. Figure 4 It can be seen that the rwkv block comprises a spatial mix module and a channel mix module.

[0036] In the rwkv block, the spatial mix module replaces the self-attention mechanism of the Transformer to model sequence dependence with linear complexity. By mixing current and historical information through a recursive structure, a learnable decay factor is introduced to realize efficient calculation. The channel mix module replaces the feedforward network to perform nonlinear transformation in the feature dimension, and enhances the expression ability through a gating mechanism. This design effectively improves the utilization efficiency of the model for multi-spectral channel features, so that the encoder can learn more discriminative channel interaction features.

[0037] For step S3, specifically, the detection image is obtained by detecting the first standard image and the second standard image through the improved RWKV model, specifically comprising: According to the first standard image and the second standard image, a first feature map set and a second feature map set are respectively extracted through the two weight-shared encoding modules; According to the first feature map set and the second feature map set, feature splicing is performed through the decoding module to obtain a detection image.

[0038] Further, the first feature map set and the second feature map set are respectively extracted through the encoding module according to the first standard image and the second standard image, specifically comprising: The first 1st-order feature map and the second 1st-order feature map are obtained by extracting the first standard image and the second standard image through the first stage encoder, respectively; The first 2nd-order feature map and the second 2nd-order feature map are obtained by extracting the first 1st-order feature map and the second 1st-order feature map through the second stage encoder, respectively; The first 3rd-order feature map and the second 3rd-order feature map are obtained by extracting the first 2nd-order feature map and the second 2nd-order feature map through the third stage encoder, respectively; The first 4th-order feature map and the second 4th-order feature map are obtained by extracting the first 3rd-order feature map and the second 3rd-order feature map through the first stage encoder, respectively; The first 1st-order feature map, the first 2nd-order feature map, the first 3rd-order feature map, and the first 4th-order feature map are combined to obtain the first feature map set, and the second 1st-order feature map, the second 2nd-order feature map, the second 3rd-order feature map, and the second 4th-order feature map are combined to obtain the second feature map set.

[0039] In a preferred embodiment, to realize the dimension reduction of the image resolution, the encoder adopts a 4-stage downsampling process. In each stage, a 4x4 convolution operation with a step size of 2 is used, so that the image resolution is sequentially reduced from the original resolution to 1 / 4, 1 / 8, 1 / 16, and 1 / 32. Finally, a series of multi-scale feature maps, i.e., the first feature map set and the second feature map set, are output, which are represented as (wherein represents the order of the feature map, and ). The generation of multi-scale feature maps lays a solid foundation for subsequent capture of spatio-temporal features at different scales.

[0040] Further, the first feature map set and the second feature map set are respectively extracted through the encoding module according to the first standard image and the second standard image, specifically comprising: The first 4-order feature map and the second 4-order feature map are spliced by a sub-module in the first stage decoder to obtain a 1-order sub-spliced map, and then the first 1-order sub-spliced map is up-sampled to obtain a 1-order spliced map; The first 3-order feature map and the second 3-order feature map are spliced by a sub-module in the second stage decoder to obtain a 2-order sub-spliced map, and then the 1-order spliced map and the 2-order sub-spliced map are pre-set fused, and the result after the pre-set fusion is up-sampled to obtain a 2-order spliced map; The first 2-order feature map and the second 2-order feature map are spliced by a sub-module in the third stage decoder to obtain a 3-order sub-spliced map, and then the 2-order spliced map and the 3-order sub-spliced map are pre-set fused, and the result after the pre-set fusion is up-sampled to obtain a 3-order spliced map; The first 1-order feature map and the second 1-order feature map are spliced by a sub-module in the fourth stage decoder to obtain a 4-order sub-spliced map, and then the 3-order spliced map and the 4-order sub-spliced map are pre-set fused, and the result after the pre-set fusion is up-sampled to obtain the detection image.

[0041] Further, the specific process of the splicing processing includes: The first n-order feature map and the second n-order feature map are spliced and combined by sequential modeling, and the result after the splicing and combination is input into the rwkv block to obtain a sequential spliced map; wherein, ; The first n-order feature map and the second n-order feature map are spliced and combined by cross modeling, and the result after the splicing and combination is input into the rwkv block to obtain a cross spliced map; The first n-order feature map and the second n-order feature map are spliced and combined by parallel modeling, and the result after the splicing and combination is input into the rwkv block to obtain a parallel spliced map; The sequential spliced map, the cross spliced map and the parallel spliced map are channel spliced to obtain a sub-spliced map.

[0042] Further, the specific process of the pre-set fusion processing includes: The i-order spliced map and the i+1-order sub-spliced map are added element by element to obtain an added feature; wherein, ; The added feature is processed by a learnable convolution to obtain a convolution feature; The convolution feature is processed by batch normalization to obtain the result after the fusion processing.

[0043] In a preferred embodiment, reference is made toFigure 5 FIG. 1 is a structural schematic diagram of a decoding module according to an embodiment of the present application. The decoding module comprises a first feature map set input unit 101, a second feature map set input unit 102, a sub-module 103, and a decoding output unit 104. Figure 5 It can be seen that when the first feature map set and the second feature map set are input into the decoding module, they first need to undergo splicing processing by the sub-module.

[0044] In the sub-module, the first feature map set and the second feature map set need to undergo sequential modeling, cross modeling and parallel modeling respectively, and then the results obtained after modeling are spliced in the channel to obtain the sub-spliced map.

[0045] Sequential modeling splices image features (i.e. and ) in time sequence, thereby effectively modeling time sequence dependence and reflecting the time sequence of changes. Let represent the time corresponding image feature map (i.e. the first n-order feature map, ), represent the time corresponding image feature map (i.e. the second n-order feature map); wherein is the height of the feature map, is the width, is the number of channels. Then the splicing operation of sequential modeling can be represented as:

[0046] wherein, is the spliced feature map. This splicing method enables the model to learn the time sequence, helping the model better understand the dynamic process of changes.

[0047] Cross modeling splices the features of and alternately, thereby strengthening the spatio-temporal comparison of the same position pixels at different times and enabling more sensitive capture of pixel-level changes. For the pixel features and of the same position in the feature maps and , the splicing operation of cross modeling can be represented as:

[0048] Combining the cross-spliced features of all positions, the feature map after cross modeling is obtained, enabling the model to more clearly identify pixel feature changes and thereby improving the accuracy of pixel-level change detection.

[0049] Parallel modeling splices and The features are spliced, and the feature information of two times is jointly represented on the channel, similar to the fusion of binocular vision, which can comprehensively utilize the feature information of different times and mine richer spatiotemporal feature patterns. The splicing operation of parallel modeling can be represented as:

[0050] wherein, represents the splicing operation in the channel dimension. Parallel modeling can combine spectral reflectance and texture information of different times to more accurately determine the change of types.

[0051] The three spatiotemporal modeling mechanisms mine the spatiotemporal features of multi-temporal images from different angles through different feature splicing methods, complement each other, and improve the accuracy and reliability of remote sensing image binary change detection.

[0052] The results output after the above three modeling processes need to be further processed by the rwkv block to obtain the sequential splicing image, the cross splicing image and the parallel splicing image. Subsequently, the sequential splicing image, the cross splicing image and the parallel splicing image are subjected to a splicing operation at the channel level to obtain a sub-splicing image. Finally, the sub-splicing image is subjected to up-sampling to reduce the number of channels and increase the picture size to obtain a splicing image.

[0053] In the second stage decoder and the decoder of the subsequent stage, the sub-splicing image output by the sub-module of each stage needs to be subjected to a preset fusion process with the splicing image output by the previous stage. In the preset fusion process, the elements are sequentially added, then the features are further extracted through a learnable convolution, and finally a batch normalization operation is performed to obtain the final output result.

[0054] In a preferred embodiment, when the improved RWKV model is trained, the remote sensing change detection task is regarded as a binary classification problem (change / non-change), and the binary label of each pixel needs to be predicted. Therefore, the cross-entropy loss function is used as the loss function for training in the preferred embodiment, which is used to measure the distribution difference between the predicted probability of the model and the true label. The calculation formula of the cross-entropy loss function is:

[0055] wherein, represents the total number of training samples (each sample is usually a pair of multi-temporal images and its true label ), is the sample index ( =1,2,…, ); is the class index, here is binary classification, c = 0 represents "no change", and c = 1 represents "change"; is one-hot encoding form of is the output result of the binary change detector through the softmax activation function, that is, the predicted change probability.

[0056] The purpose of this formula design is to optimize the model parameters θ so that the predicted change probability is as close as possible to the true label .

[0057] Referring to Figure 6 , the structure schematic diagram of a remote sensing image change detection device based on the RWKV model provided by another embodiment of the application comprises a preprocessing module 101, a construction module 102, and a detection module 103. The preprocessing module 101 is configured to preprocess a first detection image and a second detection image respectively, and obtain a first standard image and a second standard image correspondingly; wherein the first detection image and the second detection image are two remote sensing images of different time phases. The construction module 102 is configured to construct an improved RWKV model based on a U-net structure; wherein the improved RWKV model comprises two weight-shared encoding modules and one decoding module. The detection module 103 is configured to detect, according to the first standard image and the second standard image, through the improved RWKV model, to obtain a detection image.

[0058] To verify the feasibility of the remote sensing image change detection method based on the RWKV model, the WHU-CD-256 dataset is selected for experiment in the embodiment of the application. The WHU-CD-256 dataset is a commonly used dataset for remote sensing image building change detection. The picture size of the dataset is 32207 x 15354 pixels, and the picture is cropped into small blocks of 256 x 256 pixels, obtaining 7634 picture blocks of 256 x 256 pixels. The dataset has the characteristics of high resolution and large size, and covers a wide range. The experimental process includes: 1. Data processing stage: divide the dataset into a training set, a validation set, and a test set, which are 5948, 744, and 745 respectively. Crop the multi-temporal image pairs and related labels to 256 x 256 pixels to input into the network, and then use the trained network on the test set to infer the data with the original size. Random rotation, left-right flipping, and up-down flipping are used as training data augmentation methods.

[0059] 2. Training network phase: the improved RWKV model used by the embodiment of the present application is a 4-layer U-Net structure, which includes four times of down-sampling and up-sampling. In the network, the input remote sensing detection image finally obtains the output detection image after passing through two weight-shared encoder modules and a feature fusion module. In the training phase, the network is optimized using the AdamW optimizer, the learning rate is 0.0001, the weight decay is 0.005, the batch size is set to 16, and the training iteration number is set to 50,000.

[0060] 3. Test set test phase: in order to evaluate the performance of the improved RWKV model, the embodiment of the present application adopts five key evaluation indicators, namely overall accuracy (OA), precision (Pre), recall (Rec), F1 score and intersection over union (IoU). OA represents the proportion of correctly predicted pixels to the total number of pixels, Pre reflects the proportion of true positive pixels in all pixels predicted as positive, and Rec represents the proportion of true positive pixels in all positive pixels in the true situation. F1 score balances precision and recall by calculating the harmonic mean of Pre and Rec. IoU measures the degree of overlap between the predicted positive region and the true positive region. These indicators are defined as follows: Pre = , Rec = , F1 = , IoU = , OA =

[0061] Where TP, TN, FP and FN represent the number of true positives, true negatives, false positives and false negatives, respectively. F1 and IoU can better reflect the generalization ability of the model.

[0062] The trained improved RWKV model obtained by the above steps is tested on the WHU-CD-256 test set. Referring to Figure 7 , a test result diagram of an improved RWKV model provided by an embodiment of the present application, in which the first two images are the input detection images of the test set, and the third image is the output monitoring image of the model. As can be seen from Figure 7 , the detection image can achieve good segmentation in structure and edge details, reflecting the changes in details before and after the remote sensing image.

[0063] Obviously, the above embodiments of the present application are merely exemplary but not intended to limit the embodiments of the present application. Based on the above description, any other variations or changes can be made by those skilled in the art without departing from the spirit and principles of the present application. It is not necessary to list all the embodiments here. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present application should be included in the protection scope of the claims of the present application.

Claims

1. A remote sensing image change detection method based on the RWKV model, characterized in that: The steps include: Preprocessing the first detection image and the second detection image respectively to obtain a first standard image and a second standard image respectively; wherein the first detection image and the second detection image are two remote sensing images of different phases; An improved RWKV model is constructed based on the U-net structure; wherein the improved RWKV model includes two weight-sharing encoding modules and one decoding module; According to the first standard image and the second standard image, detection is performed using the improved RWKV model to obtain a detection image.

2. The remote sensing image change detection method based on the RWKV model according to claim 1, characterized in that: The preprocessing of the first detection image and the second detection image to obtain the first standard image and the second standard image respectively specifically includes: The first detection image and the second detection image are respectively normalized by histogram matching to obtain the first standard image and the second standard image respectively. The specific formula of the normalization process is as follows: in, Represents the grayscale value of the detected image pixel, Represents the grayscale value of the standard image pixel, and Represent the mean and standard deviation of the grayscale value of the detection image respectively.

3. The remote sensing image change detection method based on the RWKV model according to claim 1, characterized in that: The improved RWKV model based on the U-net structure is constructed, specifically including: Constructing two encoding modules, each of which includes a first-stage encoder, a second-stage encoder, a third-stage encoder, and a fourth-stage encoder, wherein the first-stage encoder, the second-stage encoder, the third-stage encoder, and the fourth-stage encoder are connected in series; wherein the encoders in the same stage share weights; Constructing a decoding module, wherein the decoding module includes a first-stage decoder, a second-stage decoder, a third-stage decoder, and a fourth-stage decoder, and the first-stage decoder, the second-stage decoder, the third-stage decoder, and the fourth-stage decoder are connected in series in sequence; wherein each stage decoder includes a submodule; Based on the U-net structure, the output of the first-stage encoder is used as the input of the fourth-stage decoder, the output of the second-stage encoder is used as the input of the third-stage decoder, the output of the third-stage encoder is used as the input of the second-stage decoder, and the output of the fourth-stage encoder is used as the input of the first-stage decoder, finally obtaining the improved RWKV model.

4. The remote sensing image change detection method based on the RWKV model according to claim 3, characterized in that: Each stage encoder includes a convolutional downsampling block and an rwkv block, and the rwkv block includes spatial mix and channel mix modules.

5. The remote sensing image change detection method based on the RWKV model according to claim 3, characterized in that: The detecting, using the improved RWKV model, according to the first standard image and the second standard image to obtain a detected image specifically includes: According to the first standard image and the second standard image, respectively extracting a first feature atlas and a second feature atlas through the two weight-sharing encoding modules; According to the first feature atlas and the second feature atlas, feature splicing is performed by the decoding module to obtain a detection image.

6. The remote sensing image change detection method based on the RWKV model according to claim 5, characterized in that: The step of extracting the first feature atlas and the second feature atlas from the first standard image and the second standard image respectively by the encoding module specifically includes: Extracting the first standard image and the second standard image respectively through the first stage encoder to obtain a first first-order feature map and a second first-order feature map; Extracting the first first-order feature map and the second first-order feature map respectively through the second-stage encoder to obtain a first second-order feature map and a second second-order feature map; Extracting the first 2nd-order feature map and the second 2nd-order feature map respectively by the third-stage encoder to obtain a first 3rd-order feature map and a second 3rd-order feature map; Extracting the first third-order feature map and the second third-order feature map respectively by the first-stage encoder to obtain a first fourth-order feature map and a second fourth-order feature map; The first first-order feature map, the first second-order feature map, the first third-order feature map and the first fourth-order feature map are combined to obtain the first feature map set, and the second first-order feature map, the second second-order feature map, the second third-order feature map and the second fourth-order feature map are combined to obtain the second feature map set.

7. The remote sensing image change detection method based on the RWKV model according to claim 6, characterized in that: The step of performing feature stitching by the decoding module based on the first feature atlas and the second feature atlas to obtain a detection image specifically includes: Performing a splicing process on the first 4th-order feature map and the second 4th-order feature map through the submodule in the first-stage decoder to obtain a 1st-order sub-spliced ​​map, and then performing an upsampling process on the first 1st-order sub-spliced ​​map to obtain a 1st-order spliced ​​map; performing a splicing process on the first third-order feature map and the second third-order feature map through a submodule in the second-stage decoder to obtain a second-order sub-spliced ​​map, then performing a preset fusion process on the first-order spliced ​​map and the second-order sub-spliced ​​map, and upsampling the result of the preset fusion process to obtain a second-order spliced ​​map; performing a splicing process on the first second-order feature map and the second second-order feature map through a submodule in the third-stage decoder to obtain a third-order sub-spliced ​​map, then performing a preset fusion process on the second-order spliced ​​map and the third-order sub-spliced ​​map, and upsampling the result of the preset fusion process to obtain a third-order spliced ​​map; The first first-order feature map and the second first-order feature map are spliced ​​together by the sub-module in the fourth-stage decoder to obtain a fourth-order sub-spliced ​​map, and then the third-order spliced ​​map and the fourth-order sub-spliced ​​map are subjected to a preset fusion process, and the result after the preset fusion process is upsampled to obtain the detection image.

8. The remote sensing image change detection method based on the RWKV model according to claim 7, characterized in that: The specific process of the splicing process includes: The first n-order feature map and the second n-order feature map are spliced ​​and combined by sequential modeling, and the spliced ​​and combined result is input into the rwkv block to obtain a sequential splicing map; wherein, ; Splicing and combining the first n-order feature map and the second n-order feature map through cross modeling, and inputting the splicing and combining result into the rwkv block to obtain a cross-spliced ​​map; Splicing and combining the first n-order feature map and the second n-order feature map through parallel modeling, and inputting the splicing and combining result into the rwkv block to obtain a parallel splicing map; Channel stitching is performed on the sequential stitching graph, the cross stitching graph, and the parallel stitching graph to obtain a sub-stitching graph.

9. The remote sensing image change detection method based on the RWKV model according to claim 7, characterized in that: The specific process of the preset fusion processing includes: The i-order splicing graph and the i+1-order sub-splicing graph are added element by element to obtain the additive features; where, ; Performing a learnable convolution process on the added features to obtain convolution features; Batch normalization is performed on the convolution features to obtain the fusion result.

10. A remote sensing image change detection device based on the RWKV model, characterized in that: include: Preprocessing module, construction module and detection module; The preprocessing module is used to preprocess the first detection image and the second detection image respectively to obtain a first standard image and a second standard image respectively; wherein the first detection image and the second detection image are two remote sensing images of different phases; The construction module is used to construct an improved RWKV model based on the U-net structure; wherein the improved RWKV model includes two weight-sharing encoding modules and one decoding module; The detection module is used to perform detection based on the first standard image and the second standard image using the improved RWKV model to obtain a detection image.

Images