Remote sensing image change detection method and system based on multi-stage feature collaborative interaction

Abstract

The application discloses a kind of based on multi-stage feature collaborative interaction remote sensing image change detection method and system, belong to remote sensing image technical field, after the first time phase remote sensing image and second time phase remote sensing image to be detected are preprocessed, input trained remote sensing image change detection model, obtain the change prediction map of two remote sensing images.The remote sensing image change detection model is respectively carried out feature enhancement processing to first, second time phase remote sensing image, obtain the enhanced double time phase feature map;Feature exchange processing is executed to the enhanced double time phase feature map, obtain the exchanged double time phase feature map;After optimization and secondary feature enhancement are respectively carried out to the exchanged double time phase feature map, feature fusion is executed, and fusion feature is generated;Based on fusion feature decoder processing generates the change prediction map of two remote sensing images.The application can effectively inhibit false change, reduce real change missed detection, improve the accuracy and robustness of remote sensing image change detection.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image technology, and in particular to a method and system for detecting changes in remote sensing images based on multi-stage feature collaborative interaction. Background Technology

[0002] Remote sensing image change detection is one of the core technologies in the field of Earth observation and intelligent interpretation. It automatically extracts, identifies, and quantitatively assesses the spatiotemporal evolution information of land cover types and land cover features by comparing and analyzing two remote sensing images of the same geographical area acquired at different times. With the continuous improvement of the aerospace observation technology system, remote sensing satellite constellations carrying multiple types of high-precision payloads are gradually achieving large-scale network operation.

[0003] The new generation of remote sensing platforms has broken through the limitations of traditional imaging methods in terms of spatiotemporal resolution, observation dimensions, and revisit cycles. It can achieve high-frequency, multimodal, and refined observation and perception of surface targets. The resulting massive remote sensing benchmark dataset provides a sufficient data foundation for improving the training and generalization capabilities of deep learning models.

[0004] In recent years, with the rapid development of deep learning technology, various deep learning methods have been widely introduced into the field of remote sensing change detection, effectively solving many technical problems of traditional detection methods. Although significant technological progress has been made in this field, existing models struggle to effectively distinguish between real land cover changes and image style shifts caused by differences in imaging conditions. Specifically, due to differences in imaging conditions such as sensor parameters, atmospheric environment, and solar altitude angle, dual-temporal remote sensing images of the same area often exhibit significant style shifts. This phenomenon greatly increases the discrimination difficulty for change detection models based on architectures such as Siamese networks, easily leading to the model misidentifying pseudo-changes caused by differences in imaging conditions as real land cover changes. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for detecting changes in remote sensing images based on multi-stage feature collaborative interaction.

[0006] On the one hand, a remote sensing image change detection method based on multi-stage feature collaborative interaction is provided, including: Acquire the first and second phase remote sensing images to be detected and perform preprocessing; The preprocessed first-phase and second-phase remote sensing images are input into the trained remote sensing image change detection model to obtain two change prediction maps of remote sensing images. The remote sensing image change detection model is used to: enhance the features of a first-phase remote sensing image and a second-phase remote sensing image respectively; exchange the enhanced first-phase remote sensing image feature map and the enhanced second-phase remote sensing image feature map; optimize and perform secondary feature enhancement on the exchanged first-phase remote sensing image feature map and the exchanged second-phase remote sensing image feature map respectively; fuse the secondary-enhanced first-phase remote sensing image feature map and the secondary-enhanced second-phase remote sensing image feature map to obtain a fused dual-phase remote sensing image feature map; and further process the fused dual-phase remote sensing image feature map using a decoder to obtain change prediction maps of the two remote sensing images.

[0007] Furthermore, feature enhancement is performed on the first-phase remote sensing image, specifically including: First-phase remote sensing images Perform downsampling to obtain the downsampled feature map. ; First-phase remote sensing images Perform downsampling to obtain the downsampled feature map. ; First-phase remote sensing images Wavelet decomposition was performed to extract high-frequency components. , obtain Channel attention map ; feature map Using channel attention maps By weighting and introducing residual connections, the enhanced downsampled feature map is obtained. ; feature map Using channel attention maps By weighting and introducing residual connections, the enhanced downsampled feature map is obtained. By and By concatenating along the channel dimension, the enhanced downsampled feature map is obtained. .

[0008] Furthermore, the feature exchange between the enhanced first-temporal remote sensing image feature map and the enhanced second-temporal remote sensing image feature map specifically includes: The feature vectors of the first-phase remote sensing image after feature enhancement and the feature vectors of the second-phase remote sensing image after feature enhancement are calculated using global average pooling and grouped maximum pooling. The average of the semantic importance scores obtained by combining feature vectors For each pair of feature maps, a pixel-by-pixel difference score is calculated, and then normalized to obtain the feature difference score D. The exchange score is then calculated. S ; Based on exchange score S The final exchange weights are obtained. Based on the exchange weights, the two temporal features are weighted and exchanged respectively, and then reshaped into the original dimensions to obtain the exchanged first temporal remote sensing image feature map and the exchanged second temporal remote sensing image feature map.

[0009] Furthermore, the feature map of the first temporal remote sensing image after the exchange is optimized, including: optimizing the feature map of the first temporal remote sensing image through the feature interaction module before and after the exchange in the first encoder.

[0010] Furthermore, the feature interaction module before and after the exchange includes: a feature optimization submodule after the exchange, a bidirectional attention-guided optimization submodule, and a feature fusion submodule before and after the exchange.

[0011] Furthermore, in the post-exchange feature optimization submodule, the bi-temporal feature difference map before the exchange is used as a guide, and the optimized feature map of the exchanged features is obtained through a multi-head cross-attention mechanism. In the bidirectional attention-guided optimization submodule, spatial context information is extracted from the features before the swap to generate a spatial attention map. The swapped feature map is obtained by spatially calibrating the swapped feature map. A channel attention map is generated based on the spatially optimized swapped feature map. The channel attention map is then used to perform channel calibration on the pre-swap feature map to obtain a channel-optimized pre-swap feature map. In the feature fusion submodule before and after the swap, the spatially optimized swapped feature map and the channel-optimized pre-swap feature map are spliced ​​and fused to obtain the optimized first-phase remote sensing image feature map.

[0012] On the other hand, a remote sensing image change detection system based on multi-stage feature collaborative interaction is provided, including: Acquisition module: Acquires the first and second phase remote sensing images to be detected and performs preprocessing; Detection module: Input the preprocessed first-phase remote sensing image and the second-phase remote sensing image into the trained remote sensing image change detection model to obtain change prediction maps of the two remote sensing images; The remote sensing image change detection model is used to: enhance the features of a first-phase remote sensing image and a second-phase remote sensing image respectively; exchange the enhanced first-phase remote sensing image feature map and the enhanced second-phase remote sensing image feature map; optimize and perform secondary feature enhancement on the exchanged first-phase remote sensing image feature map and the exchanged second-phase remote sensing image feature map respectively; fuse the secondary-enhanced first-phase remote sensing image feature map and the secondary-enhanced second-phase remote sensing image feature map to obtain a fused dual-phase remote sensing image feature map; and further process the fused dual-phase remote sensing image feature map using a decoder to obtain change prediction maps of the two remote sensing images.

[0013] Furthermore, an electronic device is also provided, including: Memory, used for non-transitory storage of computer-readable instructions; and Processor, for executing the computer-readable instructions, When the computer-readable instructions are executed by the processor, they perform the method described in the first aspect above.

[0014] In another aspect, a storage medium is also provided for non-transitory storage of computer-readable instructions, wherein when the non-transitory computer-readable instructions are executed by a computer, the method described in the first aspect is performed.

[0015] In another aspect, a computer program product is also provided, including a computer program that, when run on one or more processors, is used to implement the method described in the first aspect above.

[0016] The above technical solution has the following advantages or beneficial effects: This invention proposes a remote sensing image change detection method and system based on multi-stage feature collaborative interaction. First, a wavelet-enhanced downsampling module is constructed to enhance the feature map. By preserving key edge and texture details, the subsequent feature exchange process benefits from rich structural information. Then, a group exchange module based on importance and dissimilarity is constructed to promote domain adaptation between two temporal images. By selectively exchanging channel groups with lower importance and higher dissimilarity, this module enriches feature diversity without compromising key semantic cues. Furthermore, a feature interaction module before and after the exchange is constructed to fully utilize the diversity generated by feature exchange. Through bidirectional enhancement and fusion, this module effectively suppresses spurious changes and highlights genuine changes, thereby generating more discriminative features. This invention achieves domain adaptation and effectively bridges the semantic gap by maximizing the feature exchange efficiency in the encoding stage. This invention not only significantly suppresses spurious changes but also effectively reduces missed detections in areas of genuine change, thus ensuring the accuracy and robustness of the change detection results. Attached Figure Description

[0017] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0018] Figure 1 This is a flowchart of a remote sensing image change detection method based on multi-stage feature collaborative interaction as described in Example 1. Detailed Implementation

[0019] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0020] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the invention. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] In this embodiment of the invention, "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Furthermore, in the description of this invention, "multiple" refers to two or more.

[0022] Furthermore, to facilitate a clear description of the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.

[0023] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0024] All data acquisition in this embodiment is carried out in accordance with laws and regulations and with user consent, and the data is used legally.

[0025] Example 1 Remote sensing image change detection is a core technology in fields such as land surface resource monitoring and ecological environment assessment. Existing remote sensing image change detection methods are generally susceptible to style shift interference between two temporal images, making it difficult to effectively distinguish between real land cover changes and spurious changes caused by differences in imaging conditions, resulting in significant limitations in detection performance.

[0026] Specifically, due to the influence of imaging factors such as differences in sensor parameters, changes in atmospheric environment, and different solar altitude angles, dual-temporal remote sensing images of the same geographical area exhibit significant feature distribution shifts in dimensions such as texture features, brightness distribution, and contrast. The feature boundaries between such style-level distribution differences and real surface changes are blurred, causing detection models to easily misjudge style differences as surface cover changes. This results in a large number of false change detection results and also easily leads to missed detections of areas with real changes.

[0027] To address the aforementioned style shift problem, existing methods mostly rely on simple domain adaptation strategies or static attention mechanisms for feature correction. However, these methods lack the ability to dynamically perceive the semantic importance and feature differences between different feature levels, and cannot adaptively adjust feature selection weights according to the degree of style shift in the input image. This makes it difficult to achieve a balance between preserving semantic information and suppressing style interference. Furthermore, the interaction between features at different stages of existing networks is poor, failing to fully explore the complementary information of multi-level features, resulting in insufficient overall feature representation capabilities. Consequently, existing models struggle to effectively suppress style interference while preserving key semantic change information, thus limiting the accuracy and robustness of remote sensing image change detection.

[0028] Therefore, to overcome the above-mentioned shortcomings of the prior art, this embodiment provides a remote sensing image change detection method based on multi-stage feature collaborative interaction, such as... Figure 1 As shown, it includes: Acquire the first and second phase remote sensing images to be detected and perform preprocessing; The preprocessed first-phase and second-phase remote sensing images are input into the trained remote sensing image change detection model to obtain two change prediction maps of remote sensing images. The remote sensing image change detection model is used to: enhance the features of a first-phase remote sensing image and a second-phase remote sensing image respectively; exchange the enhanced first-phase remote sensing image feature map and the enhanced second-phase remote sensing image feature map; optimize and perform secondary feature enhancement on the exchanged first-phase remote sensing image feature map and the exchanged second-phase remote sensing image feature map respectively; fuse the secondary-enhanced first-phase remote sensing image feature map and the secondary-enhanced second-phase remote sensing image feature map to obtain a fused dual-phase remote sensing image feature map; and further process the fused dual-phase remote sensing image feature map using a decoder to obtain change prediction maps of the two remote sensing images.

[0029] This embodiment provides a remote sensing image change detection method based on multi-stage feature collaborative interaction, the specific steps of which include: S1: Acquire the first and second phase remote sensing images to be detected and perform preprocessing; To acquire the original dual-temporal remote sensing images, the high-resolution remote sensing images have a large size. Due to limited computing resources, the original dual-temporal images need to be cropped into a first-temporal remote sensing image and a second-temporal remote sensing image of 256*256, where the first-temporal remote sensing image and the second-temporal remote sensing image are paired.

[0030] Preprocessing of the first-phase and second-phase remote sensing images includes: First, the mean and standard deviation of the three RGB channels of the first and second time-phase remote sensing images are calculated, and the image pixel values ​​are standardized. Next, perform the same data augmentation operations, such as randomly flipping the corresponding pair of bi-temporal images horizontally and vertically, and swapping the temporal order of the pair of bi-temporal images.

[0031] S2: Input the preprocessed first-phase remote sensing image and the second-phase remote sensing image into the trained remote sensing image change detection model to obtain change prediction maps of the two remote sensing images; The remote sensing image change detection model employs a multi-stage feature interaction change detection model, which is used to: enhance the features of the first-phase and second-phase remote sensing images respectively; exchange features between the enhanced first-phase and second-phase remote sensing image feature maps; optimize and perform secondary feature enhancement on the exchanged first-phase and second-phase remote sensing image feature maps respectively; fuse the secondary-enhanced first-phase and second-phase remote sensing image feature maps to obtain fused dual-phase remote sensing image feature maps; and further process the fused dual-phase remote sensing image feature maps using a decoder to obtain change prediction maps of the two remote sensing images.

[0032] Furthermore, the preprocessed first-phase and second-phase remote sensing images are input into the trained remote sensing image change detection model to obtain change prediction maps of the two remote sensing images. The trained remote sensing image change detection model employs a multi-stage feature interaction change detection model, including: First encoder, second encoder, fusion unit, and decoder; The input terminal of the first encoder is used to input the first temporal remote sensing image to be detected; The input terminal of the second encoder is used to input the second temporal remote sensing image to be detected; The output terminals of the first encoder and the second encoder are both connected to the input terminal of the fusion unit. The output of the fusion unit is connected to the input of the decoder; The decoder outputs a change prediction map of two remote sensing images.

[0033] Furthermore, the first encoder includes: convolutional and hybrid convolutional units connected in sequence, two wavelet-enhanced downsampling modules, a group exchange module based on importance and difference, a feature interaction module before and after exchange, and a wavelet-enhanced downsampling module; Furthermore, the second encoder includes: convolutional and hybrid convolutional units connected in sequence, two wavelet-enhanced downsampling modules, a group exchange module based on importance and difference, a feature interaction module before and after exchange, and a wavelet-enhanced downsampling module; The convolutional and hybrid convolutional units extract bi-temporal features of the input image without downsampling, thus preserving rich original information.

[0034] S21: Perform feature enhancement on the first time-phase remote sensing image and the second time-phase remote sensing image respectively, specifically including: performing feature enhancement on the first time-phase remote sensing image through the first encoder to obtain the enhanced first time-phase remote sensing image feature map; and performing feature enhancement on the second time-phase remote sensing image through the second encoder to obtain the enhanced second time-phase remote sensing image feature map. In one embodiment, the first temporal remote sensing image is enhanced using a first encoder to obtain an enhanced feature map of the first temporal remote sensing image. Specifically, this includes: performing feature enhancement on the first temporal remote sensing image using the wavelet enhancement downsampling module in the first encoder, with the following steps: S211: The first temporal remote sensing image is processed using 3×3 CBR cells with a step size of 2. Perform downsampling to obtain the downsampled feature map. ; S212: Processing the first temporal remote sensing image using a max-pooling layer with a step size of 2. Perform downsampling to obtain the downsampled feature map. ; S213: Remote sensing images of the first time phase Perform a single-layer Haar discrete wavelet transform (DWT) on the first-phase remote sensing image. It is decomposed into four frequency sub-bands: LL, LH, HL, and HH. Specifically, the LL sub-band captures low-frequency content related to global structural information; while the high-frequency sub-bands LH, HL, and HH correspond to edge features in the vertical, horizontal, and diagonal directions, respectively. This is crucial for recovering fine textures and sharp edges, and can be expressed by the formula: (1) Since the wavelet enhancement downsampling module is used to reconstruct high-frequency details, it only utilizes high-frequency subbands, concatenates them along the channel dimension, and then performs channel blending and feature optimization through a 1×1 CBR unit, as expressed by the formula: (2) The CBR unit consists of a 1×1 convolutional layer, a BN layer, and a ReLU activation function in sequence.

[0035] Utilizing high-frequency components As a guide, the downsampled feature map is enhanced and refined, with high-frequency components... Two distinct spatial context descriptors are aggregated using Global Average Pooling (GAP) and Global Max Pooling (GMP), and then fed into a shared multilayer perceptron. A sigmoid activation function is then applied to generate a channel attention map. This allows for selective enhancement of key channels that contain texture and edge details.

[0036] feature map Using channel attention maps By weighting and introducing residual connections, the enhanced downsampled feature map is obtained. ; feature map Using channel attention maps By weighting and introducing residual connections, the enhanced downsampled feature map is obtained. This can be expressed as a formula: (3) (4) (5) in, and These represent global max pooling and global average pooling along the spatial dimension, respectively. Represents a multilayer perceptron network. This represents the Sigmoid activation function.

[0037] Finally, by and By concatenating along the channel dimension, the enhanced downsampled feature map is obtained. This enriches the expressive power of features, which can be expressed as a formula: (6) Similarly, feature enhancement is performed on the second temporal remote sensing image using a second encoder to obtain the enhanced feature map of the second temporal remote sensing image. .

[0038] S22: Perform feature exchange on the enhanced first-phase remote sensing image feature map and the enhanced second-phase remote sensing image feature map to obtain the exchanged first-phase remote sensing image feature map and the exchanged second-phase remote sensing image feature map; In one embodiment, the feature exchange between the enhanced first-temporal remote sensing image feature map and the enhanced second-temporal remote sensing image feature map specifically includes: performing feature exchange between the enhanced first-temporal remote sensing image feature map and the enhanced second-temporal remote sensing image feature map through a group exchange module based on importance and difference in the first encoder and the second encoder, as follows: S221: Enhanced feature map of the first temporal remote sensing image and the enhanced second-phase remote sensing image feature map Divided along the channel dimension G Groups, each containing K One channel (i.e.) The enhanced first-phase remote sensing image feature map was obtained. and the enhanced second-phase remote sensing image feature map ,in Indicates group index.

[0039] S222: To evaluate the importance of each group, this embodiment simultaneously applies a global average pooling (GAP) operation and a grouped max pooling operation to capture complementary statistical information. The global average pooling operation is applied to calculate the global average of the feature groups. For the grouped max pooling operation, max pooling is first performed along the channel dimension to extract the most salient features within each group, and then the average is calculated along the spatial dimension.

[0040] For the feature map of the first temporal remote sensing image after feature enhancement and the enhanced second-phase remote sensing image feature map The vector obtained from its calculation is represented as: (7) (8) (9) (10) in, and These represent the feature vectors after global average pooling and grouped max pooling, respectively. This indicates a global average pooling operation; Indicates the dimension along the channel K The maximum pooling operation performed; Indicates at altitudeH and width W Average pooling is performed on the spatial dimension.

[0041] S223: Input the feature vectors obtained above into a shared MLP to obtain the biphase importance scores for each group. and The average score of each group The result is obtained by averaging, and can be expressed by the formula: (11) (12) (13) S224: Balance semantic importance with feature differences to determine which groups need to be swapped. Specifically, calculate the difference score pixel-by-pixel between the two feature maps in each group. D And normalize it using the maximum value along the group dimension, expressed by the formula: (14) (15) in, K , H and W These represent the number of channels, height, and width of the feature group, respectively. It is a minimal constant used to ensure numerical stability.

[0042] S225: By combining semantic importance and characteristic differences D To define the exchange score S Expressed by a formula: (16) Formula (16) shows that channel groups with lower semantic importance (i.e., non-dominant features) and greater inter-group differences will obtain higher exchange scores. This reflects the design philosophy of this module: prioritize the exchange of information with lower importance but greater diversity, thereby increasing sample diversity without destroying key semantic content.

[0043] S226: Based on exchange score S The final exchange weights It is calculated using a Sigmoid activation function and can be expressed by the following formula: (17) in, The preset exchange threshold, This is a scaling factor used to control the steepness of the transition.

[0044] S227: Based on the weighted exchange, perform weighted exchange operations on the biphase features to obtain the exchanged features. and This can be expressed as a formula: (18) (19) Finally, the processed feature sets are reshaped back to their original dimensions to obtain the swapped first-phase remote sensing image feature map. and the feature map of the second temporal remote sensing image after exchange .

[0045] S23: Optimize and perform secondary feature enhancement on the exchanged first-phase remote sensing image feature map and the exchanged second-phase remote sensing image feature map respectively to obtain the second-enhanced first-phase remote sensing image feature map and the second-enhanced second-phase remote sensing image feature map. Feature map of the first phase remote sensing image after exchange and the feature map of the second temporal remote sensing image after exchange The optimization is performed separately, specifically including: optimizing the feature map of the first time-phase remote sensing image using the first encoder to obtain the optimized feature map of the first time-phase remote sensing image; and optimizing the feature map of the second time-phase remote sensing image using the second encoder to obtain the optimized feature map of the second time-phase remote sensing image. In one embodiment, optimizing the feature map of the first temporal remote sensing image using a first encoder to obtain an optimized feature map of the first temporal remote sensing image specifically includes: optimizing the feature map of the first temporal remote sensing image using a feature exchange module before and after exchanging features in the first encoder, with the following steps: The feature interaction module before and after the exchange includes three sub-modules: a feature optimization sub-module after the exchange, a bidirectional attention-guided optimization sub-module, and a feature fusion sub-module before and after the exchange.

[0046] S231: In the post-exchange feature optimization submodule, the difference feature map is obtained by calculating the absolute difference between the two temporal feature maps before the exchange. This can be expressed as a formula: (20) A difference-aware cross-attention mechanism is used to enhance the swapped bi-temporal feature maps. To capture global spatial location information, a shared, learnable location is embedded. Injected into the feature map and broadcast along the batch dimension. Initialization is performed by truncating the normal distribution, and its spatial size is similar to the input feature map ( H × WTo maintain consistency and ensure precise element-by-element alignment.

[0047] Difference feature map and swapped feature maps Separately with position encoding By performing element-wise addition, its spatial perception representation is obtained. and The calculation method is as follows: (twenty one) (twenty two) Will and The spatial dimension is flattened to a length of The sequence; to capture diverse representation subspaces, a multi-head cross-attention mechanism is employed, where flattened differential features serve as queries, while flattened, exchanged features provide keys and values. The feature space is projected onto... h In each of the 10 independent attention heads, the dimension of each head is 1. The derivation of the query, key, and value tensors for each head is as follows: (twenty three) (twenty four) (25) in, , , They represent the first i Learnable projection weights for each attention head.

[0048] By concatenating the outputs of all attention points, we obtain the swapped features. Optimized feature map This can be expressed as a formula: (26) The calculation of each attention head is accomplished by scaling the dot product attention, as expressed by the formula: (27) This design utilizes the difference map before the swap as spatial guidance, and enhances the discriminativeness of the swapped features through the cross-attention mechanism of difference perception, thereby improving the feature's ability to detect subtle changes in details.

[0049] S232: In the bidirectional attention-guided optimization submodule, the features before and after the swap interact bidirectionally through two dimensions: space and channel. First, spatial context information is extracted from the features before the swap to generate a spatial attention map, which is used to calibrate the features after the swap and achieve spatial alignment.

[0050] Consider the feature map before swapping and swapped feature maps ,along Global average pooling (GAP) and global max pooling (GMP) are applied along the channel dimension to aggregate spatial information and obtain two feature maps. Then, the two feature maps are concatenated along the channel dimension and input into a... In the convolutional layer, to generate spatial attention maps This can be expressed as a formula: (28) (29) (30) Finally, attention maps are used through residual connections. Features after exchange Spatial calibration is performed to obtain spatially optimized feature maps. This can be expressed as a formula: (31) Given that the swapped feature map contains rich channel information, using from Optimize by extracting channel-level dependencies ,right Global average pooling (GAP) and global max pooling (GMP) are applied to concatenate the two feature maps, which are then processed by a multilayer perceptron (MLP) and a sigmoid activation function to obtain the channel attention map. This can be expressed as a formula: (32) (33) (34) Finally, the channel attention map is applied to the pre-swap features via residual connections. The channel-optimized features are obtained. This can be expressed as a formula: (35) S233: In the feature fusion submodule before and after the swap, in order to fuse the optimized features before and after the swap, their channel dimensions are concatenated, and a... Convolution is used to fuse features, resulting in an optimized first-phase remote sensing image feature map. This can be expressed as a formula: (36) Similarly, the second encoder is used to optimize the feature map of the second temporal remote sensing image, resulting in an optimized feature map of the second temporal remote sensing image. .

[0051] Feature map of the optimized first-phase remote sensing image and optimized second-phase remote sensing image feature map The two-stage feature enhancement is performed separately, specifically including: performing a second-stage feature enhancement on the optimized first-phase remote sensing image feature map using a first encoder to obtain a second-enhanced first-phase remote sensing image feature map; and performing a second-stage feature enhancement on the second-phase remote sensing image feature map using a second encoder to obtain a second-enhanced second-phase remote sensing image feature map.

[0052] In one embodiment, the step of performing secondary feature enhancement on the optimized first temporal remote sensing image feature map through the first encoder to obtain a second-enhanced first temporal remote sensing image feature map specifically includes: performing feature enhancement on the optimized first temporal remote sensing image feature map through the wavelet enhancement downsampling module in the first encoder.

[0053] In one embodiment, the step of performing secondary feature enhancement on the optimized first temporal remote sensing image feature map through the second encoder to obtain the enhanced second temporal remote sensing image feature map specifically includes: performing feature enhancement on the optimized second temporal remote sensing image feature map through the wavelet enhancement downsampling module in the second encoder.

[0054] S24: The feature map of the first time phase remote sensing image after secondary enhancement and the feature map of the second time phase remote sensing image after secondary enhancement are fused by the absolute difference strategy to obtain the fused remote sensing image feature map. By fusing the feature maps of the first temporal remote sensing image after secondary enhancement and the feature maps of the second temporal remote sensing image after secondary enhancement using the absolute difference strategy, the preliminary location of the changing target was achieved.

[0055] S25: The decoder is used to further process the feature map of the fused dual-temporal remote sensing image to obtain the change prediction map of the two remote sensing images.

[0056] This embodiment employs a dedicated decoder module that combines upsampling, skip connections, and residual connections. The decoder module uses a dual decoder to combine high-level semantics with low-level spatial details, thereby promoting the fusion of deep features, achieving accurate localization of changing targets, and ultimately generating full-resolution change prediction maps of two remote sensing images.

[0057] To shorten the gradient propagation path and stabilize the training process, all outputs of the dual decoder and the fusion branch are directly subjected to a supervision signal.

[0058] In remote sensing image change detection, compared to the static background, pixels representing actual changes occupy only a tiny fraction of the entire image area, leading to class imbalance. To address this challenge, this embodiment employs a hybrid loss function, combining binary cross-entropy (BCE) loss with Dice loss. The definitions of these two loss functions are as follows: (37) (38) in, y and y represent the predicted change plot and its corresponding true label, respectively. N The symbol represents the total number of pixels in the change graph. express The norm, the total loss can be expressed as: (39) in, ( ) represent the balance weights of the fusion branch and the dual decoder branch, respectively. ( The diagrams () represent the prediction changes for the fusion branch and the dual-decoder branch, respectively.

[0059] The effectiveness of this embodiment can be further illustrated by experimental results. The experimental data is the LEVIR-CD dataset, a publicly available dataset for remote sensing image change detection. This dataset contains 637 pairs of high-resolution remote sensing images (1024×1024 pixels), covering 20 different regions, mainly focusing on building changes. To improve computational efficiency, the images were cropped into 256×256 tiles. Finally, the training set, validation set, and test set contained 7120, 1024, and 2048 image pairs, respectively. The experimental results are shown in Table 1. Table 1 Experimental Results

[0060] As can be seen from Table 1, compared with existing representative methods, the remote sensing image change detection method based on multi-stage feature collaborative interaction adopted in this embodiment has obvious advantages in detection performance. Its IoU value, F1 value, recall (Rec) and precision (Pre) are all higher than all the comparison methods.

[0061] This embodiment proposes a remote sensing image change detection method based on multi-stage feature collaborative interaction. By maximizing the feature exchange efficiency in the encoding stage, it achieves domain adaptation and effectively bridges the semantic gap. This embodiment can not only significantly suppress spurious changes but also effectively reduce missed detections in truly changed areas, thereby ensuring the accuracy and robustness of the change detection results.

[0062] Example 2 This embodiment provides a remote sensing image change detection system based on multi-stage feature collaborative interaction, including: Acquisition module: Acquires the first and second phase remote sensing images to be detected and performs preprocessing; Detection module: Input the preprocessed first-phase remote sensing image and the second-phase remote sensing image into the trained remote sensing image change detection model to obtain change prediction maps of the two remote sensing images; The remote sensing image change detection model is used to: enhance the features of a first-phase remote sensing image and a second-phase remote sensing image respectively; exchange the enhanced first-phase remote sensing image feature map and the enhanced second-phase remote sensing image feature map; optimize and perform secondary feature enhancement on the exchanged first-phase remote sensing image feature map and the exchanged second-phase remote sensing image feature map respectively; fuse the secondary-enhanced first-phase remote sensing image feature map and the secondary-enhanced second-phase remote sensing image feature map to obtain a fused dual-phase remote sensing image feature map; and further process the fused dual-phase remote sensing image feature map using a decoder to obtain change prediction maps of the two remote sensing images.

[0063] It should be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.

[0064] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0065] The proposed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and the division of modules described above is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.

[0066] Example 3 This embodiment also provides an electronic device, including: one or more processors, one or more memories, and one or more computer programs; wherein, the processor is connected to the memory, and the one or more computer programs are stored in the memory. When the electronic device is running, the processor executes the one or more computer programs stored in the memory to cause the electronic device to perform the method described in Embodiment 1.

[0067] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0068] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0069] In the implementation process, each step of the above method can be completed by the integrated logic circuits in the processor hardware or by software instructions.

[0070] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0071] Those skilled in the art will recognize that the units and algorithm steps described in connection with the various examples of this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention.

[0072] Example 4 This embodiment also provides a storage medium for storing computer instructions, which, when executed by a processor, complete the method described in Embodiment 1.

[0073] Example 5 This embodiment also provides a computer program product, including a computer program that, when run on one or more processors, implements the method described in Embodiment 1.

[0074] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting changes in remote sensing images based on multi-stage feature collaborative interaction, characterized in that, include: Acquire the first and second phase remote sensing images to be detected and perform preprocessing; The preprocessed first-phase and second-phase remote sensing images are input into the trained remote sensing image change detection model to obtain two change prediction maps of remote sensing images. The remote sensing image change detection model is used to: enhance the features of a first-phase remote sensing image and a second-phase remote sensing image respectively; exchange the enhanced first-phase remote sensing image feature map and the enhanced second-phase remote sensing image feature map; optimize and perform secondary feature enhancement on the exchanged first-phase remote sensing image feature map and the exchanged second-phase remote sensing image feature map respectively; fuse the secondary-enhanced first-phase remote sensing image feature map and the secondary-enhanced second-phase remote sensing image feature map to obtain a fused dual-phase remote sensing image feature map; and further process the fused dual-phase remote sensing image feature map using a decoder to obtain change prediction maps of the two remote sensing images.

2. The remote sensing image change detection method based on multi-stage feature collaborative interaction according to claim 1, characterized in that, Feature enhancement is performed on the first-phase remote sensing image, specifically including: First-phase remote sensing images Perform downsampling to obtain the downsampled feature map. ; First-phase remote sensing images Perform downsampling to obtain the downsampled feature map. ; First-phase remote sensing images Wavelet decomposition was performed to extract high-frequency components. , obtain Channel attention map ; feature map Using channel attention maps By weighting and introducing residual connections, the enhanced downsampled feature map is obtained. ; feature map Using channel attention maps By weighting and introducing residual connections, the enhanced downsampled feature map is obtained. By and By concatenating along the channel dimension, the enhanced downsampled feature map is obtained. .

3. The remote sensing image change detection method based on multi-stage feature collaborative interaction according to claim 1, characterized in that, The feature exchange between the enhanced first-temporal remote sensing image feature map and the enhanced second-temporal remote sensing image feature map specifically includes: The feature vectors of the first-phase remote sensing image after feature enhancement and the feature vectors of the second-phase remote sensing image after feature enhancement are calculated using global average pooling and grouped maximum pooling. The average of the semantic importance scores obtained by combining feature vectors For each pair of feature maps, a pixel-by-pixel difference score is calculated, and then normalized to obtain the feature difference score D. The exchange score is then calculated. S ; Based on exchange score S The final exchange weights are obtained. Based on the exchange weights, the two temporal features are weighted and exchanged respectively, and then reshaped into the original dimensions to obtain the exchanged first temporal remote sensing image feature map and the exchanged second temporal remote sensing image feature map.

4. The remote sensing image change detection method based on multi-stage feature collaborative interaction according to claim 1, characterized in that, Optimizing the feature map of the first temporal remote sensing image after the exchange includes: optimizing the feature map of the first temporal remote sensing image through the feature interaction module before and after the exchange in the first encoder.

5. The remote sensing image change detection method based on multi-stage feature collaborative interaction according to claim 4, characterized in that, The feature interaction module before and after the exchange includes: a feature optimization submodule after the exchange, a bidirectional attention-guided optimization submodule, and a feature fusion submodule before and after the exchange.

6. The remote sensing image change detection method based on multi-stage feature collaborative interaction according to claim 5, characterized in that, In the post-swap feature optimization submodule, the bi-temporal feature difference map before the swap is used as a guide, and the optimized feature map of the swapped features is obtained through a multi-head cross-attention mechanism. In the bidirectional attention-guided optimization submodule, spatial context information is extracted from the features before the swap to generate a spatial attention map. The swapped feature map is obtained by spatially calibrating the swapped feature map. A channel attention map is generated based on the spatially optimized swapped feature map. The channel attention map is then used to perform channel calibration on the pre-swap feature map to obtain a channel-optimized pre-swap feature map. In the feature fusion submodule before and after the swap, the spatially optimized swapped feature map and the channel-optimized pre-swap feature map are spliced ​​and fused to obtain the optimized first-phase remote sensing image feature map.

7. A remote sensing image change detection system based on multi-stage feature collaborative interaction, characterized in that, include: Acquisition module: Acquires the first and second phase remote sensing images to be detected and performs preprocessing; Detection module: Input the preprocessed first-phase remote sensing image and the second-phase remote sensing image into the trained remote sensing image change detection model to obtain change prediction maps of the two remote sensing images; The remote sensing image change detection model is used to: enhance the features of a first-phase remote sensing image and a second-phase remote sensing image respectively; exchange the enhanced first-phase remote sensing image feature map and the enhanced second-phase remote sensing image feature map; optimize and perform secondary feature enhancement on the exchanged first-phase remote sensing image feature map and the exchanged second-phase remote sensing image feature map respectively; fuse the secondary-enhanced first-phase remote sensing image feature map and the secondary-enhanced second-phase remote sensing image feature map to obtain a fused dual-phase remote sensing image feature map; and further process the fused dual-phase remote sensing image feature map using a decoder to obtain change prediction maps of the two remote sensing images.

8. An electronic device, characterized in that, include: Memory is used to store computer-readable instructions in a non-transitory manner. as well as Processor, for executing the computer-readable instructions, When the computer-readable instructions are executed by the processor, they perform the remote sensing image change detection method based on multi-stage feature collaborative interaction as described in any one of claims 1-6.

9. A storage medium, characterized in that, Non-transitory storage of computer-readable instructions, wherein when the non-transitory computer-readable instructions are executed by a computer, the remote sensing image change detection method based on multi-stage feature collaborative interaction as described in any one of claims 1-6 is performed.

10. A computer program product, characterized in that, The method includes a computer program that, when running on one or more processors, implements the remote sensing image change detection method based on multi-stage feature collaborative interaction as described in any one of claims 1-6.

Images