Satellite and unmanned aerial vehicle cooperative remote sensing image change detection system and method

Abstract

The invention relates to the technical field of image detection, in particular to a satellite and unmanned aerial vehicle cooperative remote sensing image change detection system and method, and the method comprises the steps: obtaining a dual-time-phase satellite remote sensing image and an unmanned aerial vehicle remote sensing image of the same geographic region; the satellite branches are subjected to four-stage convolution-pooling operation, multi-stage wide-area features are output, and the unmanned aerial vehicle branches output high-resolution local features aligned with the satellite branches in space through the backbone network; performing channel interaction operation on each stage of double-time-phase features, and generating a spatial weight map to strengthen a change region; aligning adjacent level resolutions through transposition convolution, dynamically integrating multi-scale features, and outputting optimized fusion features in combination with residual connection; and performing convolution dichotomy on the fused features to generate a change detection result graph. According to the method, the problem of signal attenuation caused by heterogeneous data feature mismatch is effectively solved, the false alarm rate caused by environmental interference is remarkably reduced, and the complete detection capability of a micro-to-macro full-scale dynamic target is synchronously improved.

Description

Technical Field

[0001] The present invention relates to the field of image detection technology, and in particular to a satellite and unmanned aerial vehicle collaborative remote sensing image change detection system and method. Background Art

[0002] Satellite remote sensing, with its wide-area coverage, can periodically monitor dynamic changes in large-scale surface areas, while drone-based near-Earth observations, with their high resolution, can precisely depict local microscopic changes. The synergy between the two has become a crucial technical support for geological disaster emergency response and refined urban management. However, due to inherent differences in imaging mechanisms, satellite data, constrained by atmospheric conditions and orbital altitude, has low spatial resolution and suffers from temporal phase misalignment. While drone data offers fine resolution, its coverage is limited and susceptible to transient environmental interference, resulting in a significant scale mismatch in the feature spaces of the two data types.

[0003] Traditional technical approaches rely on a strategy of independent processing followed by fusion, using simple weighting or splicing methods, which results in significant attenuation of key subtle change signals such as local road deformation and building microcracks during the fusion process. At the same time, complex environmental noise brings severe interference - the spectral shift caused by seasonal vegetation changes in satellite images, and the instantaneous pseudo-changes caused by sudden changes in illumination in drone data, make it difficult for existing analysis methods to effectively distinguish between real ground object changes and environmental noise, and false alarms continue to occur frequently. The more fundamental challenge lies in the failure of multi-scale dynamic target capture: the standard detection network loses key information about tiny moving targets due to the feature downsampling process, and the convolutional structure with a fixed perception range finds it difficult to simultaneously model the correlation characteristics between large-scale geological deformation and local structural mutations. Ultimately, the completeness of detecting changing targets in complex scenarios is seriously insufficient, greatly restricting the actual effectiveness of the collaborative observation system. Summary of the Invention

[0004] The present invention provides a satellite and UAV collaborative remote sensing image change detection system and method, thereby effectively solving the problems pointed out in the background technology.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] A satellite and UAV collaborative remote sensing image change detection system, comprising:

[0007] A data input module for receiving dual-temporal satellite remote sensing images and UAV remote sensing images of the same geographical area;

[0008] A heterogeneous feature extraction module, comprising a satellite branch and a drone branch. The satellite branch consists of four-level convolutional blocks and outputs multi-level low-resolution wide-area features. The drone branch removes the classification layer and outputs local detail features aligned with the satellite branch hierarchy.

[0009] The cross-domain semantic alignment module includes a difference perception enhancement unit and a context enhancement unit. The difference perception enhancement unit is used to perform channel interaction reorganization on the corresponding hierarchical features of the satellite branch and the UAV branch, generate a spatial weight map and weighted fusion features. The context enhancement unit is used to superimpose the fusion features output by the difference perception enhancement unit with the high-level CDSA features, and enhance the contextual semantics through multi-expansion rate void convolution and channel attention mechanism.

[0010] The multi-scale feature aggregation module aligns the CDSA feature resolutions of adjacent layers through transposed convolution, uses an attention feature fusion mechanism for cross-layer feature integration, and introduces residual connections to optimize gradient transfer.

[0011] The change detection output module performs binary classification on the fused features after feature aggregation and generates a pixel-level change detection map.

[0012] Furthermore, a hybrid loss optimization module is included to calculate the IoU-BCE hybrid loss function, which is formulated as follows:

[0013] Loss = αLoss BCE +(1-α)Loss iou

[0014] Among them, α is a constant used to balance the IoU weight and BCE weight, α = 0.5, Where N represents the total number of pixels, and t and p represent the ground truth label and the predicted result map respectively. For each pixel I,t i ∈{0,1} represents the ground truth class, 0 represents unchanged, 1 represents changed, and the predicted probability P of pixel i being classified as changed i ∈[0,1], and 1-P i is the predicted probability of no change.

[0015] Furthermore, the difference perception enhancement unit includes:

[0016] The cross-layer feature fusion subunit upsamples the features output by the next layer and reduces the channels by half before superimposing them with the satellite and drone features of the current layer.

[0017] The dual-branch reconstruction subunit splits the superimposed features into two independent channel groups and recombines them interactively;

[0018] The spatial weight generation subunit performs convolution processing on the reorganized two branches and compresses the dimensions. After merging, it generates a spatial weight map through the S-type activation function.

[0019] Furthermore, the context enhancement unit includes:

[0020] The channel splicing and dimensionality reduction subunit splices the input features in the channel dimension and compresses the channel dimension to half of the original feature through the convolution-batch normalization-activation function processing process;

[0021] Attention channel, which calibrates the importance weight of each channel through global information aggregation and adaptive reweighting mechanism;

[0022] The context extraction subunit uses three dilated convolutions with different receptive fields to extract multi-scale context features in parallel and fuse the multi-scale features.

[0023] The feature optimization weighting subunit refines the feature expression through 1×1 convolution and performs feature weighting using the spatial weight map generated by the difference perception enhancement unit.

[0024] Furthermore, the multi-scale feature aggregation module includes:

[0025] The feature alignment subunit performs a learnable upsampling operation on the input features of adjacent levels to unify the spatial resolution;

[0026] The channel fusion subunit splices the spatially aligned features in the channel dimension;

[0027] The multi-scale extraction subunit includes a first branch and a second branch, wherein the first branch performs a multi-branch heterogeneous convolution operation on the splicing features to generate a multi-scale feature expression; and the second branch performs basic feature enhancement processing on the splicing features;

[0028] A feature fusion subunit, for dynamically weighting the multi-scale features output by the first branch;

[0029] The residual aggregation subunit performs an element-by-element addition operation on the output of the second branch and the adaptive fusion result to generate a final aggregated feature.

[0030] A satellite and UAV collaborative remote sensing image change detection method, comprising:

[0031] Acquire dual-temporal satellite remote sensing images and drone remote sensing images of the same geographical area;

[0032] The satellite branch undergoes a four-level convolution-pooling operation to output multi-level low-resolution wide-area features. The UAV branch passes through the backbone network to output high-resolution local features that are spatially aligned with the satellite branch.

[0033] Perform channel interaction operations on each level of bi-temporal features to generate a spatial weight map to enhance the change area, use multi-expansion rate dilated convolution and channel attention to fuse contextual information, and weightedly suppress background noise;

[0034] Align adjacent levels of resolution through transposed convolution, dynamically integrate multi-scale features using the attention feature fusion mechanism, and output optimized fusion features through residual connections.

[0035] Perform convolution binary classification on the fused features to generate a change detection result map.

[0036] Furthermore, it also includes loss function construction, including:

[0037] Simultaneously calculate the binary cross entropy loss and intersection-over-union loss of the predicted results and the true labels;

[0038] Dynamically balance the contribution weights of the two types of loss terms through adjustable weight coefficients;

[0039] The weighted summation result is used as the final supervisory signal to drive model optimization.

[0040] Furthermore, a spatial weight map is generated to emphasize the change areas, including:

[0041] The collaborative features output by the high-level layer are learned upsampled and fused with the satellite features and UAV features of the current layer in the channel dimension.

[0042] Split the fusion features into independent channel groups and perform interactive recombination of dual-branch features;

[0043] The reorganized dual-branch features are convolved and the channel dimension is compressed respectively, and then merged and nonlinearly activated to generate a spatial weight map.

[0044] Furthermore, we use multi-expansion rate dilated convolution and channel attention to fuse contextual information, including:

[0045] Concatenate the output fusion features with high-level collaborative features in the channel dimension;

[0046] Reduce the channel dimension of the spliced ​​features through convolution and feature compression operations;

[0047] A channel attention mechanism is used to perform global information aggregation on compressed features and adaptively calibrate channel importance weights;

[0048] Multiple groups of dilated convolutions with different receptive fields are used to extract multi-scale context features in parallel, and the extracted results are fused.

[0049] Furthermore, we align adjacent levels of resolution through transposed convolution, dynamically integrate multi-scale features using the attention feature fusion mechanism, and combine the residual connection to output the optimized fusion features, including:

[0050] Perform learnable upsampling operations on high-level features to align adjacent levels of resolution;

[0051] The multi-level features after spatial alignment are spliced ​​in the channel dimension;

[0052] Process the concatenated features in two parallel branches and perform element-by-element fusion of the output features of the two branches;

[0053] Residual connections are introduced to superimpose the fusion results with the original high-level features and output optimized aggregated features.

[0054] The technical solution of the present invention can achieve the following technical effects:

[0055] It effectively solves the signal attenuation problem caused by the mismatch of heterogeneous data features, significantly reduces the false alarm rate caused by environmental interference, and simultaneously improves the complete detection capability of dynamic targets at all scales from micro to macro. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0057] Figure 1 This is a schematic diagram of a satellite and UAV collaborative remote sensing image change detection system;

[0058] Figure 2 The figure is a flowchart of a method for detecting changes in remote sensing images using collaboration between satellites and UAVs.

[0059] Figure 3 Schematic diagram of the process built for the loss function;

[0060] Figure 4 Schematic diagram of the process for generating spatial weight maps to enhance the areas of change;

[0061] Figure 5 To fuse contextual information using multi-expansion rate dilated convolution and channel attention

[0062] Figure 6 Schematic diagram of the process of aligning adjacent layer resolutions through transposed convolution and dynamically integrating multi-scale features using the attention feature fusion mechanism. DETAILED DESCRIPTION

[0063] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0064] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention pertains. The terms used in this specification are for the purpose of describing specific embodiments only and are not intended to limit the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0065] Example 1

[0066] like Figure 1 As shown, the present invention provides a satellite and UAV collaborative remote sensing image change detection system, comprising:

[0067] A data input module for receiving dual-temporal satellite remote sensing images and UAV remote sensing images of the same geographical area;

[0068] Specifically, the data input module is used to synchronously receive satellite remote sensing images and drone remote sensing images taken at different times from the same geographical area, and to construct dual-temporal heterogeneous remote sensing data pairs for change detection. Satellite remote sensing images can be obtained from public medium- and high-resolution platforms such as Sentinel-2 or the Gaofen series. The image spatial resolution is generally between 10 and 30 meters, with a wide coverage area and multispectral information. UAV remote sensing images are obtained through high-resolution cameras carried by multi-rotor drones. The resolution is usually between 0.05 and 0.2 meters. The shooting area partially overlaps with the satellite image and is used to provide fine local change information. In order to ensure the consistency of the image in time and space, the module performs alignment and cropping operations. First, it performs coarse alignment based on the geographic coordinates in the image metadata, and further achieves fine alignment through feature point matching algorithms such as SIFT or ORB. Spatial alignment is completed in conjunction with affine or perspective transformation. Then, the corresponding area in the drone image is automatically cropped based on the observation area of ​​the satellite image to ensure that the spatial range is completely consistent. To further reduce interference caused by varying imaging conditions, the module uniformly scales images to a standard size (e.g., 256×256 pixels) and performs preprocessing such as grayscale normalization and histogram matching. In a practical application scenario, such as urban building deformation monitoring, the system can automatically call Sentinel-2 images and low-altitude drone images before and after an earthquake. After completing registration and normalization, it outputs paired input images for subsequent feature extraction and change area analysis. By effectively controlling factors such as spatiotemporal dimensions, image resolution, and illumination, this module provides a clear and consistent input foundation for collaborative detection tasks.

[0069] The heterogeneous feature extraction module includes a satellite branch and a UAV branch. The satellite branch consists of a four-level convolutional block and outputs multi-level low-resolution wide-area features. The UAV branch removes the classification layer and outputs local detail features aligned with the satellite branch hierarchy.

[0070] Specifically, the feature extraction module adopts a dual-branch architecture, processing input satellite remote sensing imagery and drone remote sensing imagery, respectively. This module's design fully considers the differences between the two image types in terms of imaging resolution, viewing angle, and semantic density, aiming to provide a structurally consistent and semantically complementary feature foundation for subsequent collaborative analysis tasks. The satellite branch primarily processes relatively low-resolution satellite imagery with wide coverage. Therefore, it employs a four-level convolutional block approach to hierarchical feature extraction. Each convolutional block consists of two consecutive 3×3 standard convolutional layers, followed by batch normalization and ReLU activation, and a max-pooling operation with a stride of 2 for spatial downsampling. After the four-level convolution, the satellite imagery is encoded into feature maps at multiple scales, capturing multiple layers of information, from edge texture to overall regional semantics. These feature maps gradually decrease in spatial resolution and increase in channel dimension, constructing a representative multi-scale low-resolution feature pyramid suitable for expressing macroscopic trends. The drone branch targets drone imagery with higher resolution and richer texture detail. To obtain feature outputs that match the structural hierarchy of the satellite branch, it employs a classic deep convolutional neural network as its backbone architecture. The original classification and global pooling layers are removed, retaining only the first four modules for convolutional encoding. At the output of each module, a convolutional layer adjusts the spatial size and number of channels of the feature map. This ensures that the processed drone image outputs, at each level, have a spatial structure and channel dimensions that match those of the satellite branch. This design ensures structural alignment of the two branches' outputs at multiple scales, providing a unified input foundation for subsequent module processing. In a typical experimental setup, the input satellite image is 256×256 pixels in size, while the original drone image resolution is 1024×1024 pixels. After processing by the feature extraction module, the two branches output feature maps with spatial sizes of 256×256, 128×128, 64×64, and 32×32 at four scale levels, and the number of channels is 64, 128, 256, and 512, respectively, achieving structurally consistent heterogeneous representation. This dual-branch extraction strategy retains both the large-scale global features of satellite images and the local detail information of drone images, and is a key basic step in realizing collaborative change detection in remote sensing images.

[0071] The cross-domain semantic alignment module includes a difference perception enhancement unit and a context enhancement unit. The difference perception enhancement unit is used to perform channel interaction and reorganization of the corresponding hierarchical features of the satellite branch and the UAV branch, generate a spatial weight map and weighted fusion features. The context enhancement unit is used to superimpose the fusion features output by the difference perception enhancement unit with the high-level CDSA features, and enhance the contextual semantics through multi-expansion rate void convolution and channel attention mechanism.

[0072] Specifically, the cross-domain semantic alignment module consists of a difference-aware enhancement unit and a context-aware enhancement unit. These two units work together to mitigate semantic inconsistencies between satellite and drone imagery due to differences in imaging mechanisms, thereby improving the expressiveness and robustness of cross-modal feature fusion. The core task of the difference-aware enhancement unit is to identify and highlight regions with significant changes between the two branches. To this end, at each specific level, the structurally aligned feature maps from the satellite and drone branches are input. A channel-by-channel reorganization mechanism is designed to achieve deep fusion of the two features. The specific process is as follows: The two feature maps are first concatenated along the channel dimension. The concatenated result is then partitioned and reorganized according to specific rules, such as cross-linking or grouped convolution operations, to ensure that the channels from different sources are mutually perceptible and complementary after reorganization. A series of convolution operations are then used to compress the channel dimension while preserving important difference information. Finally, a spatial weight map is generated through a nonlinear activation function. This weight map, acting as a saliency guide, is used to weight the response intensity of each pixel in the original feature map, thereby highlighting regions of potential real change and suppressing background noise. The context enhancement unit further explores the high-order semantic relationships in the fused features to improve the recognition accuracy of the changed areas. Its input is the weighted fused features output by the difference perception enhancement unit and the high-level CDSA features from the deeper layer at the current level. After splicing the two in the channel dimension, the channels are first compressed through convolution and normalization operations. Then, multiple sets of dilation rate dilation operations are introduced in parallel to extract short-range and long-range contextual dependency information respectively. This multi-scale receptive field design enables the model to focus on both local structure and global contour without increasing the number of parameters. In addition, to enhance the selectivity of inter-channel feature expression, a channel attention mechanism is introduced to weight the channel dimension of the fused features, strengthen the channel response with strong semantic consistency, and weaken redundant information.

[0073] The multi-scale feature aggregation module aligns the CDSA feature resolutions of adjacent layers through transposed convolution, uses the attention feature fusion mechanism for cross-layer feature integration, and introduces residual connections to optimize gradient transfer;

[0074] Specifically, first, the module receives multi-level CDSA features output from the cross-domain semantic alignment module. These features have differences in spatial resolution. In order to achieve structural alignment between levels, the module uses a trainable transposed convolution operation to upsample the high-level CDSA features to make their resolution consistent with the adjacent low-level features. The transposed convolution has the ability to learn while restoring the space, and can adaptively adjust the sampling method according to different feature contents. Compared with the traditional interpolation method, it has better effects in retaining structural edge information. After completing the spatial alignment, the features of multiple levels are input into a unified fusion path for cross-scale integration. In order to improve the fusion effect and effectively guide the focus of important information, the module introduces an attention feature fusion mechanism. This mechanism guides the network to focus on areas with more variable discriminative power, dynamically adjusts the feature contribution during the fusion process, and effectively avoids excessive interference of high-level abstract features or noise introduction of underlying features. The attention mechanism can be dynamically adjusted based on channel weights, and can also be combined with spatial distribution to guide The network maintains high response in key areas; to improve the availability of deep features in downstream tasks, the module further introduces a residual connection design, which adds the fused features to the original low-level CDSA features that have not been fused element by element, to achieve structural compensation and direct pass on the gradient path. This residual structure not only helps to retain the fine-grained texture information of the underlying layer, but also alleviates the gradient dissipation problem that occurs in the propagation of deep information flow, ensuring the stability and efficiency of end-to-end training; finally, to optimize the gradient transfer problem in the process of multi-layer feature fusion, the module introduces a residual connection mechanism in the fusion path. The specific approach is to add the attention-fused features to the original low-level CDSA features of the current layer element by element to form a residual path, so that the gradient can be directly propagated back to the lower layer during the backpropagation process, effectively avoiding feature degradation and gradient disappearance. At the same time, the residual structure also retains the edge and detail information in the original features, improving the stability and resolution of the fusion results.

[0075] The change detection output module performs binary classification on the fused features after feature aggregation and generates a pixel-level change detection map.

[0076] Specifically, the change detection output module receives the final fused features output by the multi-scale feature aggregation module and converts them into a pixel-level binary classification result map, indicating whether each location has changed. As the output of the entire system, this module's design focuses on ensuring the accuracy, detail integrity, and spatial consistency of the classification results. In practical implementation, this module first applies a 1×1 convolution to the fused features to compress the high-dimensional feature map into the single-channel or two-channel probability map required for binary classification output. Compared to large-scale convolutions, 1×1 convolutions perform linear combinations only in the channel dimension without increasing the spatial receptive field, effectively preserving the local response in the aggregated features and achieving dimensionality compression for class prediction. The output is a change probability value or a change / unchanged class score for each pixel. The output probability map or score map is then further thresholded (for example, with a threshold of 0.5) to generate a binary change detection map, where "1" indicates that the pixel is considered a changed area and "0" indicates an unchanged area. The entire output maintains the same spatial size as the input image, ensuring that the detection results can be directly aligned with the original image for subsequent visualization, annotation, or quantitative analysis.

[0077] As a preference of the above embodiment, a hybrid loss optimization module is further included, for calculating the IoU-BCE hybrid loss function, the formula is:

[0078] Loss = αLoss BCE +(1-α)Loss iou

[0079] Among them, α is a constant used to balance the IoU weight and BCE weight, α = 0.5, Where N represents the total number of pixels, and t and p represent the ground truth label and the predicted result map respectively. For each pixel I,t i ∈{0,1} represents the ground truth class, 0 represents unchanged, 1 represents changed, and the predicted probability P of pixel i being classified as changed i ∈[0,1], and 1-P i is the predicted probability of no change.

[0080] Specifically, the hybrid loss optimization module is used to guide the training process of the entire change detection network. By combining regional overlap and pixel-level classification error, the model's recognition accuracy and boundary fitting ability for the changed region are enhanced. This module weightedly fuses the IoU loss and the BCE loss to construct a hybrid supervision mechanism that takes into account both regional integrity and category discrimination. In actual implementation, the IoU loss is used to measure the degree of spatial overlap between the model's predicted map and the ground truth label. It pays more attention to the overall structural accuracy of the changed region, which helps to improve the model's ability to fully predict the target contour. The BCE loss calculates the difference between the predicted probability and the true label pixel by pixel, and pays more attention to pixel-level classification accuracy, which helps to optimize the judgment of detailed areas and boundary positions. To achieve an effective balance between the two, the system sets a fixed weight coefficient (preferably 0.5) to weight the two losses separately to form the final hybrid loss function. During the specific training process, the system first calculates the BCE loss based on the prediction result of each pixel and the corresponding true label. This loss uniformly models the changed and unchanged categories and is suitable for binary classification tasks in most cases; at the same time, the IoU loss is calculated based on the intersection over union ratio between the predicted image and the label image, which is mainly used to enhance the sensitivity to sparse change areas (such as the edge of collapsed buildings, small new objects, etc.). The hybrid loss optimization module takes the weighted sum of the two output results as the main objective function in the overall training process.

[0081] As a preferred embodiment of the above embodiment, the difference perception enhancement unit includes:

[0082] The cross-layer feature fusion subunit upsamples the features output by the next layer and reduces the channels by half before superimposing them with the satellite and drone features of the current layer.

[0083] The dual-branch reconstruction subunit splits the superimposed features into two independent channel groups and recombines them interactively;

[0084] The spatial weight generation subunit performs convolution processing on the reorganized two branches and compresses the dimensions. After merging, it generates a spatial weight map through the S-type activation function.

[0085] Specifically, first of all, the design purpose of the cross-layer feature fusion subunit is to introduce cross-level information flow to enhance the semantic expression ability of the current feature. The subunit receives the fused features output from a deeper layer (i.e., the next layer) and uses transposed convolution or bilinear interpolation to perform spatial upsampling to keep its resolution consistent with the current processing layer. The upsampled feature map is compressed in the channel dimension by 1×1 convolution, and the number of channels is halved to reduce the computational complexity and highlight important responses. Subsequently, the upsampling result is spliced ​​or element-by-element added with the satellite features and drone features of the current layer in the channel dimension to form a fused feature map containing semantic information of the upper and lower levels, which serves as the basis for subsequent processing. Next, the dual-branch reconstruction subunit divides the above-mentioned fused feature map into two independent channel groups. This division can be done uniformly or according to specific rules. The purpose is to provide structural independence for subsequent feature interactions. Within each channel group, convolution paths are constructed separately. By performing cross-convolution and information exchange operations on the features, feature reorganization between the two branches is achieved. In the specific implementation, interleaved connections, group convolutions or channel cross-splicing can be used to enhance the ability to perceive differences between features from different sources, making the model more sensitive to subtle semantic shifts and structural disturbances. Finally, the spatial weight generation subunit receives the two feature maps reconstructed by the dual branches, compresses the channel dimension and extracts the spatial response information through convolution operations respectively, and then sums the two element-by-element at the pixel position. The combined result is processed by a nonlinear activation function (preferably an S-type activation function, i.e. Sigmoid) to output a spatial weight map. This weight map is used to mark the potential change-sensitive areas in the image. Its value range is between 0 and 1. The higher the value, the more likely the position belongs to the change area, otherwise it is the background or unchanged area. This weight map will be used in subsequent modules to weight the current layer fusion features to achieve enhanced expression of key change areas.

[0086] As a preferred embodiment of the above, the context enhancement unit includes:

[0087] The channel splicing and dimensionality reduction subunit splices the input features in the channel dimension and compresses the channel dimension to half of the original feature through the convolution-batch normalization-activation function processing process;

[0088] Attention channel, which calibrates the importance weight of each channel through global information aggregation and adaptive reweighting mechanism;

[0089] The context extraction subunit uses three dilated convolutions with different receptive fields to extract multi-scale context features in parallel and fuse the multi-scale features.

[0090] The feature optimization weighted subunit refines the feature expression through 1×1 convolution and uses the spatial weight map generated by the difference perception enhancement unit to perform feature weighting.

[0091] Specifically, the channel concatenation and dimensionality reduction subunit receives the fused features from the difference-aware enhancement unit and the original semantic features of the current level and concatenates them along the channel dimension, forming an intermediate feature with twice the number of channels. To avoid computational redundancy caused by excessive channel dimensions while strengthening the information representation of important channels, this subunit compresses the concatenated features through a sequence of convolutional layers, batch normalization layers, and activation functions (such as ReLU), reducing their channel count to half that of the original features. This dimensionality reduction not only preserves the fused semantics but also improves the compactness of the feature distribution, facilitating the subsequent attention mechanism to accurately model the channels. Next, the attention channel is used to calibrate the importance of different channels in semantic representation. This module first performs a global average pooling operation on the reduced feature maps to extract statistical descriptive information for each channel. Subsequently, through an adaptive reweighting structure consisting of two fully connected layers, a set of normalized channel weight vectors is learned to adjust the response strength of each channel. This channel weighting mechanism automatically suppresses low-value redundant information and highlights semantic channels that are highly relevant to the changing target, making subsequent context modeling more focused and effective. The context extraction subunit models spatial context information at different scales. It applies three dilated convolution kernels with different receptive fields (e.g., dilation ratios of 1, 3, and 5) in parallel to the attention-calibrated feature map. This approach extracts multi-level spatial correlation information while maintaining feature resolution. The three convolution outputs are then integrated into a fused feature representation to characterize the contextual structure of the target at different spatial scales. This design is particularly suitable for remote sensing scenes with diverse change patterns and complex spatial boundaries. Finally, the feature optimization weighting subunit further refines the fused feature. This subunit performs channel compression and semantic refinement through a 1×1 convolution operation. It also incorporates the spatial weight map generated by the difference-aware enhancement unit to weight pixel values ​​at different locations in the feature map, thereby focusing the model on areas of true change in the spatial dimension. This spatial guidance mechanism effectively suppresses spurious changes caused by differences in terrain, illumination, and texture, improving the discriminability and stability of the final feature map.

[0092] As a preferred embodiment of the above, the multi-scale feature aggregation module includes:

[0093] The feature alignment subunit performs a learnable upsampling operation on the input features of adjacent levels to unify the spatial resolution;

[0094] The channel fusion subunit splices the spatially aligned features in the channel dimension;

[0095] The multi-scale extraction subunit includes a first branch and a second branch. The first branch performs a multi-branch heterogeneous convolution operation on the spliced ​​features to generate a multi-scale feature expression; the second branch performs basic feature enhancement processing on the spliced ​​features.

[0096] The feature fusion subunit is used to dynamically fuse the multi-scale features output by the first branch;

[0097] The residual aggregation subunit performs an element-by-element sum operation on the output of the second branch and the adaptive fusion result to generate the final aggregated feature.

[0098] Specifically, first, the feature alignment subunit is used to solve the problem of inconsistent spatial resolution of features at different levels. This subunit uses a learnable transposed convolution operation to upsample the higher-level CDSA feature map to make its spatial size consistent with the adjacent low-level features. By introducing trainable parameters, this upsampling process can not only restore the resolution, but also adaptively retain important structural information in the original features, such as edge contours and change area responses, thereby providing a stable spatial basis for subsequent fusion; then, the channel fusion subunit receives the aligned feature maps and splices them in the channel dimension. This splicing operation retains the semantic differences between different levels and provides richer channel expression capabilities, which is conducive to the comprehensive modeling of multi-scale contextual information. In order to prevent channel redundancy after splicing, this unit usually cooperates with the channel normalization mechanism to unify the feature scale. After the channel fusion is completed, the multi-scale extraction subunit further processes the fused features. This unit consists of two parallel branches: the first branch performs multi-branch heterogeneous convolution operations, applying convolution kernels of different sizes (e.g., 3×3, 5×5, and 7×7) to the same fused feature map to extract semantic features with different receptive fields, thereby constructing a multi-scale feature representation. The second branch uses standard convolution operations to perform basic enhancement on the concatenated features, enhancing their continuity and edge preservation. This dual-branch parallel design ensures both semantic hierarchy expansion and preservation of local details. Subsequently, the feature fusion subunit dynamically fuses the multi-scale features extracted by the first branch. This process is typically accomplished through an attention mechanism, which assigns a set of learnable weights to each branch output, adaptively adjusting the contribution of information at each scale based on contextual responses. The fused result has enhanced change perception and effectively suppresses interference from redundant scale features. Finally, the residual aggregation subunit performs element-wise summation of this fused result with the basic enhanced features output by the second branch. This residual structure not only provides a stable gradient path, facilitating network training, but also effectively compensates for detail loss during the fusion process, further improving the output feature's performance in boundary localization and subtle change recognition.

[0099] Example 2

[0100] Based on the same inventive concept as the satellite and UAV collaborative remote sensing image change detection system in the aforementioned embodiment, the present invention also provides a satellite and UAV collaborative remote sensing image change detection method, such as Figure 2 As shown, the method includes:

[0101] S1: Acquire dual-temporal satellite remote sensing images and UAV remote sensing images of the same geographical area;

[0102] S2: The satellite branch undergoes a four-level convolution-pooling operation to output multi-level low-resolution wide-area features. The UAV branch passes through the backbone network and outputs high-resolution local features that are spatially aligned with the satellite branch.

[0103] S3: Perform channel interaction operations on each level of bi-temporal features to generate a spatial weight map to enhance the change area, use multi-expansion rate dilated convolution and channel attention to fuse contextual information, and weightedly suppress background noise;

[0104] S4: Align adjacent layer resolutions through transposed convolution, dynamically integrate multi-scale features using the attention feature fusion mechanism, and output the optimized fusion features through residual connections;

[0105] S5: Perform convolution binary classification on the fused features to generate a change detection result map.

[0106] Specifically, in step S1, two temporal phases of remote sensing imagery are acquired from the same geographic area, one set from a satellite platform and the other from an unmanned aerial vehicle (UAV) platform. These images undergo spatial registration and preprocessing to ensure effective comparison in subsequent steps. In step S2, the satellite and UAV images are fed into two feature extraction branches, respectively. Four-level convolution and pooling operations are used to extract multi-scale, low-resolution, wide-area features from the satellite images, while a backbone network is used to extract high-resolution, local features that match the resolution of the satellite images. In step S3, each pair of features from the two branches is interactively processed to generate a spatial weight map to enhance the true change regions. Multi-scale contextual information and a channel-wise attention mechanism are also introduced to further enhance the ability to identify change regions in complex backgrounds. In step S4, features from different levels are upsampled and fused to unify their spatial resolution. The attention mechanism dynamically integrates multi-scale information, and the fusion process incorporates residual connections to optimize information transfer and training stability. In step S5, the fused features are fed into a classification module to generate a change detection result map for each pixel, ultimately outputting a pixel-level prediction map that clearly marks the change regions.

[0107] As a preferred embodiment of the above, Figure 3 As shown, it also includes loss function construction, including:

[0108] A10: Simultaneously calculate the binary cross entropy loss and intersection-over-union loss of the predicted results and the true labels;

[0109] A20: Dynamically balances the contribution weights of the two types of loss terms through adjustable weight coefficients;

[0110] A30: Use the weighted summation result as the final supervisory signal to drive model optimization.

[0111] As a preferred embodiment of the above, Figure 4 As shown, a spatial weight map is generated to emphasize the areas of change, including:

[0112] B10: Perform learnable upsampling on the collaborative features output by the high-level layer and perform channel-dimensional fusion with the satellite features and UAV features of the current layer;

[0113] B20: Split the fused features into independent channel groups and perform interactive recombination of dual-branch features;

[0114] B30: The reorganized dual-branch features are convolved separately and the channel dimension is compressed, and then merged and nonlinearly activated to generate a spatial weight map.

[0115] As a preferred embodiment of the above, Figure 5 As shown in Figure 2, we use multi-expansion rate dilated convolution and channel attention to fuse contextual information, including:

[0116] C10: Concatenate the output fusion features with the high-level collaborative features in the channel dimension;

[0117] C20: Reduce the channel dimension of the concatenated features through convolution and feature compression operations;

[0118] C30: Uses a channel attention mechanism to perform global information aggregation on compressed features and adaptively calibrates channel importance weights;

[0119] C40: Use multiple sets of dilated convolutions with different receptive fields to extract multi-scale contextual features in parallel and fuse the extracted results.

[0120] As a preferred embodiment of the above, Figure 6 As shown in the figure, the adjacent layer resolutions are aligned by transposed convolution, the multi-scale features are dynamically integrated by the attention feature fusion mechanism, and the optimized fusion features are output by residual connection, including:

[0121] S41: Perform a learnable upsampling operation on high-level features to align adjacent levels of resolution;

[0122] S42: Splice the multi-level features after spatial alignment in the channel dimension;

[0123] S43: Process the splicing features in parallel using two branches, and perform element-by-element fusion of the output features of the two branches;

[0124] S44: Introduce residual connection to superimpose the fusion result with the original high-level features and output optimized aggregated features.

[0125] The above detection method in the present invention can be effectively applied to the satellite and UAV collaborative remote sensing image change detection system, and the technical effects that can be achieved are as described in the above embodiments and will not be repeated here.

[0126] Although the present application has been described with reference to specific features and embodiments thereof, it is apparent that various modifications and combinations thereof may be made without departing from the spirit and scope of the present application. Accordingly, this specification and the drawings are merely illustrative of the present application as defined herein, and are intended to cover any and all modifications, variations, combinations or equivalents within the scope of the present application.

[0127] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalents, the present application is intended to include these modifications and variations.

Claims

1. A satellite and UAV collaborative remote sensing image change detection system, characterized in that: include: A data input module for receiving dual-temporal satellite remote sensing images and UAV remote sensing images of the same geographical area; A heterogeneous feature extraction module, comprising a satellite branch and a drone branch. The satellite branch consists of four-level convolutional blocks and outputs multi-level low-resolution wide-area features. The drone branch removes the classification layer and outputs local detail features aligned with the satellite branch hierarchy. The cross-domain semantic alignment module includes a difference perception enhancement unit and a context enhancement unit. The difference perception enhancement unit is used to perform channel interaction reorganization on the corresponding hierarchical features of the satellite branch and the UAV branch, generate a spatial weight map and weighted fusion features. The context enhancement unit is used to superimpose the fusion features output by the difference perception enhancement unit with the high-level CDSA features, and enhance the contextual semantics through multi-expansion rate void convolution and channel attention mechanism. The multi-scale feature aggregation module aligns the CDSA feature resolutions of adjacent layers through transposed convolution, uses an attention feature fusion mechanism for cross-layer feature integration, and introduces residual connections to optimize gradient transfer. The change detection output module performs binary classification on the fused features after feature aggregation and generates a pixel-level change detection map.

2. The satellite and UAV collaborative remote sensing image change detection system according to claim 1, characterized in that: It also includes a hybrid loss optimization module for calculating the IoU-BCE hybrid loss function, the formula is: Loss=αLoss BCE +(1-α)Loss iou Among them, α is a constant used to balance the IoU weight and BCE weight, α = 0.5, Where N represents the total number of pixels, and t and p represent the ground truth label and the predicted result map respectively. For each pixel I,t i ∈{0,1} represents the groundtruth class, 0 represents unchanged, 1 represents changed, and the predicted probability P of pixel i being classified as changed i ∈[0,1], and 1-P i is the predicted probability of no change.

3. The satellite and UAV collaborative remote sensing image change detection system according to claim 1, characterized in that: The difference perception enhancement unit includes: The cross-layer feature fusion subunit upsamples the features output by the next layer and reduces the channels by half, and then superimposes them with the satellite and drone features of the current layer; The dual-branch reconstruction subunit splits the superimposed features into two independent channel groups and recombines them interactively; The spatial weight generation subunit performs convolution processing on the reorganized two branches and compresses the dimensions. After merging, it generates a spatial weight map through the S-type activation function.

4. The satellite and UAV collaborative remote sensing image change detection system according to claim 1, characterized in that: The context enhancement unit includes: The channel splicing and dimensionality reduction subunit splices the input features in the channel dimension and compresses the channel dimension to half of the original feature through the convolution-batch normalization-activation function processing process; Attention channel, which calibrates the importance weight of each channel through global information aggregation and adaptive reweighting mechanism; The context extraction subunit uses three dilated convolutions with different receptive fields to extract multi-scale context features in parallel and fuse the multi-scale features. The feature optimization weighting subunit refines the feature expression through 1×1 convolution and performs feature weighting using the spatial weight map generated by the difference perception enhancement unit.

5. The satellite and UAV collaborative remote sensing image change detection system according to claim 1, characterized in that: The multi-scale feature aggregation module includes: The feature alignment subunit performs a learnable upsampling operation on the input features of adjacent levels to unify the spatial resolution; The channel fusion subunit splices the spatially aligned features in the channel dimension; The multi-scale extraction subunit includes a first branch and a second branch, wherein the first branch performs a multi-branch heterogeneous convolution operation on the splicing features to generate a multi-scale feature expression; and the second branch performs basic feature enhancement processing on the splicing features; A feature fusion subunit, for dynamically weighting the multi-scale features output by the first branch; The residual aggregation subunit performs an element-by-element addition operation on the output of the second branch and the adaptive fusion result to generate a final aggregated feature.

6. A satellite and UAV collaborative remote sensing image change detection method, characterized in that: include: Acquire dual-temporal satellite remote sensing images and drone remote sensing images of the same geographical area; The satellite branch undergoes a four-level convolution-pooling operation to output multi-level low-resolution wide-area features. The UAV branch passes through the backbone network to output high-resolution local features that are spatially aligned with the satellite branch. Perform channel interaction operations on each level of bi-temporal features to generate a spatial weight map to enhance the change area, use multi-expansion rate dilated convolution and channel attention to fuse contextual information, and weightedly suppress background noise; Align adjacent levels of resolution through transposed convolution, dynamically integrate multi-scale features using the attention feature fusion mechanism, and output optimized fusion features through residual connections. Perform convolution binary classification on the fused features to generate a change detection result map.

7. The satellite and UAV collaborative remote sensing image change detection method according to claim 6, characterized in that: It also includes loss function construction, including: Simultaneously calculate the binary cross entropy loss and intersection-over-union loss of the predicted results and the true labels; Dynamically balance the contribution weights of the two types of loss terms through adjustable weight coefficients; The weighted summation result is used as the final supervisory signal to drive model optimization.

8. The satellite and UAV collaborative remote sensing image change detection method according to claim 6, characterized in that: Generate spatial weight maps to emphasize areas of change, including: The collaborative features output by the high-level layer are learned upsampled and fused with the satellite features and UAV features of the current layer in the channel dimension. Split the fusion features into independent channel groups and perform interactive recombination of dual-branch features; The reorganized dual-branch features are convolved and the channel dimension is compressed respectively, and then merged and nonlinearly activated to generate a spatial weight map.

9. The satellite and UAV collaborative remote sensing image change detection method according to claim 6, characterized in that: Multi-expansion rate dilated convolution and channel attention are used to fuse contextual information, including: Concatenate the output fusion features with high-level collaborative features in the channel dimension; Reduce the channel dimension of the spliced ​​features through convolution and feature compression operations; A channel attention mechanism is used to perform global information aggregation on compressed features and adaptively calibrate channel importance weights; Multiple groups of dilated convolutions with different receptive fields are used to extract multi-scale context features in parallel, and the extracted results are fused.

10. The satellite and UAV collaborative remote sensing image change detection method according to claim 6, characterized in that: The resolution of adjacent layers is aligned through transposed convolution, and the attention feature fusion mechanism is used to dynamically integrate multi-scale features. The optimized fusion features are output through residual connection, including: Perform learnable upsampling operations on high-level features to align adjacent levels of resolution; The multi-level features after spatial alignment are spliced ​​in the channel dimension; Process the concatenated features in two parallel branches and perform element-by-element fusion of the output features of the two branches; Residual connections are introduced to superimpose the fusion results with the original high-level features and output optimized aggregated features.

Images