Remote sensing image change detection method and system based on edge perception multi-scale difference

Abstract

This invention relates to the field of remote sensing image change detection technology, specifically to a method and system for remote sensing image change detection based on edge-aware multi-scale difference. The method includes: acquiring a pair of dual-temporal remote sensing images and a trained edge-aware multi-scale difference network; inputting the dual-temporal remote sensing images into an edge-aware central difference feature encoder; preserving high-frequency edge information and extracting multi-scale edge features through a three-level cascaded central difference convolution during the input stage; inputting the multi-scale edge features into a multi-scale spatiotemporal transform module to generate differential enhancement features; inputting the differential enhancement features into an adaptive region-aware progressive decoder; progressively fusing the features from coarse to fine through region-aware attention; and outputting the change detection result after multi-level supervision. This invention solves the problems of boundary blurring, false change detection, and missed detection of small targets in high-resolution remote sensing image change detection, improving the accuracy and robustness of change detection.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image change detection technology, specifically to a remote sensing image change detection method and system based on edge-aware multi-scale difference. Background Technology

[0002] Remote sensing image change detection has become a key technology for monitoring Earth's surface dynamics, widely applied in urban expansion monitoring, disaster assessment, environmental protection, and land resource management. High-resolution satellite systems such as WorldView, GeoEye, and Sentinel provide sub-meter spatial resolution images and short observation cycles, creating favorable conditions for fine-grained change analysis. Current urban planning, disaster response, and environmental protection all rely on remote sensing image change detection algorithms for monitoring and decision-making. However, achieving robust and high-precision change detection results in complex real-world scenarios remains a significant challenge, particularly in monitoring scenarios with varying lighting conditions, seasonal changes, and different types of landforms.

[0003] While high-resolution remote sensing imagery brings rich spatial details, it also presents numerous technical challenges. Firstly, pseudo-change phenomena are a major source of false positives due to non-semantic factors: differences in imaging phases can cause variations in surface reflectance of up to 20%-30%; seasonal vegetation phenology can easily cause spectral shifts similar to surface changes; atmospheric haze and cloud shadows can lead to local brightness differences; and calibration differences between sensors on different platforms can introduce systemic biases. Secondly, actual changes often take many forms: urban changes include the construction and demolition of single buildings (occupying only 10-50 pixels) to large building complexes, as well as subtle changes in geometric details; natural changes include vegetation growth and deforestation, seasonal shifts in water boundaries, and the conversion of farmland into urban areas. The coexistence of these pseudo-changes and heterogeneous real changes, coupled with the unavoidable geometric and registration errors in multi-temporal imaging, makes maintaining high recall while ensuring high accuracy extremely difficult.

[0004] Despite years of development, existing change detection methods still suffer from fundamental limitations. Traditional methods rely on handcrafted features and rules, lacking the ability to understand complex semantics and spatial context, making it difficult to stably distinguish between real and pseudo-change regions in high-resolution images. Deep learning has significantly improved change detection capabilities through hierarchical feature learning: Convolutional Neural Networks (CNNs) establish a data-driven foundation for hierarchical feature extraction from raw images, with Siamese architecture and explicit differencing being particularly effective in highlighting temporal differences. However, common CNN backbones use large convolutional kernels and pooling in shallow layers, equivalent to low-pass filtering, which irreversibly weakens high-frequency edge information crucial for boundary localization. While Transformer-based self-attention layers can provide global contextual information for better network learning, most existing attention methods rely on cumbersome lexical representations, resulting in a quadratic increase in computational cost, and such block operations can easily blur spatial details. Another problem is that current network designs often use multi-scale fusion at the decoding end to perceive the location of different regions in the image as an average, leading to sparse change signals being easily submerged by the dominant background gradient under extreme class imbalances.

[0005] The aforementioned structural problems can be summarized into three main challenges: (i) the degradation of the shallow edge information extraction capability of the encoder leads to blurred boundaries and missed detection of small targets; (ii) insufficient modeling of spectral-spatial-temporal differences makes it susceptible to spurious changes caused by illumination and seasons; and (iii) unified multi-scale fusion under extreme class imbalance makes sparse variation signals dominated by background information and difficult to detect. To address these three technical problems, this invention proposes an edge-aware multi-scale difference network (EMDNet), which uses ECF-ResNet18 to preserve boundary information at the encoder input; employs MSST-Module for joint spectral-spatial-temporal inference to suppress spurious changes; and uses ARP-Decoder for progressive fusion of region perception and multi-level supervision to address class imbalance. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a remote sensing image change detection method and system based on edge-aware multi-scale difference, aiming to solve the problems of edge degradation, false change detection, and class imbalance in high-resolution remote sensing image change detection, and significantly improve the accuracy and robustness of change detection.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] On one hand, embodiments of the present invention provide a remote sensing image change detection method based on edge-aware multi-scale difference, the method comprising the following steps:

[0009] S100, acquire a pair of dual-temporal remote sensing images and a trained edge-aware multi-scale difference network, wherein the edge-aware multi-scale difference network includes an edge-aware center difference feature encoder, a multi-scale spatiotemporal transformer module and an adaptive region-aware progressive decoder.

[0010] S200, the dual-temporal remote sensing images are respectively input into the edge-aware center difference feature encoder. High-frequency edge information is retained in the input stage through three-level cascaded center difference convolution, and multi-scale edge features are extracted through pooling layers and residual blocks.

[0011] S300, the multi-scale edge features are input into the multi-scale spatiotemporal transformer module, and joint spectral-spatial-temporal inference is performed through multi-scale spectral pooling, dual-temporal self-attention and adaptive feature fusion to generate differential enhancement features;

[0012] S400, the differential enhancement features are input into the adaptive region-aware progressive decoder. The features are divided into change paths, background paths and global paths through region-aware attention and processed separately. Progressive fusion from coarse to fine is performed, and multi-level supervision is applied to output the change detection results.

[0013] On the other hand, embodiments of the present invention provide a remote sensing image change detection system based on edge-aware multi-scale difference, comprising:

[0014] At least one processor;

[0015] At least one memory for storing at least one program;

[0016] When the at least one program is executed by the at least one processor, the at least one processor performs the method described above.

[0017] On the other hand, embodiments of the present invention provide a computer-readable storage medium storing a processor-executable program, which, when executed by a processor, is used to perform the above-described method.

[0018] The embodiments of the present invention have the following beneficial effects:

[0019] The edge-aware center differential feature encoder (ECF-ResNet18) proposed in this invention adopts a three-level cascaded center differential convolution at the encoder input to form a learnable high-pass filter, which retains high-frequency edge information in the shallow input stage, effectively solving the problem of edge information attenuation caused by early low-pass filtering in traditional encoders, and significantly improving the boundary positioning accuracy and small target detection effect.

[0020] The multi-scale spatiotemporal transformer module (MSST-Module) designed in this invention achieves joint spectral-spatial-temporal inference through multi-scale spectral pooling, dual-temporal self-attention and adaptive feature fusion. It can effectively distinguish between real changes and pseudo changes caused by factors such as illumination and seasons, and improve the robustness of the model in complex scenarios.

[0021] The adaptive region-aware progressive decoder (ARP-Decoder) designed in this invention divides features into change / background / global pathways for separate processing. It employs boundary-focused attention for change regions and noise suppression for background regions through region-aware attention. Furthermore, it adopts a coarse-to-fine progressive fusion strategy and a multi-level supervision mechanism to effectively address the problem of extreme class imbalance and significantly improve the detection capability of sparse change regions.

[0022] This invention achieved excellent performance on three public datasets: LEVIR-CD, SYSU-CD, and WHU-CD, with F1 scores of 91.09%, 83.01%, and 90.70%, respectively. This verifies the effectiveness of edge-preserving coding, spectral-temporal difference modeling, and region-aware fusion, providing a new technical approach for change detection in high-resolution remote sensing images. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the remote sensing image change detection method based on edge-aware multi-scale difference in an embodiment of the present invention.

[0025] Figure 2 This is a diagram showing the overall structure of the Edge-Aware Multi-Scale Differential Network (EMDNet) in an embodiment of the present invention.

[0026] Figure 3 This is a structural diagram of the Multi-Scale Spacetime Transformer Module (MSST-Module) in an embodiment of the present invention;

[0027] Figure 4 This is a structural diagram of the Multiscale Spectral-Spatial Attention Module (MSSA-Module) in an embodiment of the present invention;

[0028] Figure 5 This is a structural diagram of the Adaptive Region Attention (ARA) module in an embodiment of the present invention;

[0029] Figure 6This is a schematic diagram illustrating the qualitative results of change detection using EMDNet proposed in this embodiment of the invention on the LEVIR-CD dataset. Detailed Implementation

[0030] The following will provide a clear and complete description of the concept, specific structure, and technical effects of the present invention in conjunction with embodiments and accompanying drawings, so as to fully understand the purpose, solution, and effects of the present invention. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of this invention; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this invention as detailed in the appended claims.

[0032] It is understood that the terms "first," "second," etc., used in this invention may be used to describe various concepts, but unless specifically stated otherwise, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of embodiments of this invention, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words "if" or "when" as used herein may be interpreted as "when," "in response to determination," or "in the event of a determination."

[0033] The terms “at least one,” “multiple,” “each,” “any,” etc., used in this invention, “at least one” includes one, two, or more than two; “multiple” includes two or more than two; “each” refers to each of the corresponding multiple; and “any” refers to any one of the multiple.

[0034] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this invention is for descriptive purposes only and is not intended to limit the invention.

[0035] refer to Figure 1 ,like Figure 1 The figure shown is a remote sensing image change detection method based on edge-aware multi-scale difference provided by an embodiment of the present invention. The method includes the following steps:

[0036] S100, acquire a pair of dual-temporal remote sensing images and a trained edge-aware multi-scale difference network, wherein the edge-aware multi-scale difference network includes an edge-aware center difference feature encoder, a multi-scale spatiotemporal transformer module and an adaptive region-aware progressive decoder.

[0037] S200, the dual-temporal remote sensing images are respectively input into the edge-aware center difference feature encoder. High-frequency edge information is retained in the input stage through three-level cascaded center difference convolution, and multi-scale edge features are extracted through pooling layers and residual blocks.

[0038] S300, the multi-scale edge features are input into the multi-scale spatiotemporal transformer module, and joint spectral-spatial-temporal inference is performed through multi-scale spectral pooling, dual-temporal self-attention and adaptive feature fusion to generate differential enhancement features;

[0039] S400, the differential enhancement features are input into the adaptive region-aware progressive decoder. The features are divided into change paths, background paths and global paths through region-aware attention and processed separately. Progressive fusion from coarse to fine is performed, and multi-level supervision is applied to output the change detection results.

[0040] This invention provides a method and system for remote sensing image change detection based on edge-aware multi-scale differential encoding. By using an edge-aware center differential feature encoder to retain high-frequency edge information during the input stage, it effectively solves the edge information attenuation problem caused by early low-pass filtering in traditional encoders. Through a multi-scale spatiotemporal transformer module, it performs joint spectral-spatial-temporal inference, effectively distinguishing between real changes and pseudo-changes caused by factors such as illumination and season. An adaptive region-aware progressive decoder performs progressive fusion and multi-level supervision of region perception, effectively addressing the problem of extreme class imbalance. This invention effectively addresses core challenges in high-resolution remote sensing image change detection, such as boundary blurring, false change detection, and missed detection of small targets, significantly improving the accuracy and robustness of change detection. It provides more reliable technical support for applications such as urban expansion monitoring, disaster assessment, environmental protection, and land resource management.

[0041] The core of this embodiment lies in the Edge-aware Multi-scale Difference Network (EMDNet) framework, which achieves high-precision change detection through edge-aware encoding, spatiotemporal reasoning, and region-aware decoding.

[0042] Data Acquisition and Network Preparation: Acquire a pair of dual-temporal remote sensing images, T1 and T2. Load a pre-trained edge-aware multi-scale difference network.

[0043] like Figure 2As shown, the method proposed in this invention employs a twin encoder-decoder process, with the dual-temporal images sequentially passing through three key modules. The network input is a pair of dual-temporal remote sensing images T1 and T2 (both with dimensions H×W×C), and the output is a variation map with dimensions H×W×1.

[0044] Edge-aware center differential feature encoder (ECF-ResNet18): The improved ResNet18 structure is used to extract multi-scale edge features from dual-temporal images. The encoder front end is designed with an edge-aware center differential convolution module (ECF-Module), in which a three-level cascaded center differential convolution is used to preserve edge information.

[0045] Multi-scale Spatiotemporal Transformer Module (MSST-Module): Performs joint spectral-spatial-temporal reasoning between the encoder and decoder, and distinguishes between real changes and pseudo changes caused by illumination through multi-scale spectral pooling, dual-temporal self-attention and adaptive feature fusion.

[0046] Adaptive Region-Aware Progressive Decoder (ARP-Decoder): It divides features into variation / background / global pathways and processes them separately, achieving coarse-to-fine cross-scale fusion and applying multi-level supervision.

[0047] This method mainly includes three stages: edge-aware feature extraction, multi-scale spatiotemporal reasoning, and region-aware decoding.

[0048] In some embodiments, S200, the step of preserving high-frequency edge information at the input stage through a three-level cascaded central difference convolution and extracting multi-scale edge features via pooling layers and residual blocks includes:

[0049] S210, the two temporal remote sensing images in the dual-temporal remote sensing image are independently transformed by the first layer of central difference convolution, and the first edge feature maps of the two temporal phases are output accordingly.

[0050] S220, the first edge feature map is transformed by the second layer of central difference convolution to output the second edge feature map;

[0051] S230, the second edge feature map is transformed by the third layer of central difference convolution to output the edge-aware feature map; wherein, each layer of central difference convolution first performs a 3×3 convolution, and then subtracts the depth 1×1 low-frequency estimate constructed by weight statistics to form a learnable high-pass filter.

[0052] S240, after the edge-aware feature map is processed by the pooling layer, it is sequentially passed through four residual stages to extract multi-scale edge features at four resolutions. The four resolutions are H / 2×W / 2, H / 4×W / 4, H / 8×W / 8, and H / 16×W / 16, with corresponding channel numbers of 64, 128, 256, and 512, respectively, where H and W are the height and width of the input image.

[0053] Specifically, current mainstream standard encoders use 7×7 convolutions (stride 2) in the input layer. However, this early low-pass and downsampling is irreversible. Once edge details are smoothed and aliased, subsequent layers struggle to recover mutual information with the true boundaries. This design is acceptable for image classification requiring only coarse semantics, but it is not optimal for change detection that requires precise boundary localization. Most change detection networks directly use this type of classification backbone (ResNet, VGG, Transformer, etc. encoders) without modifying the input level, only refining edges at the back end through multi-scale fusion or dedicated thinning heads. These strategies work on already blurred features and cannot fully recover fine boundaries.

[0054] In contrast, the edge-aware center-difference feature encoder employs an edge-aware center-difference convolutional module followed by multiple residual stages, ensuring that edge information is still well received at the network's end. The core component within the module is the center-difference convolution (CDConv), which highlights local intensity differences. In the one-dimensional case, the classical center-difference operator approximates the first derivative, with its discrete Fourier amplitude frequency being 0 at zero and increasing with frequency. Extending this to two dimensions, it suppresses responses in slowly varying regions and generates large responses at edges. The CDConv block designed in this invention first performs a 3×3 convolution, then subtracts a 1×1 low-frequency estimate constructed from weight statistics, thereby achieving a learnable high-pass filter while preserving edge-related mid-frequency frequencies.

[0055] Overall, the effective spatial support of this central difference convolutional input layer is comparable to that of a 7×7 convolutional kernel. However, due to the cascading of the central difference construction and the high-pass response, its overall transfer function has stronger low-frequency suppression and more complete retention of edge-related mid-frequency information. The edge-aware feature map output from the input layer is then processed through pooling and multiple residual stages to obtain a four-layer feature pyramid. The multi-scale edge features from two temporal phases are then fed into a multi-scale spatiotemporal transformer module for differential modeling.

[0056] In some embodiments, in S300, the step of inputting the multi-scale edge features into the multi-scale spatiotemporal transformer module, and generating differentially enhanced features through joint spectral-spatial-temporal inference via multi-scale spectral pooling, dual-temporal self-attention, and adaptive feature fusion, includes:

[0057] S310 processes the dual-temporal features contained in the multi-scale edge features of each scale through three parallel paths, and outputs the preliminary fusion features, temporal enhancement features and single-temporal refinement features respectively.

[0058] S320, after concatenating the preliminary fusion feature with the single-phase refined feature, one path is generated by pooling and Softmax normalization to produce channel adaptive weights, and the other path is generated by Sigmoid activation to produce spatial adaptive weights. The channel adaptive weights and the spatial adaptive weights are multiplied to obtain the adaptive weights.

[0059] S330, Multiply the adaptive weights element-wise with the preliminary fusion features to obtain the weighted concatenation features;

[0060] S340, the temporal enhancement feature and the weighted concatenation feature are added element by element to obtain the differential enhancement feature.

[0061] Specifically, such as Figure 3 As shown, after extracting edge features from dual-temporal images using an edge-aware center differential feature encoder, these features need to be transformed into a differential representation that highlights real changes and suppresses pseudo-changes. While simple absolute value differencing or convolution after stitching can highlight intensity differences, it is sensitive to illumination and seasonal influences and does not explicitly utilize the joint relationship between spectrum, space, and temporal sequence. Therefore, this scheme inserts a multi-scale spatiotemporal transformer module between the backbone network and the decoder. This module suppresses pseudo-changes in three ways: (i) multi-scale spectral pooling: utilizing the fact that illumination primarily changes the overall intensity while maintaining relative spectral relationships, and that changes in ground features alter spectral morphology at multiple spatial scales; (ii) self-attention modeling of temporal context: real structural changes are spatially correlated, while pseudo-changes are mostly spatially uncorrelated; and (iii) adaptive fusion and learnable gating, weighting the contributions of different feature streams.

[0062] The multi-scale spatiotemporal transformer module processes dual-temporal features at each scale through three parallel paths. The upper path concatenates the dual-temporal features and then performs five layers of 1×1 convolution to preserve temporal and symbolic information. The middle path calculates the differential features and obtains temporal enhancement features through multi-scale spectral pooling, spatial attention, self-attention, and feedforward. The lower path extracts the first-temporal features through five layers of 1×1 convolution. Overall, it emphasizes the temporal enhancement features and the element-wise summation of the concatenated features by adaptive weights, thereby achieving the fusion of temporal context, spatial coherence, and the original signal to realize robust change detection.

[0063] In some embodiments, in S310, the multi-scale edge features at each scale are processed through three parallel paths, which contain dual-temporal features, to output preliminary fusion features, temporal enhancement features, and single-temporal refined features, respectively, including:

[0064] The upper path concatenates the two temporal features along the channel dimension and then processes them through five layers of 1×1 convolution to output preliminary fused features; the two temporal features are the corresponding scale feature pairs extracted from the two temporal remote sensing images by the edge-aware center differential feature encoder, including the first temporal feature and the second temporal feature;

[0065] The intermediate path calculates the difference features of the dual-temporal features, which are then passed through the multi-scale spectral-spatial attention module to output spectral-spatial enhancement features, and then through the temporal context modeling module to obtain temporal enhancement features; the difference features are the absolute value difference between the first temporal feature and the second temporal feature by subtracting each element from the first temporal feature;

[0066] The lower path performs a five-layer 1×1 convolution feature extraction on the first phase feature in the dual-temporal features, and outputs a single-temporal refined feature; the single-temporal feature is either the first phase feature or the second phase feature in the dual-temporal remote sensing image.

[0067] In some embodiments, the sequential output of spectral-spatial enhancement features via a multi-scale spectral-spatial attention module includes:

[0068] S311, the differential features are input into the spectral branch, and adaptive average pooling with three windows (1×1, 3×3, and 5×5) is used to obtain multi-scale spectral information. The 1×1 pooling result is then averaged along the width and height and concatenated along the spatial dimension, followed by 1×1 convolution dimensionality reduction and Sigmoid activation to obtain channel attention. The 3×3 and 5×5 pooling results are added and upsampled back to the original resolution to obtain multi-scale contextual information.

[0069] S312, the differential features are input into the spatial branch, projected into a query vector, a key vector and a value vector by three 1×1 convolutions, and obtained by scaling dot product attention to obtain spatial weighted features;

[0070] S313, the channel attention and spatial weighted features are fused to output spectral-spatial enhancement features.

[0071] Specifically, such as Figure 4 As shown, the Multi-Scale Spectral-Spatial Attention Module (MSSA-Module) takes differential features as input and outputs enhanced features of the same size. This module captures spectral and spatial information through two parallel branches. The spectral branch uses adaptive average pooling with 1×1, 3×3, and 5×5 windows to obtain multi-scale spectral information. The 1×1 pooling result is then averaged along width and height and concatenated along the spatial dimension. After 1×1 convolution dimensionality reduction and sigmoid activation, channel attention is obtained to distinguish between changes in ground features and changes in illumination. The 3×3 and 5×5 pooling results are added and upsampled back to the original resolution to obtain multi-scale contextual information.

[0072] The spatial branch projects the differential features into a query (Q), key (K), and value (V) through three 1×1 convolutions. Scaling dot product attention then yields spatially weighted features, highlighting spatially coherent change patterns. The final output is the sum of the differential features and a learnable scalar (the sum of channel-weighted spatial features and multi-scale context terms), injecting channel weighting and multi-scale spatial context while preserving the original differential signal.

[0073] In some embodiments, the process of obtaining temporal enhancement features via the temporal context modeling module includes:

[0074] S314, the spectral-spatial enhancement features are flattened into a sequence and then projected into a query vector, a key vector, and a value vector;

[0075] S315 establishes the correspondence between the two temporal feature states by scaling the dot product attention, and obtains the attention output;

[0076] S316, the attention output is processed through a linear layer, residual connections and a feedforward network to obtain temporal enhancement features.

[0077] Specifically, the Temporal Context Modeling Module (TCM-Module) establishes a correspondence between bi-temporal feature states through self-attention. Real structural changes are spatially coherent (neighboring pixels change together), while pseudo-changes caused by transient phenomena are mostly spatially uncorrelated. The spectral-spatial enhanced features output by the multi-scale spectral-spatial attention module are flattened into a sequence and projected as query, key, and value. After scaling dot product attention to obtain the attention output, the temporal enhanced features are obtained through linear layer residual connections and a feedforward network (1×1 convolution and GELU activation layer) and a learnable scalar, thus completing the temporal context modeling.

[0078] In some embodiments, S400, the step of dividing features into change paths, background paths, and global paths through region-aware attention and processing them separately includes:

[0079] S410 reduces the dimensionality of multi-scale difference features to 128 dimensions through 3×3 convolution, batch normalization, and ReLU activation.

[0080] S420 inputs the normalized features into the adaptive region attention module, which processes them through four parallel paths: Path 1 is a 1×1 convolution, a depthwise separable convolution, and a 1×1 convolution; Paths 2 and 3 each generate a gated mask; Path 4 is a 1×1 convolution, a depthwise separable convolution, a 1×1 convolution, and a 1×1 convolution.

[0081] The S430 combines the four outputs element-wise through multiplication and addition, then performs 3×3 convolution, ReLU, 1×1 convolution, and batch normalization to obtain refined features for region awareness. It applies boundary focusing attention to changing regions and performs noise suppression on background regions.

[0082] Specifically, the Adaptive Region-Aware Progressive Decoder (ARP-Decoder) addresses the problem of multi-scale fusion under extreme region imbalance. In change detection, the vast majority of pixels are unchanged (background), with only a few showing changes. If the changed regions and background regions are treated the same and fused uniformly, the dominant background signal will suppress the sparse change features during aggregation such as averaging or stitching. Most existing decoders treat all spatial locations equally, resulting in background dominance during fusion and the suppression of small targets and fine boundaries. Some methods use weighted loss or focal loss, but these operate at the loss level rather than the feature representation, and intermediate features may still be dominated by the background.

[0083] Adaptive region-aware progressive decoders can highlight the representation of changing regions and suppress noise in background regions, thereby addressing class imbalance at the architectural level. The adaptive region-aware progressive decoder incorporates three design elements: region-aware attention (for background regions and global context); coarse-to-fine progressive cross-scale fusion; and multi-level supervision to provide auxiliary training signals at each scale to enhance feature learning and convergence.

[0084] The multi-scale spatiotemporal transform module outputs multi-scale difference feature channels of 64, 128, 256, and 512, respectively. All scale difference features are uniformly represented in 128 dimensions after 3×3 convolution, batch normalization (BN), and ReLU. This unifies the dimensions across scales for easier weight sharing and smooths high-frequency noise while preserving structural boundaries through 3×3 convolution. A soft region mask (two layers of inner convolution, ReLU activation, and Sigmoid activation) is predicted for each normalized feature to indicate the probability of change.

[0085] refer to Figure 5 Adaptive Region Attention (ARA) processes each normalized feature through four parallel paths (including depthwise separable convolution DSConv and gating) to obtain refined region-aware features: Path 1 consists of 1×1 convolution, DSConv, and 1×1 convolution; Paths 2 and 3 each generate gating masks; Path 4 consists of 1×1 convolution, DSConv, 1×1 convolution, DSConv, and 1×1 convolution. After element-wise multiplication and addition, the four paths are combined and then processed by 3×3 convolution, ReLU, 1×1 convolution, and batch normalization (BN) to obtain refined features. Boundary-focused attention is applied to change regions, while noise-suppressing attention is applied to background regions, thus addressing class imbalance from an architectural perspective.

[0086] In some embodiments, S400, the step of performing coarse-to-fine progressive fusion, applying multi-level supervision, and outputting change detection results includes:

[0087] S440 starts from the coarsest scale, normalizes the differential features, and then obtains refined features through adaptive region attention;

[0088] S450 directly upsamples the refined features of the deepest scale; for shallower scales, it upsamples the refined features of the previous scale and concatenates them with the refined features of the current scale, then performs 3×3 convolution, batch normalization, ReLU activation, and finally adds them to the refined features of the current scale according to learnable weights to obtain the refined output of the current scale.

[0089] S460 adds auxiliary prediction heads at each scale for multi-level supervision, and finally upsamples all refined features at all scales to the original resolution and stitches them together. After convolutional dimensionality reduction, it outputs a binary classification result.

[0090] Specifically, the decoding process employs a coarse-to-fine progressive strategy: starting from the deepest scale (i=4), each scale normalizes the differential features and then refines them through adaptive region attention to obtain refined features; the deepest scale is directly upsampled; for i=3, 2, and 1, the refined features from the previous scale are upsampled and concatenated with the refined features of the current scale, then processed through 3×3 convolution, batch normalization (BN), ReLU activation, and finally added to the refined features of the current scale with learnable weights to obtain the refined output of the current scale. The coarse scale first establishes semantic understanding of the changing positions, and then guides the fine scale to refine the boundaries within the identified regions. Each scale is supplemented with an auxiliary prediction head (3×3 convolution, batch normalization (BN), ReLU activation, 1×1 convolution) for multi-level supervision. Finally, the refined features from all scales are upsampled to the original resolution and concatenated (512 channels), then processed through three layers of 3×3 and 1×1 convolutions to reduce the channels to 256 and 128, respectively, finally yielding a binary classification as the output.

[0091] Experiments and Results:

[0092] This invention was evaluated on three publicly available remote sensing change detection datasets, representing different change detection scenarios:

[0093] (1) LEVIR-CD dataset: A dataset for building change detection, containing 637 pairs of images with a spatial size of 1024×1024 pixels and a spatial resolution of 0.5 meters. The original images were obtained from 20 regions in Texas, USA, through the Google Earth platform. Each image was cropped into a non-overlapping image patch of 256×256 pixels. Building changes in LEVIR-CD show significant scale variations, ranging from single small houses to large building complexes.

[0094] (2) SYSU-CD dataset: Contains various types of changes, such as buildings, vegetation, roads, and marine structures, posing more challenges to change detection tasks. The dataset contains 20,000 pairs of 0.5-meter aerial images, with a size of 256×256 pixels, taken in Hong Kong from 2007 to 2014. SYSU-CD presents challenges such as complex urban scenes with changes in illumination, seasonality, and mixed types of changes.

[0095] (3) WHU-CD dataset: A dataset for building change detection, generated from a pair of aerial images with a size of 32507×15354 pixels and a spatial resolution of 0.7 meters. The original images were cropped into 256×256 pixel image patches. WHU-CD focuses on building construction and demolition, with the changed area accounting for about 5% of the image area.

[0096] To quantitatively evaluate the effectiveness of change detection methods, five widely used metrics were used to compare predicted change maps with ground truth annotations, including precision (Pre), recall (Rec), F1 score (F1), intersection over union (IoU), and overall accuracy (OA).

[0097] Training was performed using the AdamW optimizer with an initial learning rate of 10^-4 and a batch size of 8. The encoder residuals were initialized from ImageNet pre-trained weights (at layer structure matching), while the edge-aware center-difference convolutional module (input level) and task-related modules (multi-scale spatiotemporal transform module and adaptive region-aware progressive decoder) were randomly initialized using the He initialization method. Standard data augmentation techniques were employed, including random horizontal / vertical flips (probability 0.5), random rotations at multiples of 90° (probability 0.5), and random cropping (scale 0.8–1.0) to improve generalization. Experiments were conducted using a single NVIDIA RTX 4090 GPU within the PyTorch framework.

[0098] The experimental results of the proposed EMDNet on three datasets are as follows:

[0099] (1) LEVIR-CD dataset: EMDNet achieved an F1 score of 91.09% and an Intersection over Union (IoU) of 83.63%, with a precision of 92.34% and a recall of 89.87%. The high IoU validates the effectiveness of the edge-preserving design of the edge-aware center differential feature encoder, the high precision reflects the role of the multi-scale spatiotemporal transformer module in suppressing spurious changes, and the strong recall confirms the advantage of the adaptive region-aware progressive decoder in small target detection. Qualitative results of change detection on the LEVIR-CD dataset are as follows: Figure 6 As shown. Figure 6 The meaning of each row of data is as follows: (a) row contains 6 sets of T1-stage images, (b) row contains 6 sets of T2-stage images, (c) row contains 6 sets of ground truth changes (manually annotated areas of change; white indicates change, black indicates no change), and (d) row contains 6 sets of change detection results using the method of this invention. For ease of comparison, (d) is overlaid with a difference marker from the ground truth: green indicates a missed detection (ground truth changed but prediction remained unchanged), and red indicates a false positive (ground truth remained unchanged but prediction changed). Figure 6 As can be seen in (d), there are fewer green and red areas, indicating that the present invention can effectively detect real changes and suppress false alarms on LEVIR-CD, which is consistent with the effect reflected by quantitative indicators.

[0100] (2) SYSU-CD dataset: EMDNet achieved an F1 score of 83.01% and an Intersection over Union (IoU) of 70.91%, with a precision of 85.77% and a recall of 80.42%. In complex scenes containing changes in illumination and seasonality, EMDNet demonstrated strong robustness to pseudo-changes, with both precision and recall reaching high levels.

[0101] (3) WHU-CD dataset: EMDNet achieved an F1 score of 90.70% and an Intersection over Union (IoU) of 82.83%, with a precision of 89.22% and a recall of 92.24%. Even in extreme class imbalance where the variable region accounts for only about 5%, EMDNet still maintains good detection performance. The high recall rate particularly demonstrates the effectiveness of the adaptive region-aware progressive decoder in handling class imbalance.

[0102] Overall experimental results show that the proposed EMDNet achieves high F1 scores on all three datasets, validating the effectiveness of edge-preserving encoding, spectral-temporal difference modeling, and region-aware fusion. Qualitative results demonstrate robust performance under challenging scenarios such as illumination variations, seasonal changes, and small buildings.

[0103] This invention also provides a remote sensing image change detection system based on edge-aware multi-scale difference, comprising:

[0104] At least one processor;

[0105] At least one memory for storing at least one program;

[0106] When the at least one program is executed by the at least one processor, the at least one processor performs the method described above.

[0107] The content of the above method embodiments is applicable to this embodiment. The specific functions implemented in this embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments. Therefore, they will not be repeated here.

[0108] This invention also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0109] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0110] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0111] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0112] This invention also provides a computer program product, including a computer program or computer instructions, which are stored in a memory. A processor of a computer device reads the computer program or computer instructions from the memory and executes the computer program or computer instructions, causing the computer device to perform the above-described method.

[0113] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0114] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0115] It will be understood by those skilled in the art that all or some of the steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically include computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0116] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

Claims

1. A method for detecting changes in remote sensing images based on edge-aware multi-scale difference, characterized in that, The method includes the following steps: S100, acquire a pair of dual-temporal remote sensing images and a trained edge-aware multi-scale difference network, wherein the edge-aware multi-scale difference network includes an edge-aware center difference feature encoder, a multi-scale spatiotemporal transformer module and an adaptive region-aware progressive decoder. S200, the dual-temporal remote sensing images are respectively input into the edge-aware center difference feature encoder. High-frequency edge information is retained in the input stage through three-level cascaded center difference convolution, and multi-scale edge features are extracted through pooling layers and residual blocks. S300, the multi-scale edge features are input into the multi-scale spatiotemporal transformer module, and joint spectral-spatial-temporal inference is performed through multi-scale spectral pooling, dual-temporal self-attention and adaptive feature fusion to generate differential enhancement features; S400, the differential enhancement features are input into the adaptive region-aware progressive decoder. The features are divided into change paths, background paths and global paths through region-aware attention and processed separately. Progressive fusion from coarse to fine is performed, and multi-level supervision is applied to output the change detection results.

2. The method according to claim 1, characterized in that, In S200, the process of preserving high-frequency edge information at the input stage through a three-level cascaded central difference convolution and extracting multi-scale edge features via pooling layers and residual blocks includes: S210, the two temporal remote sensing images in the dual-temporal remote sensing image are independently transformed by the first layer of central difference convolution, and the first edge feature maps of the two temporal phases are output accordingly. S220, the first edge feature map is transformed by the second layer of central difference convolution to output the second edge feature map; S230, the second edge feature map is transformed by the third layer of central difference convolution to output the edge-aware feature map; wherein, each layer of central difference convolution first performs a 3×3 convolution, and then subtracts the depth 1×1 low-frequency estimate constructed by weight statistics to form a learnable high-pass filter. S240, after the edge-aware feature map is processed by the pooling layer, it is sequentially passed through four residual stages to extract multi-scale edge features at four resolutions. The four resolutions are H / 2×W / 2, H / 4×W / 4, H / 8×W / 8, and H / 16×W / 16, with corresponding channel numbers of 64, 128, 256, and 512, respectively, where H and W are the height and width of the input image.

3. The method according to claim 1, characterized in that, In S300, the process of inputting the multi-scale edge features into the multi-scale spatiotemporal transformer module, and performing joint spectral-spatial-temporal inference through multi-scale spectral pooling, dual-temporal self-attention, and adaptive feature fusion to generate differentially enhanced features includes: S310 processes the dual-temporal features contained in the multi-scale edge features of each scale through three parallel paths, and outputs the preliminary fusion features, temporal enhancement features and single-temporal refinement features respectively. S320, after concatenating the preliminary fusion feature with the single-phase refined feature, one path is generated by pooling and Softmax normalization to produce channel adaptive weights, and the other path is generated by Sigmoid activation to produce spatial adaptive weights. The channel adaptive weights and the spatial adaptive weights are multiplied to obtain the adaptive weights. S330, Multiply the adaptive weights element-wise with the preliminary fusion features to obtain the weighted concatenation features; S340, the temporal enhancement feature and the weighted concatenation feature are added element by element to obtain the differential enhancement feature.

4. The method according to claim 3, characterized in that, In S310, the multi-scale edge features at each scale are processed through three parallel paths, which output preliminary fusion features, temporal enhancement features, and single-temporal refined features, including: The upper path concatenates the two temporal features along the channel dimension and then processes them through five layers of 1×1 convolution to output preliminary fused features; the two temporal features are the corresponding scale feature pairs extracted from the two temporal remote sensing images by the edge-aware center differential feature encoder, including the first temporal feature and the second temporal feature; The intermediate path calculates the difference features of the dual-temporal features, which are then passed through the multi-scale spectral-spatial attention module to output spectral-spatial enhancement features, and then through the temporal context modeling module to obtain temporal enhancement features; the difference features are the absolute value difference between the first temporal feature and the second temporal feature by subtracting each element from the first temporal feature; The lower path performs a five-layer 1×1 convolution feature extraction on the first phase feature in the dual-temporal features, and outputs a single-temporal refined feature; the single-temporal feature is either the first phase feature or the second phase feature in the dual-temporal remote sensing image.

5. The method according to claim 4, characterized in that, The spectral-spatial enhancement features output sequentially through the multi-scale spectral-spatial attention module include: S311, the differential features are input into the spectral branch, and adaptive average pooling with three windows (1×1, 3×3, and 5×5) is used to obtain multi-scale spectral information. The 1×1 pooling result is then averaged along the width and height and concatenated along the spatial dimension, followed by 1×1 convolution dimensionality reduction and Sigmoid activation to obtain channel attention. The 3×3 and 5×5 pooling results are added and upsampled back to the original resolution to obtain multi-scale contextual information. S312, the differential features are input into the spatial branch, projected into a query vector, a key vector and a value vector by three 1×1 convolutions, and obtained by scaling dot product attention to obtain spatial weighted features; S313, the channel attention and spatial weighted features are fused to output spectral-spatial enhancement features.

6. The method according to claim 5, characterized in that, The temporal enhancement features obtained by the temporal context modeling module include: S314, the spectral-spatial enhancement features are flattened into a sequence and then projected into a query vector, a key vector, and a value vector; S315 establishes the correspondence between the two temporal feature states by scaling the dot product attention, and obtains the attention output; S316, the attention output is processed through a linear layer, residual connections and a feedforward network to obtain temporal enhancement features.

7. The method according to claim 1, characterized in that, In S400, the step of dividing features into change paths, background paths, and global paths through region-aware attention and processing them separately includes: S410 reduces the dimensionality of multi-scale difference features to 128 dimensions through 3×3 convolution, batch normalization, and ReLU activation. S420 inputs the normalized features into the adaptive region attention module, which processes them through four parallel paths: Path 1 is a 1×1 convolution, a depthwise separable convolution, and a 1×1 convolution; Paths 2 and 3 each generate a gated mask; Path 4 is a 1×1 convolution, a depthwise separable convolution, a 1×1 convolution, and a 1×1 convolution. The S430 combines the four outputs element-wise through multiplication and addition, then performs 3×3 convolution, ReLU, 1×1 convolution, and batch normalization to obtain refined features for region awareness. It applies boundary focusing attention to changing regions and performs noise suppression on background regions.

8. The method according to claim 7, characterized in that, In S400, the progressive fusion from coarse to fine, and the application of multi-level supervision, to output change detection results, includes: S440 starts from the coarsest scale, normalizes the differential features, and then obtains refined features through adaptive region attention; S450 directly upsamples the refined features of the deepest scale; for shallower scales, it upsamples the refined features of the previous scale and concatenates them with the refined features of the current scale, then performs 3×3 convolution, batch normalization, ReLU activation, and finally adds them to the refined features of the current scale according to learnable weights to obtain the refined output of the current scale. S460 adds auxiliary prediction heads at each scale for multi-level supervision, and finally upsamples all refined features at all scales to the original resolution and stitches them together. After convolutional dimensionality reduction, it outputs a binary classification result.

9. A remote sensing image change detection system based on edge-aware multi-scale difference, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor performs the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 8.

Images