Remote sensing image change understanding and detection method based on global perception convolution and double-phase interaction

Abstract

The application discloses a remote sensing image change understanding and detection method based on global perception convolution and double-time interaction, utilizes an interframe cross-scale feature aggregation module to uniformly model different scale features in the same time point, fuses high-level semantic information and low-level spatial details, and obtains two time-scale consistent feature pyramids; and utilizes an interframe cross-time feature interaction module to establish a feature interaction relationship between two time nodes, and obtains four-level time fusion features through symmetric aggregation. The application effectively suppresses false changes, and significantly improves the precision and efficiency balance of remote sensing change detection.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image change detection technology, specifically a method for understanding and detecting changes in remote sensing images based on global perception convolution and dual temporal interaction. Background Technology

[0002] With the rapid development of remote sensing technology and the widespread application of satellite and unmanned aerial vehicle (UAV) platforms, Remote Sensing Change Detection (RSCD) has become one of the core technologies in land resource monitoring, urban planning management, and disaster emergency response, attracting increasing attention. RSCD, through pixel-level comparative analysis of multi-temporal remote sensing images, can accurately identify changes in land cover, providing crucial data support for applications such as dynamic land use monitoring, illegal construction investigation, forest cover change assessment, and natural disaster damage evaluation. Therefore, in recent years, efficient and accurate remote sensing change detection methods have gradually become a research hotspot.

[0003] With the development of remote sensing change detection technology, research focus has shifted from early manual feature methods to end-to-end learning based on CNNs and visual Transformers. However, in practical deployments, existing methods face a trade-off between global context modeling and computational efficiency. While CNN methods are efficient, they are limited by local receptive fields and struggle to suppress spurious changes caused by illumination and seasonal variations. Although visual Transformers can model long-range dependencies, their quadratic complexity is insufficient for real-time processing. Therefore, achieving efficient global context modeling within a convolutional framework has become a key challenge in improving the practicality of remote sensing change detection.

[0004] To address these issues, most existing methods attempt to fuse convolutional and attention mechanisms through hybrid architectures, but limitations remain in feature fusion efficiency and global routing mechanisms. Some methods employ large-kernel convolution or dynamic filtering to expand the receptive field, but neglect the characteristics of varying land cover scales and complex change patterns in remote sensing imagery. Other methods compensate for the global modeling deficiencies of state-space models through spatiotemporal relationship modules, but designing these modules significantly increases the number of parameters and computational overhead. These limitations are mainly reflected in two aspects: First, existing linear complexity models compress spatial features into sequential representations, leading to indirect image-level global inference and loss of spatial structural information required for dense prediction tasks; second, in the process of cross-temporal feature fusion, simple stitching or shallow interactions are insufficient to effectively distinguish between real structural changes and pseudo-changes, ultimately affecting the accuracy and robustness of change detection. Summary of the Invention

[0005] The purpose of this invention is to provide a remote sensing image change detection method based on global perceptual convolution, which solves the problems of high computational complexity and difficulty in processing high-resolution remote sensing images in existing change detection methods based on self-attention mechanisms. This invention achieves global context modeling while maintaining linear computational complexity, thereby improving the accuracy and efficiency of change detection.

[0006] The technical solution for achieving the objective of this invention is: a method for understanding and detecting changes in remote sensing images based on global perceptual convolution and dual temporal interaction, comprising:

[0007] The dual-temporal remote sensing images are input into the corresponding dual-temporal initial features of the encoder. The encoder is used to encode the dual-temporal initial features layer by layer to obtain the multi-scale feature pyramids of the two temporal phases.

[0008] By using the intra-frame cross-scale feature aggregation module, features of different scales are uniformly modeled within the same time point, enabling the fusion of high-level semantic information and low-level spatial details to obtain two temporally scale-consistent feature pyramids.

[0009] The feature interaction relationship between two time nodes is established by using the inter-frame cross-temporal feature interaction module, and four-level temporal fusion features are obtained through symmetrical aggregation;

[0010] The four-level temporal fusion features are aggregated and upsampled by the decoder to generate a change probability map. The change probability map is then thresholded to obtain a binary change mask.

[0011] The remote sensing image detection model, consisting of an encoder, an intra-frame cross-scale feature aggregation module, an inter-frame cross-temporal feature interaction module, and a decoder, is trained. The detection results are obtained by inputting the corresponding dual-temporal initial features of the real-time dual-temporal remote sensing images into the remote sensing image detection model.

[0012] Compared with the prior art, the significant advantages of this invention are:

[0013] This invention achieves global routing of convolution by dynamically generating globally perceptive convolution kernels, while maintaining linear computational complexity, enabling local feature aggregation to integrate overall contextual information. Simultaneously, it introduces a dual-temporal fusion strategy based on symmetric modulation, performing channel-level and spatial-level mutual conditionalization on features from dual-temporal images. This dynamically enhances the response to real structural changes and suppresses spurious changes caused by illumination, seasonality, and radiation variations, effectively alleviating the problems of limited receptive field in traditional convolution methods and high computational overhead in attention methods. This results in efficient and accurate change detection and robust fusion.

[0014] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0015] Figure 1The flowchart is shown below for a remote sensing image change understanding and detection method based on global perception convolution and dual temporal interaction.

[0016] Figure 2 This is a schematic diagram of the global perception convolution principle in Embodiment 1 of the present invention.

[0017] Figure 3 This is a schematic diagram of the inter-frame cross-temporal feature interaction principle in Embodiment 1 of the present invention. Detailed Implementation

[0018] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0019] A remote sensing image change understanding and detection method based on global-aware convolution and dual-temporal interaction aims to achieve efficient global context modeling through a convolutional global routing mechanism and explicitly model cross-scale and cross-temporal interactions through a dual-temporal fusion module, effectively suppressing spurious changes and significantly improving the balance between accuracy and efficiency in remote sensing change detection. The specific steps are as follows:

[0020] Step 1: Hierarchical feature extraction based on global awareness convolution: By dynamically generating global awareness convolution kernels, local convolutions can integrate global contextual information when aggregating neighborhood features, thereby achieving global semantic modeling while maintaining linear computational complexity, and avoiding the secondary complexity problem caused by explicit inter-element interactions in the self-attention mechanism.

[0021] In practice, the first step is to combine the two-temporal remote sensing images. Input the initial downsampling module to obtain dual-time initial features The dual-temporal remote sensing image pair includes a pre-event remote sensing image T1 and a post-event remote sensing image T2.

[0022] Using the first global perception convolutional module as the basic building block, a four-level hierarchical backbone network is constructed as the encoder. Each level consists of the first global perception convolutional module. The encoder encodes the initial features of the two temporal phases layer by layer, resulting in a multi-scale feature pyramid for the two temporal phases. and{ ;

[0023] At each level of the encoder, the first global-aware convolutional module implements adaptive convolution through a dynamic routing mechanism: firstly, through the projection matrix... and global average pooling from input features Extract compact global descriptors The descriptor encodes scene-level statistics; these are then processed by a lightweight generator. Activation with sigmoid Convert descriptors to routing coefficients , used for a set of learnable convolutional kernel bases We perform weighted combinations to construct a dynamic convolutional kernel K that is relevant to the input. The specific calculation process is as follows:

[0024]

[0025]

[0026]

[0027] This globally aware convolutional kernel can achieve global information propagation with linear complexity when performing local aggregation. When the convolutional kernel is at a pixel location... neighborhood When performing feature aggregation within a neighborhood, the features of the neighboring pixel q The output features of this layer are obtained by weighted summation based on the dynamic convolution kernel. .

[0028]

[0029] In this way, each local convolution not only utilizes local spatial information when performing neighborhood feature aggregation, but also perceives the global semantic structure of the entire feature map. This allows for global convolutional routing without explicit inter-token attention, while maintaining the linear computational efficiency of standard convolutions.

[0030] Step 2: Intra-frame cross-scale feature aggregation: The intra-frame cross-scale feature aggregation module is used to uniformly model features of different scales within the same time point, so that high-level semantic information and low-level spatial details can be fully integrated, thereby improving the expressive ability of changing regions.

[0031] Specifically, the four-level feature pyramid receives two temporal phases from the encoder output. ,in Represents a timestamp. Indicates the scale level. For each time phase... First, feature maps at all scales are bilinearly upsampled to the highest spatial resolution, then concatenated along the channel dimension and fused using a second global awareness convolutional module. The specific design process is as follows:

[0032]

[0033]

[0034] in, This represents a globally perceptive convolution. Indicates bilinear upsampling. This indicates concatenation along the channel dimension. This indicates scale-specific downsampling. Fusion features. After downsampling and projecting back to each scale, a scale-consistent feature pyramid is obtained. This process enables cross-resolution information exchange, preserving spatial details while enhancing semantic consistency.

[0035] Step 3: Inter-frame cross-temporal feature interaction: The inter-frame cross-temporal feature interaction module is used to establish feature interaction relationships between two time nodes, so that the features of each time node can use the information of the other time node, thereby strengthening the change area and suppressing pseudo-changes.

[0036] To effectively model cross-temporal semantic relationships, this invention receives two temporally scale-consistent feature pyramids output from intra-frame cross-scale feature aggregation. and{ For each scale j, let = , = As input feature pairs at this scale, this invention employs a symmetric interaction mechanism to conditionally pair the features from the two periods.

[0037] Firstly, at the channel dimension, the channel modulator extracts statistical descriptors for features at each scale through global average pooling and max pooling, and then uses a lightweight bottleneck network. Generate channel weights. These weights are used to modulate the feature response at another time point, thereby highlighting regions of real change and suppressing spurious changes.

[0038] In the spatial dimension, the spatial modulator performs channel aggregation on the two temporal-scale consistent feature pyramids at this scale to generate a spatial description map, and then uses a lightweight spatial gating function. Spatial modulation weights are generated. These weights are used to emphasize the spatial response of varying regions. Fusion features are obtained through symmetric aggregation. The specific design of cross-temporal feature interaction is as follows:

[0039] , ,

[0040] , ,

[0041]

[0042] in, Indicates channel modulation weights; Indicates spatial modulation weights; The characteristics after channel modulation, This represents the feature representation after spatial modulation. This indicates element-wise multiplication.

[0043] Step 4: Use a lightweight decoder to process the four-level temporal fusion features output by the inter-frame cross-temporal feature interaction module. Aggregate upsampling to generate a probability map of change. Then, analyze this probability map... The binary transformation mask can be obtained by thresholding. .

[0044] Step 5: Train the remote sensing image detection model consisting of the encoder, intra-frame cross-scale feature aggregation module, inter-frame cross-temporal feature interaction module, and decoder. Input the corresponding dual-temporal initial features of the real-time acquired dual-temporal remote sensing images into the remote sensing image detection model to obtain the detection results.

[0045] When training the remote sensing image detection model, a joint optimization of the loss function is employed: through pixel-level supervision and region-level overlap constraints, the change detection accuracy is improved, and the ability to identify small target change regions is enhanced. The binary change mask is then compared with the ground truth label. Optimize using a common input loss function.

[0046] To enhance the model's ability to identify changing regions, this invention employs a joint optimization strategy combining binary cross-entropy (BCE) loss and Dice similarity loss. BCE loss constrains the consistency between the predicted results and the true labels through pixel-level supervision, while Dice loss measures the degree of matching between the predicted region and the actual changing region through region overlap. By jointly optimizing these two loss functions, pixel-level accuracy and region-level detection performance can be simultaneously guaranteed, effectively improving the model's ability to detect changing regions of small targets.

[0047] Specifically, based on binary transformation masks and real labels The BCE item is defined as follows:

[0048]

[0049] Further employing the soft Dice coefficient and its corresponding loss:

[0050]

[0051]

[0052] The ultimate training objective is:

[0053]

[0054] in, Represents the binary cross-entropy loss. This indicates the loss of similarity between Des and S. This represents the loss weighting coefficient (default value is 1). H is a smoothing factor, and H and W represent the image height and width, respectively.

[0055] This joint loss function simultaneously constrains pixel-level classification accuracy and region-level overlap quality. By optimizing the entire network parameters through end-to-end backpropagation, the change probability map output by this method can be thresholded to obtain an accurate binary change mask, thus realizing the complete process of remote sensing image change detection.

[0056] In terms of global semantic modeling capabilities, this invention introduces a global-aware convolution mechanism, which dynamically modulates the convolution kernel using global semantic information during the convolution calculation process. This enables the aggregation of local features to integrate the overall contextual information, thereby improving the expressive ability of remote sensing image change features.

[0057] In terms of change discrimination capability, this invention constructs an intra-frame cross-scale feature aggregation method and an inter-frame cross-temporal feature interaction method, enabling features of different scales and different temporal phases to establish stable semantic associations, effectively highlighting real structural changes and suppressing pseudo-change interference caused by illumination changes and seasonal changes.

[0058] In terms of computational efficiency, this invention uses a global awareness modeling method based on convolutional structures to achieve global context modeling while avoiding the secondary complexity problem of self-attention mechanisms, thereby reducing computational overhead and improving the application efficiency of the method on resource-constrained platforms.

[0059] Figure 1 This paper presents an overall framework diagram of a remote sensing image change understanding and detection method based on globally perceptive convolution and dual-temporal interaction. The framework consists of three main components: a hierarchical encoder based on globally perceptive convolution, a dual-temporal fusion module, and a lightweight decoder. The hierarchical encoder encodes the dual-temporal remote sensing images into four-level feature pyramids; the dual-temporal fusion module sequentially performs intra-frame cross-scale aggregation and inter-frame cross-temporal interaction to achieve multi-scale semantic enhancement and highlight true changes; the lightweight decoder upsamples the fused multi-scale features level by level to generate the final change probability map.

[0060] Figure 2 This section describes the design of an intra-frame cross-scale feature aggregation method. For each timestamp, a four-level feature pyramid is first upsampled to the highest spatial resolution, then concatenated along the channel dimension and fused using a globally perceptive convolutional block to generate fused features. These fused features are then projected back to their original scales to obtain a scale-consistent refined representation. This process enables cross-resolution information exchange, preserving spatial details while enhancing semantic consistency, and effectively solving the problem of semantic inconsistency in single-scale features.

[0061] Figure 3This section describes a method for cross-temporal feature interaction between frames. Given scale-aligned bi-temporal features, channel descriptors are first calculated using global average pooling and max pooling to generate cross-temporal channel gating coefficients. The features themselves are then modulated using the weights of the other channel, achieving channel-level mutual conditionalization. Subsequently, spatial descriptors are generated through channel-level pooling and convolutional prediction. The features are further modulated using the spatial attention map of the other channel, achieving spatial-level mutual conditionalization. Finally, the symmetrically modulated bi-temporal features are added element-wise to obtain temporally aware fusion features. This symmetrical interaction scheme allows bi-temporal features to mutually remind attention to important channels and changing locations, effectively highlighting real structural changes and suppressing spurious change interference.

Claims

1. A method for understanding and detecting changes in remote sensing images based on global perception convolution and dual temporal interaction, characterized in that, include: The dual-temporal remote sensing image pairs are input into the encoder corresponding to the dual-temporal initial features. The encoder encodes the dual-temporal initial features layer by layer to obtain the multi-scale feature pyramids of the two temporal phases. The encoder adopts a four-level hierarchical backbone network. Each level of the network adopts a first global perception convolutional module. The first global perception convolutional module achieves adaptive convolution through a dynamic routing mechanism. By utilizing an intra-frame cross-scale feature aggregation module to uniformly model features at different scales within the same time point, high-level semantic information is fused with low-level spatial details, resulting in two temporally scale-consistent feature pyramids. The specific method is as follows: A four-level feature pyramid receiving two temporal phases from the encoder output. ,in Represents a timestamp. Indicates scale level; For each time phase The feature maps of all scales are bilinearly upsampled to the highest spatial resolution, and then stitched together along the channel dimension before being fused by the second global perception convolutional module. An inter-frame cross-temporal feature interaction module is used to establish feature interaction relationships between two time nodes. Four-level temporal fusion features are obtained through symmetrical aggregation. Specifically, at the channel dimension, statistical descriptors of features at each scale of the feature pyramid with consistent temporal scales are extracted using global average pooling and max pooling. A lightweight bottleneck network is then used to further refine these features. Generate channel weights; In the spatial dimension, channel aggregation is performed on the two temporally scaled consistent feature pyramids to generate spatial description maps, and a lightweight spatial gating function is used to achieve this. Generate spatial modulation weights; Fusion features are obtained through symmetric aggregation. ; The four-level temporal fusion features are aggregated and upsampled by the decoder to generate a change probability map. The change probability map is then thresholded to obtain a binary change mask. The remote sensing image detection model, consisting of an encoder, an intra-frame cross-scale feature aggregation module, an inter-frame cross-temporal feature interaction module, and a decoder, is trained. The detection results are obtained by inputting the corresponding dual-temporal initial features of the real-time dual-temporal remote sensing images into the remote sensing image detection model.

2. The remote sensing image change understanding and detection method based on global perception convolution and dual temporal interaction according to claim 1, characterized in that, The specific process by which the first globally aware convolutional module implements adaptive convolution through a dynamic routing mechanism is as follows: Through projection matrix and global average pooling from input features Extract compact global descriptors ; Through lightweight generator Activation with sigmoid Convert global descriptors to routing parameters ; The routing coefficient For a set of learnable convolutional kernel bases We perform weighted combinations to construct a dynamic convolutional kernel K that is relevant to the input; When the convolution kernel is at pixel position neighborhood When performing feature aggregation, the features V(q) of the neighboring pixel q will be weighted and summed according to the dynamic convolution kernel to obtain the output feature Y(p) of the corresponding level.

3. The remote sensing image change understanding and detection method based on global perception convolution and dual temporal interaction according to claim 1, characterized in that, Fusion features Specifically: ; , , ； ， , ； In the formula, This is a feature pyramid feature map consistent across two temporal scales. and These are the global average pooling and global max pooling functions, respectively. For channel modulation functions based on global perception convolution, Indicates channel modulation weights; and These are the channel averaging and channel max pooling functions, respectively. For concatenation functions, For spatial modulation functions based on global perception convolution, Indicates spatial modulation weights; The characteristics after channel modulation, This represents the feature representation after spatial modulation. This represents element-wise multiplication. Represents a timestamp.

4. The remote sensing image change understanding and detection method based on global perception convolution and dual temporal interaction according to claim 1, characterized in that, When training the remote sensing image detection model, a joint optimization strategy of binary cross-entropy loss and Dessian similarity loss is adopted. The binary cross-entropy loss constrains the consistency between the prediction results and the real labels through pixel-level supervision, while the Dessian similarity loss measures the degree of matching between the predicted region and the real changed region through the region overlap.

5. The remote sensing image change understanding and detection method based on global perception convolution and dual temporal interaction according to claim 4, characterized in that, Based on the binary transformation mask and the true label, the specific binary cross-entropy loss is determined as follows: ； In the formula, Authentic labels The binary transformation mask predicted by the detection model, where H and W represent the image height and width, respectively. As a smoothing factor, , The real label and the binary transformation mask are located at the following positions: The value at that location.

6. The remote sensing image change understanding and detection method based on global perception convolution and dual temporal interaction according to claim 4, characterized in that, The specific Des similarity loss is as follows: ； In the formula, Authentic labels The binary change mask predicted by the detection model, denoted as the Dessian similarity coefficient.

7. The remote sensing image change understanding and detection method based on global perception convolution and dual temporal interaction according to claim 6, characterized in that, The Des similarity coefficient is as follows: ； As a smoothing factor, , The real label and the predicted mask are located at the following positions. The value at that location.

Images