A Remote Sensing Method and System for Glacier Change Detection Based on Local Detail Feature Aggregation

Abstract

This invention relates to the field of remote sensing image change detection technology, and particularly to a remote sensing method and system for glacier change detection based on local detail feature aggregation. By effectively aggregating and optimizing local detail information, this method significantly improves the accuracy of glacier change detection under complex terrain and variable environmental conditions, and enhances the ability to capture and identify the fine structures at glacier edges. The system includes: a multi-scale feature extraction module for extracting multi-scale features from high-resolution dual-temporal glacier images; a local detail aggregation module for aggregating local glacier detail information to detect edges and transition regions in glacier images; a shape perception module for flexibly capturing the complex structure and contours of glaciers; and a multi-scale difference prediction module for simultaneously processing large-scale and small-scale changes, providing more comprehensive detection results. Compared with existing technologies, this invention can accurately detect the complex shapes and boundary information of glaciers and accurately identify minute change regions located at glacier edges.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image change detection technology, and in particular to a remote sensing method and system for glacier change detection based on local detail feature aggregation. Background Technology

[0002] Glaciers are not only important indicators of global climate but also the largest freshwater reservoir on Earth. Accurate monitoring of the scale and distribution of glacier changes is crucial for coastline stability and the protection of natural ecosystems. Glaciers are mostly located in cold, high-altitude regions with harsh climates, making large-scale field surveys of glacier changes extremely difficult. Remote sensing imagery, due to its increasingly abundant data, high availability, and wide coverage, has become an indispensable tool for studying glacier changes. However, current automated glacier change detection based on deep learning still faces some challenges:

[0003] (1) Edge changes are difficult to detect. The terrain at the edge of a glacier is complex, with cliffs, slopes and other complex features, making the edge more susceptible to changes in terrain and gravity, thus increasing the likelihood of melting and collapse. At the same time, the edge of a glacier is often disturbed by shadows, snow cover and other environmental factors, affecting the accurate detection of edge changes and increasing the difficulty of conducting research and monitoring in the edge area.

[0004] (2) Complex geometry. Glacier edges are complex and varied, often jagged or wavy, containing multiple tributary glaciers, forming a complex edge structure. Their surfaces are covered with cracks, ice towers, and ice caves, exhibiting irregular surface features due to factors such as ice flow velocity and temperature changes. Furthermore, glaciers are constantly moving, with significant differences in their speed across different regions and time points. Melting and accumulation processes exhibit seasonal and interannual variations, leading to continuous changes in glacier shape. Glaciers are typically located in valleys or plateaus with complex topography, significantly impacting their shape. Current mainstream change detection methods struggle to detect irregularly shaped targets. Summary of the Invention

[0005] In view of this, the purpose of this invention is to propose a remote sensing detection method and system for glacier changes based on local detail feature aggregation, so as to solve the problem that existing detection networks have poor detection performance in edge regions.

[0006] To achieve the above objectives, this invention provides a remote sensing method for glacier change detection based on local detail feature aggregation, comprising the following steps:

[0007] S1. Extract multi-scale feature maps from dual-temporal glacier images;

[0008] S2. Enhance the features of the edge parts of the multi-scale feature map to obtain the enhanced multi-scale feature map.

[0009] S3. Perform shape perception on the enhanced multi-scale feature map of complex glaciers to obtain an enhanced dual-temporal glacier feature map.

[0010] S4. The enhanced dual-temporal glacier feature map is used to predict multi-scale glacier changes through multi-scale difference prediction, and the final binary map of changes is obtained through a classifier to obtain the detection results.

[0011] Preferably, step S1 further includes:

[0012] S1.1 Input the first-phase glacier image into the backbone network to obtain multi-scale feature maps F1 1 , The sizes of the multi-scale feature maps are 2, 4, 8, and 16 times the resolution downsampled from the original feature map, respectively.

[0013] S1.2 Input the second-phase glacier image into the backbone network to extract multi-scale feature maps F1 2 F2 2 F3 2 , The sizes of the multi-scale feature maps are 2, 4, 8, and 16 times the resolution of the original feature map, respectively.

[0014] Preferably, step S2 further includes:

[0015] S2.1, Combine encoder feature maps of the same scale. With decoder feature map Feature maps P1 and P2 are obtained by passing them through the channel attention module respectively;

[0016] S2.2, Connect feature maps P1 and P2 with... Perform Hadamard product, and the resulting feature map is compared with... Adding them together yields feature maps P3 and P4;

[0017] S2.3. Pass feature maps P3 and P4 through the spatial attention module and multiply them element-wise with the original feature maps to obtain feature maps M1 and M2.

[0018] S2.4. Add feature maps M1 and M2 to feature maps P3 and P4 element by element to obtain feature maps M3 and M4 respectively.

[0019] S2.5. Perform average pooling on feature map M3 and max pooling on feature map M4. Then, perform element-wise subtraction on the two pooled feature maps to obtain feature map C1.

[0020] Preferably, step S3 further includes:

[0021] S3.1 The input feature map C1 is first processed by a 3×3 convolution to obtain feature map B1;

[0022] S3.2, Feature map B1 is processed by a deformable large convolution kernel and then added to the original feature map to obtain feature map B2;

[0023] S3.3. Feature map B2 is multiplied element-wise by feature map C1 after a 3×3 convolution to obtain feature map B3.

[0024] Preferably, step S4 further includes:

[0025] S4.1. Obtain dual-temporal, multi-scale glacier prediction maps at different stages of the decoder.

[0026] S4.2 Subtract each element from the two-temporal multi-scale glacier feature maps to obtain multi-scale difference maps D1, D2, D3, and D4;

[0027] S4.3 Obtain a multi-scale binary map of glacier changes using a classifier and calculate the loss for backpropagation.

[0028] The present invention also provides a remote sensing detection system for glacier changes based on local detail feature aggregation, the system comprising:

[0029] Multi-scale feature extraction module: used to extract multi-scale features of dual-temporal glacier images using existing backbone network models;

[0030] Local detail aggregation module: used to aggregate local detail information of glacier edges and transition areas;

[0031] Shape-aware module: used to flexibly capture the complex structure and contours of glaciers;

[0032] Multi-scale difference prediction module: used to process large-scale and small-scale changes simultaneously and fuse information from different scales.

[0033] Preferably, the multi-scale feature extraction module includes:

[0034] The first feature extraction unit is used to input the first temporal glacier image into the backbone network to obtain multi-scale feature maps with gradually downsampled resolution of 2x, 4x, 8x and 16x.

[0035] The second feature extraction unit is used to input the second temporal glacier image into the backbone network to obtain a multi-scale feature map corresponding to the first temporal image, so that the features of the two temporal phases can be matched one-to-one at each scale.

[0036] Preferably, the local detail aggregation module includes:

[0037] The channel attention submodule is used to apply channel attention operations to the encoder feature map and decoder feature map of the same scale respectively to generate channel-enhanced features;

[0038] The feature fusion submodule is used to perform Hadamard product and element-wise addition operations on the feature maps enhanced by channel attention, and to further enhance the features using spatial attention to highlight the subtle changes at the glacier edge and transition zone.

[0039] The pooling and differential enhancement submodule is used to perform average pooling and max pooling on the enhanced features, and to perform element-wise subtraction and convolution on the pooling results, thereby further highlighting and synthesizing local detail features to generate the final local detail aggregated feature map.

[0040] Preferably, the shape-aware module includes:

[0041] The initial convolutional unit is used to perform convolution processing on the input feature map to obtain the initial feature map;

[0042] Deformable large convolution kernel units are used to perform large-scale morphological adaptation of initial features using deformable convolution and add them to the original feature map to capture the complex morphology of glaciers.

[0043] The feature selection unit is used to perform convolution and element-wise multiplication operations on the fused features, thereby improving the ability to perceive and characterize complex glacier contours and shape features.

[0044] Preferably, the multi-scale difference prediction module includes:

[0045] A multi-scale prediction unit is used to output dual-temporal multi-scale glacier prediction maps at different stages of the decoder;

[0046] The difference map generation unit is used to obtain a multi-scale difference map by subtracting elements from the dual-temporal multi-scale glacier feature map;

[0047] The classification and discrimination unit is used to make decisions on multi-scale difference features through a classifier, thereby outputting a binary detection result map of glacier changes.

[0048] The beneficial effects of this invention are:

[0049] This invention first extracts multi-scale features from dual-temporal glacier images to ensure rich texture and structural information at different spatial scales. Then, a local detail aggregation module applies channel and spatial attention processing to the features at matching scales to highlight edges and subtle changes, thereby enhancing the representational power of local features. Building upon this, a shape-aware module is introduced to perform deformable kernel convolution processing on the enhanced features, making the feature distribution more closely match the real glacier morphology and improving the accuracy of depicting the complex structure and contours of glaciers. Finally, the optimized dual-temporal features are fed into a multi-scale difference prediction module, which outputs a binary map of glacier changes through multi-scale fusion and a classifier, resulting in higher accuracy and robustness in glacier change detection tasks.

[0050] This invention combines a multi-scale feature extraction strategy with a local detail aggregation mechanism to achieve refined detection of glacier changes in complex environments. Channel and spatial attention modules are introduced to enhance local features at a fine-grained level, effectively highlighting glacier edges and subtle textures. A shape-aware module and deformable convolution operations are used to ensure that the feature spatial distribution closely matches the actual glacier morphology, thereby improving the accuracy of glacier shape modeling and change capture. Finally, multi-scale difference prediction and classifier decision-making achieve precise detection of glacier changes. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart of a remote sensing system for glacier change based on local detail aggregation, according to an embodiment of the present invention.

[0053] Figure 2 This is a diagram of a remote sensing network model for glacier change detection based on local detail aggregation, according to an embodiment of the present invention.

[0054] Figure 3 This is a schematic diagram of the composition structure of the local detail aggregation module in an embodiment of the present invention;

[0055] Figure 4 This is a schematic diagram of the shape sensing module structure according to an embodiment of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0057] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0058] like Figure 1 The diagram shown is a flowchart of a method according to an embodiment of the present invention, including the following processes: Dual-temporal glacier remote sensing images are input into a multi-scale feature extraction module to obtain a glacier multi-scale feature map; the glacier feature map is then sequentially input into a local detail aggregation module, a shape perception module, and a multi-scale difference prediction module to obtain the final predicted value. The predicted value and the true value are compared using a loss function, and backpropagation is used to optimize the network parameters to obtain a trained neural network model; test data is input into the trained neural network model to obtain a binary prediction image for glacier change detection. The network model diagram is shown below. Figure 2 As shown. The method specifically includes the following steps:

[0059] Step 1: Input the dual-temporal glacier images into the multi-scale feature extraction module to extract multi-scale features.

[0060] Step 1.1: Input the first-phase glacier image into the backbone network to obtain multi-scale feature maps. The sizes of the multi-scale feature maps are 2, 4, 8, and 16 times the resolution downsampled from the original feature map, respectively.

[0061] Step 1.2: Input the second-phase glacier image into the backbone network to extract multi-scale feature maps. The sizes of the multi-scale feature maps are 2, 4, 8, and 16 times the resolution downsampled from the original feature map, respectively.

[0062] Step 2: Input the multi-scale feature map into the local detail aggregation module to enhance the features of the edge regions and obtain the enhanced feature map. The aggregation module is as follows: Figure 3 As shown.

[0063] Step 2.1: Encoder feature maps with the same scale With decoder feature map Input the local detail aggregation module, and Feature maps P1 and P2 are obtained by passing through the channel attention module respectively.

[0064] Step 2.2: Feature maps P1, P2 and Perform Hadamard product, and the resulting feature map is compared with... Adding them together yields feature maps P3 and P4;

[0065] Step 2.3: Feature maps P3 and P4 are passed through the spatial attention module and multiplied element-wise with the original feature maps to obtain feature maps M1 and M2.

[0066] Step 2.4: Add feature maps M1 and M2 element by element to feature maps P3 and P4 respectively to obtain feature maps M3 and M4;

[0067] Step 2.5: Perform average pooling on feature map M3, max pooling on feature map M4, and then perform element-wise subtraction on the two pooled feature maps to obtain feature map C1;

[0068] Step 3: Input the enhanced multi-scale feature map into the shape perception module to perform shape perception of complex glaciers and obtain the enhanced dual-temporal glacier feature map. The shape perception module is as follows: Figure 4 As shown.

[0069] S3.1 The input feature map C1 is first processed by a 3×3 convolution to obtain feature map B1;

[0070] S3.2, Feature map B1 is processed by a deformable large convolution kernel and then added to the original feature map to obtain feature map B2;

[0071] S3.3. Feature map B2 is multiplied element-wise by feature map C1 after a 3×3 convolution to obtain feature map B3.

[0072] Step 4: Input the enhanced dual-temporal glacier feature map into the multi-scale difference prediction module to predict multi-scale glacier changes, and obtain the final binary map of changes through a classifier.

[0073] Step 4.1: Obtain dual-temporal, multi-scale glacier prediction maps at different stages of the decoder.

[0074] Step 4.2: Subtract each element from the two-temporal multi-scale glacier feature maps to obtain multi-scale difference maps D1, D2, D3, and D4;

[0075] Step 4.3: Obtain multi-scale binary maps P and P' of glacier change using a classifier.s2 P s4 P s8 And calculate the loss for backpropagation.

[0076] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in detail for the sake of brevity. Any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A remote sensing method for detecting glacier changes based on local detail feature aggregation, characterized in that, Includes the following steps: S1. Extract multi-scale feature maps from dual-temporal glacier images; S2. Enhance the features of the edge parts of the multi-scale feature map to obtain the enhanced multi-scale feature map. S3. Perform shape perception on the enhanced multi-scale feature map of complex glaciers to obtain an enhanced dual-temporal glacier feature map. S4. The enhanced dual-temporal glacier feature map is used to predict multi-scale glacier changes through multi-scale difference prediction, and the final binary map of changes is obtained through a classifier to obtain the detection results. Step S1 includes: S1.1 Input the first-phase glacier image into the backbone network to obtain multi-scale feature maps. , , , The sizes of the multi-scale feature maps are 2, 4, 8, and 16 times the resolution of the original feature map, respectively. S1.2 Input the second-phase glacier image into the backbone network to extract multi-scale feature maps. , , , The sizes of the multi-scale feature maps are 2, 4, 8, and 16 times the resolution of the original feature map, respectively. Step S2 includes: S2.1, Combine encoder feature maps of the same scale. With decoder feature map Feature maps are obtained by passing them through the channel attention module. , ; S2.2, Feature map , respectively with , Perform Hadamard product, and the feature maps after product are respectively compared with... , Add them together to obtain the feature map. , ; S2.3, Feature map , The feature maps are obtained by passing them through the spatial attention module and performing element-wise multiplication with feature maps P4 and P3 respectively. , ; S2.4, will , Feature maps and , The feature map is obtained by adding elements one by one. , ; S2.5, will The feature map is then subjected to average pooling. The feature maps are max-pooled, and then the two pooled feature maps are subtracted element-wise to obtain the final feature map. ; Step S3 includes: S3.1 Input Feature Map First, go through a Convolution yields feature maps. ; S3.2, Feature Map After applying a deformable large convolution kernel, the feature map is added to the original feature map to obtain the feature map. ; S3.3, Feature Map After a Convolution and feature map Element-wise multiplication yields the feature map. ; Step S4 includes: S4.

1. Obtain enhanced dual-temporal multi-scale glacier feature maps at different stages of the decoder. , , , , , , , ; S4.

2. Subtract the enhanced dual-temporal multi-scale glacier feature maps element-wise at the same scale to obtain a multi-scale difference map. , , , ; S4.3 Obtain a multi-scale binary map of glacier changes using a classifier and calculate the loss for backpropagation.

2. A remote sensing system for glacier change detection based on local detail feature aggregation, characterized in that, The system for performing the method as described in claim 1 includes: Multi-scale feature extraction module: used to extract multi-scale features of dual-temporal glacier images using existing backbone network models; Local detail aggregation module: used to aggregate local detail information of glacier edges and transition areas; Shape-aware module: used to flexibly capture the complex structure and contours of glaciers; Multi-scale difference prediction module: used to process large-scale and small-scale changes simultaneously and fuse information from different scales.

3. The glacier change remote sensing detection system based on local detail feature aggregation according to claim 2, characterized in that, The multi-scale feature extraction module includes: The first feature extraction unit is used to input the first phase glacier image into the backbone network to obtain multi-scale feature maps with gradually downsampled resolution of 2x, 4x, 8x and 16x. The second feature extraction unit is used to input the second temporal glacier image into the backbone network to obtain a multi-scale feature map corresponding to the first temporal image, so that the features of the two temporal phases can be matched one-to-one at each scale.

4. The glacier change remote sensing detection system based on local detail feature aggregation according to claim 2, characterized in that, The local detail aggregation module includes: The channel attention submodule is used to apply channel attention operations to the encoder feature map and decoder feature map of the same scale respectively to generate channel-enhanced features; The feature fusion submodule is used to perform Hadamard product and element-wise addition operations on the feature maps enhanced by channel attention, and to further enhance the features using spatial attention to highlight the subtle changes at the glacier edge and transition zone. The pooling and differential enhancement submodule is used to perform average pooling and max pooling on the enhanced features, and to perform element-wise subtraction and convolution on the pooling results, thereby further highlighting and synthesizing local detail features to generate the final local detail aggregated feature map.

5. The glacier change remote sensing detection system based on local detail feature aggregation according to claim 2, characterized in that, The shape sensing module includes: The initial convolutional unit is used to perform convolution processing on the input feature map to obtain the initial feature map; Deformable large convolution kernel units are used to perform large-scale morphological adaptation of initial features using deformable convolution and add them to the original feature map to capture the complex morphology of glaciers. The feature selection unit is used to perform convolution and element-wise multiplication operations on the fused features, thereby improving the ability to perceive and characterize complex glacier contours and shape features.

6. The glacier change remote sensing detection system based on local detail feature aggregation according to claim 2, characterized in that, The multi-scale difference prediction module includes: A multi-scale prediction unit is used to output dual-temporal multi-scale glacier prediction maps at different stages of the decoder; The difference map generation unit is used to obtain a multi-scale difference map by subtracting elements from the dual-temporal multi-scale glacier feature maps at the same scale. The classification and discrimination unit is used to make decisions on multi-scale difference features through a classifier, thereby outputting a binary detection result map of glacier changes.

Images