Remote sensing image change detection method based on measurement fusion and scale perception enhancement

Abstract

The invention relates to a remote sensing image change detection method based on measurement fusion and scale perception enhancement. The remote sensing image change detection with high accuracy, high detail integrity and high robustness in a complex environment is realized. Preprocessing a remote sensing image data set formed by the remote sensing images at different moments in the same area, inputting the preprocessed remote sensing image data set into a preset neural network, and extracting to obtain a first moment feature map and a second moment feature map at different levels; performing non-parameter and feature enhancement on the first moment feature map and the second moment feature map on each level to obtain a first moment enhanced feature map and a second moment enhanced feature map; performing cross fusion on the first moment enhanced feature map and the second moment enhanced feature map through different input sequences on each level, and integrating the levels after fusion to obtain coding features; and performing lightweight multi-scale feature extraction on the coding features, performing fusion conversion after extraction to obtain a pixel-level change probability graph, and performing threshold segmentation by adopting a preset threshold to obtain a binary change graph.

Description

Technical Field

[0001] This application relates to the field of satellite remote sensing technology, and in particular to a method for detecting changes in remote sensing images based on metric fusion and scale perception enhancement. Background Technology

[0002] Currently, parallel optimization methods for large-scale remote sensing image change detection tasks are based on three implementation approaches: cluster computing, grid computing, and cloud computing. Cluster computing organizes multiple computing nodes into a working cluster and uses scheduling software to control the system. Unlike cluster computing systems, grid computing can not only perform distributed parallel computing processing but also utilize heterogeneous computing resources for remote sensing image change detection. Cloud computing evolved from grid computing and can process large amounts of remote sensing image data in a highly fault-tolerant environment. Compared to grid computing, cloud computing has lower computing costs and offers advantages in elasticity and scalability.

[0003] With the significant improvement in remote sensing image acquisition capabilities, changes in ground features in high-resolution images exhibit characteristics such as subtlety, multi-scale, and enhanced nonlinear features. While existing deep learning change detection methods can automatically learn semantic information, they generally lack explicit difference measurement mechanisms. Modeling feature similarity and differences relies on automatic learning by the network. When the change contrast is low or the noise is strong, these methods struggle to accurately extract true change features, leading to missed detections of subtle changes. Furthermore, existing models lack sufficient multi-scale feature fusion, making it difficult to identify small changes and inaccurately locate change edges. In addition, differences in illumination, season, and imaging angle between images from different time phases introduce spurious changes into the feature space. Deep learning networks often produce different results when the input order of two time phases is interchanged, indicating significant deficiencies in their robustness and stability.

[0004] A search of existing technical literature revealed a patent application (application number 202210866783.6) entitled "A Method and Apparatus for Detecting Changes in Remote Sensing Images." This patent extracts feature maps from different time phases and fuses the change features from these feature maps. It uses an unsupervised learning module to reduce the differences in change region features across different time phases, thereby strengthening the change features and improving the accuracy of change detection. However, this patent suffers from problems such as a lack of explicit difference measurement, making it difficult to identify subtle image changes and handle high noise interference. Summary of the Invention

[0005] Therefore, it is necessary to provide a remote sensing image change detection method based on metric fusion and scale perception enhancement to address the above-mentioned technical problems, so as to achieve high accuracy, high detail integrity and high robustness in remote sensing image change detection under complex environments.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for detecting changes in remote sensing images based on metric fusion and scale-aware enhancement, the method comprising: S1: Preprocess the remote sensing image dataset to obtain a preprocessed remote sensing image dataset, wherein the remote sensing image dataset includes pairs of remote sensing images of the same area at different times; S2: Input the remote sensing image pairs in the preprocessed remote sensing image dataset into the trained preset neural network, and output the first time feature map and the second time feature map at different levels. S3: Calculate the feature map difference degree for the first time-time feature map and the second time-time feature map at each level to obtain a difference degree feature map. Perform feature enhancement processing on the first time-time feature map and the difference degree feature map, and the second time-time feature map and the difference degree feature map at each level to obtain a first time-time enhanced feature map and a second time-time enhanced feature map. S4: The enhanced feature map at the first time step and the enhanced feature map at the second time step are concatenated along the channel dimension at each level according to different input orders to obtain the first concatenated feature and the second concatenated feature. The first concatenated feature and the second concatenated feature are respectively subjected to feature fusion processing to obtain the first fused feature and the second fused feature. The cross-fusion feature is calculated based on the first fused feature and the second fused feature. The cross-fusion feature is integrated at all levels to obtain the encoded feature. S5: A lightweight convolutional network is used to extract multi-scale features from the encoded features. The extracted features at different scales are fused and transformed to obtain a pixel-level change probability map. The pixel-level change probability map is then segmented using a preset threshold to obtain a binary change map.

[0007] Preferably, the preprocessing includes: Based on the one-to-one correspondence of all pixels in the remote sensing image pair in spatial location, geometric registration and radiometric correction are performed on the remote sensing image dataset to obtain the first preprocessed remote sensing image dataset. The first preprocessed remote sensing image dataset is cropped to a uniform size using a sliding window method or a random cropping method to obtain the second preprocessed remote sensing image dataset. The second preprocessed remote sensing image dataset was subjected to image enhancement and normalization processes using random flipping, random rotation, random cropping, and channel-based normalization to obtain the preprocessed remote sensing image dataset.

[0008] Preferably, the remote sensing image dataset further includes change label maps corresponding to the remote sensing image pairs, and before step S1, it further includes: The remote sensing image dataset is divided into a training set, a validation set, and a test set according to a preset ratio; The loss function is constructed using the cross-entropy loss function, the Dice loss function, or a weighted combination of the two. Based on the loss function and the change label map, the training set is input into the preset neural network for training to obtain the trained preset neural network, and the performance of the trained preset neural network is evaluated using the validation set and the test set.

[0009] Preferably, the trained pre-trained neural network includes a first convolutional neural network, a residual network, and a second convolutional neural network. The first convolutional neural network and the second convolutional neural network are twin convolutional neural networks with shared weights. The residual network is pre-trained with parameters using a public dataset. The preprocessed remote sensing image dataset includes remote sensing images at a first time and remote sensing images at a second time. The remote sensing images at the first time are sequentially processed by a first convolutional neural network, a residual network, and a second convolutional neural network to output first-time feature maps at different levels. The remote sensing images at the second time are sequentially processed by a first convolutional neural network, a residual network, and a second convolutional neural network to output second-time feature maps at different levels.

[0010] Preferably, after step S2, the method further includes: Variable convolution is used to perform prior feature enhancement on the feature maps at the first and second time points.

[0011] Preferably, the difference feature map The calculation expression is as follows:

[0012]

[0013] in, This represents the feature map at the first moment. This represents the feature map at the second time step. Represents the cosine similarity function. Indicates the first Hierarchy; The feature enhancement process includes: The first time-step feature map and the difference feature map, the second time-step feature map and the difference feature map are respectively concatenated along the channel dimension and subjected to a first convolution at each level to obtain the first time-step enhanced feature map and the second time-step enhanced feature map.

[0014] Preferably, a third convolutional neural network or a cross-attention mechanism is used to perform feature fusion processing on the first concatenated features and the second concatenated features, respectively.

[0015] Preferably, the cross-fusion feature The calculation expression is as follows:

[0016] in, Indicates the first fusion feature, Indicates the second fusion feature, Indicates the first Hierarchy; The integration process across all levels includes: The encoded features are obtained by combining the cross-fusion features from all levels through upsampling or interpolation, concatenating the channel dimensions, and performing a second convolution.

[0017] Preferably, the fusion and conversion process includes: The extracted features at different scales are then subjected to spatial alignment, channel dimension concatenation, and a third convolution to obtain multi-scale enhanced features. The multi-scale enhanced features are processed sequentially through a fourth convolution and a nonlinear activation function to obtain a pixel-level change probability map.

[0018] Preferably, after obtaining the binary transformation map, the method further includes post-processing the binary transformation map, the post-processing including: Morphological opening and closing operations are used to remove isolated noise points from the binary transformation map to obtain a denoised binary transformation map. Connectivity analysis is then used to merge small regions in the denoised binary transformation map to obtain a post-processed binary transformation map. An edge smoothing strategy is used to optimize the boundary of the change region in the binary change map, resulting in a post-processed binary change map.

[0019] Compared with the prior art, the beneficial effects of the present invention are: This invention provides a remote sensing image change detection method based on metric fusion and scale-aware enhancement. A parameterless feature metric mechanism for acquiring difference feature maps at different levels for remote sensing image pairs of the same area at different times enables explicit modeling of change difference features. A cross-fusion strategy using different input orders at different levels mitigates interference from non-ground object differences and ensures that feature extraction is insensitive to temporal order. Finally, a lightweight multi-scale feature enhancement structure is employed, balancing local and global scale changes, improving the ability to identify targets changing at different scales without significantly increasing model parameters and computational load. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of a remote sensing image change detection method based on metric fusion and scale-aware enhancement in one embodiment. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0022] Example 1 like Figure 1 As shown in the figure, this embodiment proposes a remote sensing image change detection method based on metric fusion and scale-aware enhancement. The method includes: S1: Preprocess the remote sensing image dataset to obtain a preprocessed remote sensing image dataset, wherein the remote sensing image dataset includes pairs of remote sensing images of the same area at different times; S2: Input the remote sensing image pairs in the preprocessed remote sensing image dataset into the trained preset neural network, and output the first time feature map and the second time feature map at different levels. S3: Calculate the feature map difference degree for the first time-time feature map and the second time-time feature map at each level to obtain a difference degree feature map. Perform feature enhancement processing on the first time-time feature map and the difference degree feature map, and the second time-time feature map and the difference degree feature map at each level to obtain a first time-time enhanced feature map and a second time-time enhanced feature map. The specific implementation of this step is as follows: First-time feature maps and second-time feature maps at different levels are obtained. That is, after the basic feature extraction is completed, a parametric feature enhancement module is introduced to explicitly measure the difference in high-dimensional features of the dual-temporal images at each level. The input of this module is the dual-temporal features at the same level. and The output is the difference feature map of the corresponding level. And the enhanced feature representation. At each spatial location, the feature vectors for time A and time B are denoted as follows: and This method uses cosine similarity as the metric function to calculate feature similarity. Cosine similarity is calculated on the feature vectors expanded along each channel dimension, and can be achieved through vector dot product and norm operations. The resulting similarity value ranges between 0 and 1, with lower values ​​indicating greater feature differences at that location. To transform the similarity metric into a difference metric, this embodiment performs a first-order transformation on the cosine similarity to obtain a difference feature map. This allows a larger difference value to indicate a more significant change. This transformation process relies solely on mathematical operations and does not introduce any trainable parameters, making it a parameterless metric calculation structure that effectively suppresses the increase in model complexity. The aforementioned parameterless metric can operate simultaneously at multiple levels. In this embodiment, difference metrics are calculated at four levels for both shallow detail and deep semantic features, resulting in a set of multi-level difference feature sequences. This allows for a rich depiction of information about changes at different spatial scales.

[0023] Obtain the difference feature map of each layer Then, the difference feature maps of each layer are fused with the corresponding basic semantic features to form enhanced features containing difference information, i.e., feature enhancement processing, resulting in the enhanced feature map at time A. Enhanced feature map at time B This operation allows explicit metric information to be embedded into the feature space in a controlled manner, improving the ability to respond to subtle changes without increasing the size of the network parameters.

[0024] S4: The enhanced feature map at the first time step and the enhanced feature map at the second time step are concatenated along the channel dimension at each level according to different input orders to obtain the first concatenated feature and the second concatenated feature. The first concatenated feature and the second concatenated feature are respectively subjected to feature fusion processing to obtain the first fused feature and the second fused feature. The cross-fusion feature is calculated based on the first fused feature and the second fused feature. The cross-fusion feature is integrated at all levels to obtain the encoded feature. The specific implementation of this step is as follows: To further improve the robustness and stability of the feature encoding process, a robust semantic cross-fusion module is set up on the basis of parametric feature enhancement. The goal of this module is to symmetrically fuse the bi-temporal enhanced features in a high-dimensional semantic space. On the one hand, this alleviates the instability caused by unilateral feature shift, and on the other hand, the symmetrical structure ensures that the output result remains unchanged when the input temporal sequence is exchanged.

[0025] For each level, the feature map at time A is enhanced. Enhanced feature map at time B As input, the two are concatenated along the channel dimension to obtain the concatenated feature. By inputting convolutional units or cross-attention units with shared parameters, the fused features of the first branch are obtained. Simultaneously, the input order is reversed, and the feature map at time B is enhanced. Enhanced feature map at time A By splicing along the channel dimension, we obtain By inputting the same convolutional unit or cross-attention unit, the fused features of the second branch are obtained. Since the two branches share the same parameters, the structures of the first and second branches are strictly symmetrical.

[0026] The outputs of the first and second branches are averaged or maximized element-wise at corresponding positions to obtain the symmetrically fused hierarchical semantic features, i.e., the cross-fusion features. This feature simultaneously incorporates high-level semantic information from both temporal phases and utilizes a symmetrical structure to eliminate the influence of a single input order on the results, thus ensuring consistent network output when image inputs at time A and time B are swapped. Cross-fusion features across all levels. After integration, a comprehensive representation of the change characteristics in the coding stage is obtained. This encoding feature As input to the decoding and discrimination networks, it will be used to generate the final transformation map.

[0027] S5: A lightweight convolutional network is used to extract multi-scale features from the encoded features. The extracted features at different scales are fused and transformed to obtain a pixel-level change probability map. The pixel-level change probability map is then segmented using a preset threshold to obtain a binary change map.

[0028] The specific implementation of this step is as follows, in encoding features Building upon this foundation, this embodiment constructs a lightweight multi-scale feature extraction and enhancement module to jointly model changes within different receptive fields. This module employs a multi-branch convolutional structure, achieving multi-scale feature perception through combinations of different convolutional kernel sizes, dilation rates, and downsampling paths. The lightweight multi-scale feature extraction and enhancement module comprises three parallel sub-branches. These sub-branches utilize depthwise separable convolutions or other lightweight convolutional structures for feature extraction, preserving multi-scale information while significantly reducing parameter and computational loads. The first sub-branch is a standard convolutional layer, extracting local detail changes and encoding features. The first branch directly uses a standard 3×3 convolution, combined with batch normalization and a non-linear activation function to extract detailed features of local changes at the original scale; the second sub-branch is a dilated convolution layer, which expands the receptive field to extract large-scale changes and enhances the encoded features. Dilated convolutions are used with an dilation rate greater than 1 to expand the effective receptive field of the convolution kernel, thereby capturing a wide range of spatial contextual change information; the third sub-branch consists of downsampling and upsampling layers to encode features. First, perform downsampling operations, such as reducing the feature map resolution through convolution or pooling with a stride greater than 1. Then, perform convolution in the low-resolution space to extract coarse-scale variation features. Finally, upsample back to the original resolution through interpolation or deconvolution.

[0029] The features at different scales output from the three sub-branches are fused at multiple scales. After fusion, the feature dimensions are further compressed through a change discrimination sub-network, and a pixel-level change probability map is output. The pixel-level change probability map is then thresholded, preferably using a threshold of 0.5, converting it into a binary change map. In the binary map, pixels with a value of 1 indicate a detected change, and pixels with a value of 0 indicate no detected change. The threshold can be adjusted appropriately according to task requirements to balance the false negative rate and the false positive rate.

[0030] Example 2 This embodiment further elaborates on the remote sensing image change detection method based on metric fusion and scale perception enhancement proposed in Embodiment 1.

[0031] The preprocessing includes: Based on the one-to-one correspondence of all pixels in the remote sensing image pair in spatial location, geometric registration and radiometric correction are performed on the remote sensing image dataset to obtain the first preprocessed remote sensing image dataset. The first preprocessed remote sensing image dataset is cropped to a uniform size using a sliding window method or a random cropping method to obtain the second preprocessed remote sensing image dataset. The second preprocessed remote sensing image dataset was subjected to image enhancement and normalization processes using random flipping, random rotation, random cropping, and channel-based normalization to obtain the preprocessed remote sensing image dataset.

[0032] The specific implementation of this step is as follows: The remote sensing image pair includes images at time A and time B. Geometric registration and radiometric correction are performed on all image pairs to ensure a one-to-one correspondence between the images at time A and time B in pixel-level spatial location. The images are then uniformly cropped, for example, into fixed-size image blocks, with the block size set to 256×256 pixels. This size can be adjusted according to the actual GPU memory capacity. During cropping, a sliding window or random cropping method can be used to simultaneously slice image pairs and labels covering the same area, forming standardized change detection samples. In the data loading stage, image enhancement and normalization processing is performed on the input images, such as random flipping, random rotation, random cropping, and channel-based normalization. This ensures the consistency of the dual-temporal image enhancement strategy, improves the model's generalization ability, and ensures that the change label maps corresponding to the remote sensing image pairs in the remote sensing image dataset undergo only geometric transformations and not intensity transformations.

[0033] The remote sensing image dataset also includes change label maps corresponding to the remote sensing image pairs. Before step S1, it also includes: The remote sensing image dataset is divided into a training set, a validation set, and a test set according to a preset ratio; The loss function is constructed using the cross-entropy loss function, the Dice loss function, or a weighted combination of the two. Based on the loss function and the change label map, the training set is input into the preset neural network for training to obtain the trained preset neural network, and the performance of the trained preset neural network is evaluated using the validation set and the test set.

[0034] The specific implementation of this step is as follows: high-resolution remote sensing image pairs are stored locally or on a server in the form of image blocks. Each pair of samples contains remote sensing images of the same area at two different times, denoted as image at time A and image at time B, and is accompanied by a change label map. The image data is organized into training set, validation set and test set folder structure according to the dataset type. Each type is further divided into independent subdirectories according to time A, time B and change label. The directory structure is unified as follows: Train / Val / Test → three types of subfolders: A, B and Label.

[0035] During the training phase, this embodiment uses remote sensing image pairs with known change labels as training samples. Image A, image B, and their corresponding change labels are input as supervisory signals into a pre-defined neural network. A loss function is constructed based on the cross-entropy loss function, the Dice loss function, or a weighted combination of both to optimize the trainable parameters in the network. For each training batch, the dual-temporal image pairs are first input into the pre-defined neural network, and the difference features are calculated and enhanced features are generated through the parameterless metric feature enhancement module in step S3. Then, the robust semantic cross-fusion module in step S4 and the lightweight multi-scale feature extraction and enhancement module in step S5 are passed sequentially, finally outputting a pixel-level change probability map. The output change probability is compared with the true labels, the loss value is calculated, and the network parameters are updated through the backpropagation algorithm. This process adopts a supervised learning approach, using cross-entropy loss or Dice loss as the optimization objective, and updating the model parameters through backpropagation. In this process, parameterless metric feature enhancement does not introduce additional trainable parameters during training, thereby ensuring the stability and controllability of training.

[0036] After model training, performance is evaluated using validation and test sets. Performance evaluation metrics include, but are not limited to, pixel-level accuracy, overall accuracy, recall, mean intersection-over-union ratio (CIU), and F1 score. By comparing with baseline change detection models, the performance improvement of this invention in subtle change detection, multi-scale target recognition, and robustness in complex scenes can be verified. In the actual deployment phase, the trained model is loaded into the inference environment. Only dual-temporal remote sensing images of the area to be detected need to be input, and the corresponding change detection results will be automatically output, eliminating the need for manual threshold setting or complex feature design. Because this embodiment employs a collaborative design of parametric feature enhancement and lightweight multi-scale structure, it maintains low model complexity and high inference efficiency while ensuring detection accuracy and boundary clarity, making it suitable for widespread application in large-scale remote sensing monitoring and operational systems.

[0037] The pre-trained neural network includes a first convolutional neural network, a residual network, and a second convolutional neural network. The first convolutional neural network and the second convolutional neural network are twin convolutional neural networks with shared weights. The residual network is pre-trained with parameters using a public dataset. The preprocessed remote sensing image dataset includes remote sensing images at a first time and remote sensing images at a second time. The remote sensing images at the first time are sequentially processed by a first convolutional neural network, a residual network, and a second convolutional neural network to output first-time feature maps at different levels. The remote sensing images at the second time are sequentially processed by a first convolutional neural network, a residual network, and a second convolutional neural network to output second-time feature maps at different levels.

[0038] The specific implementation of this step is as follows: a Siamese convolutional neural network is constructed as the basic feature extraction backbone, and a residual network structure is used as the main network. Since the residual network has a large number of parameters, the parameters are trained on a public dataset such as ImageNet, so that the residual network has a certain initial state. The backbone network contains multiple layers of convolution and downsampling modules. In this embodiment, four feature maps of different levels are output, representing the image representation from shallow details to deep semantics.

[0039] For each pair of input images, the image at time A and the image at time B are respectively input into a Siamese backbone network with shared weights. Through layer-by-layer convolution and downsampling operations, the feature map of the image at time A is obtained at each layer. Feature map of time B image .

[0040] After step S2, the following is also included: Variable convolution is used to perform prior feature enhancement on the feature maps at the first and second time points.

[0041] The specific implementation of this step is as follows: to enhance the expressive power of basic semantic features, the feature map of the image at time A is obtained. Feature map of time B image Subsequently, residual enhancement units are added to stack the convolution results of the same layer step by step. For example, three-branch convolution and step-by-step addition are used to form prior enhancement features, thereby improving the nonlinear representation capability of high-dimensional features and providing a more stable basic semantic representation for subsequent metric feature enhancement, which reduces the impact of image feature alignment bias to a certain extent.

[0042] Example 3 This embodiment further elaborates on the remote sensing image change detection method based on metric fusion and scale perception enhancement proposed in Embodiment 1.

[0043] The difference feature map The calculation expression is as follows:

[0044]

[0045] in, This represents the feature map at the first moment. This represents the feature map at the second time step. Represents the cosine similarity function. Indicates the first Hierarchy; The feature enhancement process includes: The first time-step feature map and the difference feature map, the second time-step feature map and the difference feature map are respectively concatenated along the channel dimension and subjected to a first convolution at each level to obtain the first time-step enhanced feature map and the second time-step enhanced feature map.

[0046] The specific implementation of this step is as follows: for the feature map at time A of each layer. Feature map at time B and difference characteristics The extended feature representation is obtained by concatenating the channels, and then the number of channels is compressed and reorganized by 1×1 convolution to obtain the enhanced feature map at time A. Enhanced feature map at time B .

[0047] The first concatenated feature and the second concatenated feature are respectively processed by a third convolutional neural network or a cross-attention mechanism.

[0048] The specific implementation of this step is as follows: cross-fusion is achieved within an attention framework. Using features from one side as the query and features from the other side as the key and value, attention weights across features are calculated. Long-distance dependencies are modeled in the multi-head cross-attention unit, and the output features are further enhanced through residual connections and a multilayer perceptron. This embodiment separately addresses the splicing of features. , Using features from one side as queries and features from the other side as keys and values, a multi-head cross-attention mechanism is employed for feature fusion. After fusion, residual connections and a multilayer perceptron are used for feature enhancement to obtain the fused features. , This approach can improve the expressive power of features for complex scene variations without significantly increasing the number of parameters.

[0049] The cross-fusion feature The calculation expression is as follows:

[0050] in, Indicates the first fusion feature, Indicates the second fusion feature, Indicates the first Hierarchy; The integration process across all levels includes: The encoded features are obtained by combining the cross-fusion features from all levels through upsampling or interpolation, concatenating the channel dimensions, and performing a second convolution.

[0051] The specific implementation of this step is as follows, involving the cross-fusion features of all levels. The process involves integration, aligning deep features to a higher resolution space step by step through upsampling or interpolation operations, aggregating them along the channel dimension, and using convolutional layers for channel compression and feature reshaping to obtain a comprehensive variation feature representation from the encoding stage. .

[0052] The fusion and conversion process includes: The extracted features at different scales are then subjected to spatial alignment, channel dimension concatenation, and a third convolution to obtain multi-scale enhanced features. The multi-scale enhanced features are processed sequentially through a fourth convolution and a nonlinear activation function to obtain a pixel-level change probability map.

[0053] The specific implementation of this step is as follows: After aligning the different scale features output by the three sub-branches in terms of spatial scale, they are concatenated along the channel dimension to form a multi-scale fused feature map. To avoid increasing the model complexity due to an excessive number of channels, a 1×1 convolution is performed on the fused feature map to reduce its dimensionality, resulting in the enhanced multi-scale features, i.e., the multi-scale enhanced features. To further enhance the ability to model inter-channel dependencies, a lightweight channel attention structure can be introduced onto multi-scale fused features. This structure establishes dependencies within the local channel neighborhood using only one-dimensional convolution, thereby enhancing the response strength of useful changing features at a lower cost and suppressing redundant features.

[0054] Multi-scale enhancement features The data is fed into a change-discrimination subnetwork, which can consist of several layers of convolutions and non-linear activation functions to further compress the feature dimensions and output a change probability map. This embodiment... Perform a 1×1 convolution to compress the number of channels to 1 or 2, and then use the Sigmoid or Softmax function to obtain a pixel-level change probability map.

[0055] After obtaining the binary transformation map, the method further includes post-processing of the binary transformation map, the post-processing including: Morphological opening and closing operations are used to remove isolated noise points from the binary transformation map to obtain a denoised binary transformation map. Connectivity analysis is then used to merge small regions in the denoised binary transformation map to obtain a post-processed binary transformation map. An edge smoothing strategy is used to optimize the boundary of the change region in the binary change map, resulting in a post-processed binary change map.

[0056] The specific implementation of this step is as follows. In practical applications, post-processing operations can be performed on the binary change map, such as using morphological opening and closing operations to remove isolated noise points, using connected component analysis methods to merge small regions, or using edge smoothing strategies to optimize the coherence of the boundary of the change region, thereby obtaining change detection results that are easier to interpret.

Claims

1. A method for detecting changes in remote sensing images based on metric fusion and scale-aware enhancement, characterized in that, include: S1: Preprocess the remote sensing image dataset to obtain a preprocessed remote sensing image dataset, wherein the remote sensing image dataset includes pairs of remote sensing images of the same area at different times; S2: Input the remote sensing image pairs in the preprocessed remote sensing image dataset into the trained preset neural network, and output the first time feature map and the second time feature map at different levels. S3: Calculate the feature map difference degree for the first time-time feature map and the second time-time feature map at each level to obtain a difference degree feature map. Perform feature enhancement processing on the first time-time feature map and the difference degree feature map, and the second time-time feature map and the difference degree feature map at each level to obtain a first time-time enhanced feature map and a second time-time enhanced feature map. S4: The enhanced feature map at the first time step and the enhanced feature map at the second time step are concatenated along the channel dimension at each level according to different input orders to obtain the first concatenated feature and the second concatenated feature. The first concatenated feature and the second concatenated feature are respectively subjected to feature fusion processing to obtain the first fused feature and the second fused feature. The cross-fusion feature is calculated based on the first fused feature and the second fused feature. The cross-fusion feature is integrated at all levels to obtain the encoded feature. S5: A lightweight convolutional network is used to extract multi-scale features from the encoded features. The extracted features at different scales are fused and transformed to obtain a pixel-level change probability map. The pixel-level change probability map is then segmented using a preset threshold to obtain a binary change map.

2. The remote sensing image change detection method based on metric fusion and scale-aware enhancement according to claim 1, characterized in that, The preprocessing includes: Based on the one-to-one correspondence of all pixels in the remote sensing image pair in spatial location, geometric registration and radiometric correction are performed on the remote sensing image dataset to obtain the first preprocessed remote sensing image dataset. The first preprocessed remote sensing image dataset is cropped to a uniform size using a sliding window method or a random cropping method to obtain the second preprocessed remote sensing image dataset. The second preprocessed remote sensing image dataset was subjected to image enhancement and normalization processes using random flipping, random rotation, random cropping, and channel-based normalization to obtain the preprocessed remote sensing image dataset.

3. The remote sensing image change detection method based on metric fusion and scale-aware enhancement according to claim 1, characterized in that, The remote sensing image dataset also includes change label maps corresponding to the remote sensing image pairs. Before step S1, it also includes: The remote sensing image dataset is divided into a training set, a validation set, and a test set according to a preset ratio; The loss function is constructed using the cross-entropy loss function, the Dice loss function, or a weighted combination of the two. Based on the loss function and the change label map, the training set is input into the preset neural network for training to obtain the trained preset neural network, and the performance of the trained preset neural network is evaluated using the validation set and the test set.

4. The remote sensing image change detection method based on metric fusion and scale-aware enhancement according to claim 3, characterized in that, The pre-trained neural network includes a first convolutional neural network, a residual network, and a second convolutional neural network. The first convolutional neural network and the second convolutional neural network are twin convolutional neural networks with shared weights. The residual network is pre-trained with parameters using a public dataset. The preprocessed remote sensing image dataset includes remote sensing images at a first time and remote sensing images at a second time. The remote sensing images at the first time are sequentially processed by a first convolutional neural network, a residual network, and a second convolutional neural network to output first-time feature maps at different levels. The remote sensing images at the second time are sequentially processed by a first convolutional neural network, a residual network, and a second convolutional neural network to output second-time feature maps at different levels.

5. The remote sensing image change detection method based on metric fusion and scale-aware enhancement according to claim 1, characterized in that, After step S2, the following is also included: Variable convolution is used to perform prior feature enhancement on the feature maps at the first and second time points.

6. The remote sensing image change detection method based on metric fusion and scale-aware enhancement according to claim 1, characterized in that, The difference feature map The calculation expression is as follows: in, This represents the feature map at the first moment. This represents the feature map at the second time step. Represents the cosine similarity function. Indicates the first Hierarchy; The feature enhancement process includes: The first time-step feature map and the difference feature map, the second time-step feature map and the difference feature map are respectively concatenated along the channel dimension and subjected to a first convolution at each level to obtain the first time-step enhanced feature map and the second time-step enhanced feature map.

7. The remote sensing image change detection method based on metric fusion and scale-aware enhancement according to claim 1, characterized in that, The first concatenated feature and the second concatenated feature are respectively processed by a third convolutional neural network or a cross-attention mechanism.

8. The remote sensing image change detection method based on metric fusion and scale-aware enhancement according to claim 1, characterized in that, The cross-fusion feature The calculation expression is as follows: in, Indicates the first fusion feature, Indicates the second fusion feature, Indicates the first Hierarchy; The integration process across all levels includes: The encoded features are obtained by combining the cross-fusion features from all levels through upsampling or interpolation, concatenating the channel dimensions, and performing a second convolution.

9. The remote sensing image change detection method based on metric fusion and scale-aware enhancement according to claim 1, characterized in that, The fusion and conversion process includes: The extracted features at different scales are then subjected to spatial alignment, channel dimension concatenation, and a third convolution to obtain multi-scale enhanced features. The multi-scale enhanced features are processed sequentially through a fourth convolution and a nonlinear activation function to obtain a pixel-level change probability map.

10. The remote sensing image change detection method based on metric fusion and scale-aware enhancement according to claim 1, characterized in that, After obtaining the binary transformation map, the method further includes post-processing of the binary transformation map, the post-processing including: Morphological opening and closing operations are used to remove isolated noise points from the binary transformation map to obtain a denoised binary transformation map. Connectivity analysis is then used to merge small regions in the denoised binary transformation map to obtain a post-processed binary transformation map. An edge smoothing strategy is used to optimize the boundary of the change region in the binary change map, resulting in a post-processed binary change map.

Images