Glacier and lake automatic identification method based on multi-source remote sensing and domain self-adaption

Abstract

The invention relates to a glacier and lake automatic identification method based on multi-source remote sensing and domain self-adaption. The method comprises the steps that remote sensing image data are acquired and preprocessed; a glacier and glacier lake classification deep learning model is constructed, the model comprises an encoder part and a decoder part, the encoder part is composed of a convolution layer, a Swin Transform module and a channel attention mechanism, and the decoder part adopts transposition convolution and multi-scale jump connection to fuse a feature map; training a glacier-lake classification deep learning model by adopting the preprocessed remote sensing image data, and migrating the trained glacier-lake classification deep learning model to a target field for field adaptation adjustment; and inputting a to-be-predicted remote sensing image into the glacier-lake classification deep learning model subjected to field adaptation adjustment, outputting a glacier-lake extraction result, and performing reduction and noise reduction processing on the glacier-lake extraction result to obtain a glacier-lake recognition result. Effective extraction of glacier and glacier lake features in the remote sensing image is realized.

Description

Technical Field

[0001] This application relates to the field of remote sensing image processing technology, and in particular to an automatic identification method for glaciers and glacial lakes based on multi-source remote sensing and domain adaptation. Background Technology

[0002] With the rapid development of remote sensing technology and deep learning, automatic identification technology based on remote sensing imagery has become an important means of Earth monitoring. In the extraction and identification of glaciers and glacial lakes, existing technologies mainly employ a single remote sensing data source or traditional image processing methods, combined with shallow machine learning models. For example, some methods utilize the spectral features and texture information of optical images to extract glacier boundaries through threshold segmentation or machine learning classification; other studies have built semantic segmentation models for glaciers and glacial lakes based on convolutional neural networks.

[0003] However, existing technical solutions have significant drawbacks: on the one hand, their ability to perceive local details and overall structural features of glaciers and glacial lakes is limited, making it difficult to accurately capture complex boundaries and small-scale changes. Furthermore, the multi-source remote sensing data fusion methods are not perfect enough to fully leverage the advantages of multi-sensor and multi-band information, resulting in difficulty in improving recognition accuracy. On the other hand, in cross-regional application scenarios, due to significant differences in spectral response, terrain background, and climate conditions among remote sensing data from different regions, and the fact that existing methods have not fully utilized domain adaptation techniques to alleviate the problem of inconsistent data distribution across different regions, the generalization ability of the training model is insufficient, which seriously affects the recognition accuracy and stable application of the model in large-scale complex environments.

[0004] Therefore, there is an urgent need for a technology that can improve the accuracy and robustness of glacier and glacial lake identification in a wide range of diverse geographical environments. Summary of the Invention

[0005] Therefore, it is necessary to address the aforementioned technical issues by providing an automatic glacier and glacier lake identification method based on multi-source remote sensing and domain adaptation that can improve the accuracy and robustness of glacier and glacier lake identification in large-scale and diverse geographical environments.

[0006] Firstly, this application provides an automatic identification method for glaciers and glacial lakes based on multi-source remote sensing and domain adaptation. The method includes: Acquire remote sensing image data and perform preprocessing; A deep learning model for classifying glaciers and lakes is constructed. The model includes an encoder and a decoder. The encoder consists of convolutional layers, a Swing Transformer module, and a channel attention mechanism. The decoder uses transposed convolution and multi-scale skip connections to fuse feature maps. The deep learning model for glacier and glacial lake classification is trained using preprocessed remote sensing image data, and then the trained deep learning model for glacier and glacial lake classification is transferred to the target domain for domain adaptation adjustment. The remote sensing image to be predicted is input into the domain-adapted deep learning model for glacier and lake classification, which outputs the glacier and lake extraction results. The glacier and lake extraction results are then subjected to restoration and noise reduction processing to obtain the glacier and lake identification results.

[0007] Optionally, in one embodiment of this application, the preprocessing includes: The remote sensing image data is subjected to radiometric calibration, atmospheric correction, geometric correction, and resampling operations. The remote sensing image data is normalized and sliced ​​using a sliding window.

[0008] Optionally, in one embodiment of this application, the Swing Transformer module includes a sliding window multi-head self-attention mechanism and a window self-attention mechanism, which calculate multi-head attention both inside and between windows to obtain global and local semantic information of the feature image.

[0009] Optionally, in one embodiment of this application, the channel attention mechanism includes an average pooling layer, a max pooling layer, a multilayer perceptron, and an activation function.

[0010] Optionally, in one embodiment of this application, the noise reduction processing of the glacier-lake extraction results includes: Image stitching and coordinate system restoration were performed on the extracted glacier and glacial lake results; The restored image is then subjected to noise removal, internal holes in the extracted results are filled, and non-glacial lake water is removed.

[0011] Optionally, in one embodiment of this application, the image stitching and coordinate system restoration of the glacier-lake extraction results includes: The total geographic extent of the stitched area and the image geographic transformation matrix are determined based on the geographic transformation information extracted from glaciers and glacial lakes. Image stitching is performed based on the total geographic range of the stitching area and the image geographic transformation matrix, and the overlapping area data is fused using a weighted average method. Restore the projection, coordinate system, and spatial reference information of the image stitching result.

[0012] Optionally, in one embodiment of this application, the noise removal of the restored image includes: Morphology-based erosion and dilation operations are used to remove discrete noise points and restore the overall outline of ground features; The overall contour boundary is smoothed using the Gaussian blur method.

[0013] Secondly, this application also provides an automatic glacier and glacial lake identification device based on multi-source remote sensing and domain adaptation. The device includes: The data acquisition and preprocessing module is used to acquire remote sensing image data and perform preprocessing. The model building module is used to build a deep learning model for glacier and lake classification. The model includes an encoder and a decoder. The encoder consists of convolutional layers, a Swing Transformer module, and a channel attention mechanism. The decoder uses transposed convolution and multi-scale skip connections to fuse feature maps. The model training and transfer module is used to train the deep learning model for glacier and lake classification using preprocessed remote sensing image data, and to transfer the trained deep learning model for glacier and lake classification to the target domain for domain adaptation. The glacier and glacial lake identification module is used to input the remote sensing image to be predicted into the domain-adapted deep learning model for glacier and glacial lake classification, output the glacier and glacial lake extraction results, and perform restoration and noise reduction processing on the glacier and glacial lake extraction results to obtain the glacier and glacial lake identification results.

[0014] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the steps of the methods described in the various embodiments above.

[0015] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps of the methods described in the various embodiments above.

[0016] The aforementioned automatic glacier and glacial lake identification method based on multi-source remote sensing and domain adaptation first acquires and preprocesses remote sensing image data. Then, a deep learning model for glacier and glacial lake classification is constructed, comprising an encoder and a decoder. The encoder consists of convolutional layers, a Swing Transformer module, and a channel attention mechanism, while the decoder uses transposed convolution and multi-scale skip connections to fuse feature maps. Next, the preprocessed remote sensing image data is used to train the deep learning model, which is then adapted to the target domain. Finally, the remote sensing image to be predicted is input into the domain-adapted deep learning model, which outputs the extracted glacier and glacial lake results. These results are then subjected to noise reduction and restoration processing to obtain the glacier and glacial lake identification result. In other words, by constructing a novel deep learning model, effective extraction of glacier and glacial lake features from remote sensing images is achieved. Combined with domain adaptation technology, this improves the model's generalization ability and recognition accuracy under cross-regional and multi-source data conditions. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an automatic glacier and glacial lake identification method based on multi-source remote sensing and domain adaptation in one embodiment. Figure 2 This is a schematic diagram of the dataset distribution in one embodiment; Figure 3 This is a schematic diagram of a sliding window slice in one embodiment; Figure 4 This is a schematic diagram of the structure of a deep learning model for glacier and glacial lake classification in one embodiment. Figure 5 This is a schematic diagram of hole filling in one embodiment; Figure 6 This is a schematic diagram of image stitching and coordinate restoration in one embodiment; Figure 7 This is a schematic diagram illustrating the noise removal effect in one embodiment; Figure 8 This is a schematic diagram illustrating the glacier extraction effect in one embodiment; Figure 9 This is a schematic diagram illustrating the performance evaluation of the model using confusion matrix, OA, F1, and MIoU in one embodiment. Figure 10 This is a schematic diagram illustrating the evaluation of the model's glacier identification accuracy using confusion matrix, OA, F1, and MIoU in one embodiment. Figure 11 This is a schematic diagram illustrating the evaluation of the model's accuracy in identifying ice lakes using confusion matrix, OA, F1, and MIoU in one embodiment. Figure 12 This is a schematic diagram illustrating the effect of domain adaptation on improving the extraction accuracy of glaciers and glacial lakes in one embodiment; Figure 13 This is a structural block diagram of an automatic glacier and glacial lake identification device based on multi-source remote sensing and domain adaptation in one embodiment; Figure 14 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0018] 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.

[0019] In one embodiment, such as Figure 1 As shown, an automatic identification method for glaciers and glacial lakes based on multi-source remote sensing and domain adaptation is provided, including the following steps: S101: Acquire remote sensing image data and perform preprocessing.

[0020] In this embodiment of the application, firstly, remote sensing image data is acquired, such as... Figure 2 As shown, remote sensing satellite imagery, Sentinel-2 MSI and Sentinel-1 GRD imagery data packages, and TanDEM-X radar elevation data were downloaded using the Google Satellite Engine cloud platform (GEE). The acquired data underwent spatiotemporal filtering, cloud cover filtering, band synthesis, and dimensionless operations. Sliding window data slicing was also performed, and the cataloged data underwent visual correction and rasterization. Remote sensing imagery served as samples, and cataloged data served as labels. The data sources used are shown in Table 1.

[0021] Table 1 Data Sources

[0022] Specifically, in one embodiment of this application, the preprocessing includes: Spatiotemporal filtering, cloud cover filtering, band synthesis, and unified dimension operations are performed on the remote sensing image data. The remote sensing image data is normalized and sliced ​​using a sliding window.

[0023] In one embodiment of this application, the Google Satellite Engine (GEE) platform was used to download Sentinel-2 MSI and Sentinel-1 SAR GRD images. First, time filtering (time periods set to 2022.07-2022.09 and 2023.07-2203.09) was performed, followed by spatial filtering (based on the study area), cloud cover filtering (<5%), band selection, multi-image fusion, and cropping to obtain complete image data within the study area. All bands were resampled to 10 m spatial resolution, exported, and downloaded locally. The dimensions of all band data were unified to 0-255.

[0024] A multi-dimensional feature input tensor was constructed from remote sensing imagery, including 11 channels of data: red band, green band, blue band, near-infrared (NIR), shortwave infrared (SWIR), normalized difference vegetation index (NDVI), normalized difference water index (NDWI), snow index (NDSI), digital elevation model (DEM), slope map, aspect map, and VV polarimetric radar imagery. To avoid the influence of data units, all bands were normalized to 0-255. Figure 3 As shown, to reduce the amount of data per training sample, a scanning method with a step size of 201 pixels was used to crop the remote sensing image into several 224×224×11 image blocks, and 10% lateral overlap and vertical overlap were retained to avoid model misjudgment caused by insufficient feature information around edge pixels. The constructed sample database is shown in Tables 2 and 3.

[0025] Table 2 Glacier Sample Database

[0026] Table 3. Glacier Lake Sample Database

[0027] S103: Construct a deep learning model for classifying glaciers and lakes. The model includes an encoder and a decoder. The encoder consists of convolutional layers, a Swing Transformer module, and a channel attention mechanism. The decoder uses transposed convolution and multi-scale skip connections to fuse feature maps.

[0028] In this embodiment, an improved UNet semantic segmentation model, Trans-CAUnet, is constructed as a deep learning model for glacier and lake classification, with the following structure: Figure 4As shown, the model comprises an encoder and a decoder. The encoder consists of convolutional layers, a Swing Transformer module, and a channel attention mechanism. Convolutions are used to construct feature blocks, generating patch blocks with a pathsize of 2. This setup accelerates the model's computation while minimizing the loss of feature map information during encoding. The patch blocks are then input into the Swing Transformer module to extract local and global contextual information, and the channel attention mechanism is fused to enhance the modeling ability for spectral features. The decoder uses transposed convolutions and multi-scale skip connections to fuse feature maps, achieving image size restoration and detail enhancement. The model output is a binary classification mask for glaciers / glacial lakes with the same size as the input image.

[0029] Specifically, in one embodiment of this application, the Swing Transformer module includes a sliding window multi-head self-attention mechanism and a window self-attention mechanism, which calculate multi-head attention both inside and between windows to obtain global and local semantic information of the feature image.

[0030] In one embodiment of this application, such as Figure 4 As shown, the sliding window multi-head self-attention mechanism (SW-MSA) and the window self-attention mechanism (W-MSA) replace the global multi-head attention mechanism in ViT. They calculate multi-head attention both inside and between windows, and obtain global and local semantic information of the feature image. This solves the problems of high computational cost of Transformer and network susceptibility to influence from outside the region of interest.

[0031] In W-MAS, the feature map is first divided into multiple fixed-size windows, and then a multi-head attention mechanism is computed within each window to obtain the positional relationship information of each pixel in the window. Figure 4 (c) Each window incorporates local information from the feature map.

[0032] In SW-MSA ( Figure 4 (c) First, the rows and columns of the feature map are slid up and to the left as a whole, dividing the shifted feature map into windows of fixed size. In each window, the multi-head attention mechanism of each window is calculated using a mask matrix. In this way, each window integrates the information between multiple windows, realizing the connection of information between feature map windows and the understanding of contextual semantics. Then, the feature map is translated back to its original position, each window is restored, and a new feature map is obtained.

[0033] As shown in the formula below, one of the key steps in the above two window multi-head self-attention methods is to calculate the query (Q), key (K), and value (V) matrices. The Q vector of an element is multiplied by the K vectors of all other elements to measure the correlation between this element and other elements in the sequence.

[0034]

[0035]

[0036] Where X represents the input data, W Q W K W V It is a learnable weight matrix, where Q, K, and V are the generated query, key, and value matrices, respectively, and d k This represents the dimension of the key vector.

[0037] Specifically, in one embodiment of this application, the channel attention mechanism includes an average pooling layer, a max pooling layer, a multilayer perceptron, and an activation function.

[0038] In one embodiment of this application, the Channel Attention Mechanism (CAM) obtains the average and maximum values ​​for each channel by performing global average pooling and global max pooling on the spatial features. These averages and maximum values ​​are then passed through a multilayer perceptron, and the outputs of the multilayer perceptron are fused and processed by an activation function to generate channel attention weights. Finally, these weights are used to weight the features of each channel to enhance important features and suppress unimportant features, thereby improving the convolutional neural network's ability to capture important channel features. In remote sensing technology, this translates to improved model perception of spectral information of ground objects.

[0039] S105: The preprocessed remote sensing image data is used to train the deep learning model for glacier and lake classification, and the trained deep learning model for glacier and lake classification is transferred to the target domain for domain adaptation adjustment.

[0040] In this embodiment, preprocessed remote sensing image data is input into a deep learning model for glacier and glacial lake classification for training. The Adam optimizer is used, with a learning rate of 0.01 for the first 15 iterations to accelerate convergence, and then set to 0.001 after 15 iterations to achieve stable convergence. The batch size is 16, and the loss function uses a combination of cross-entropy and Dice loss. Training is conducted for 50 epochs, with an input size of 224×224 pixels. After learning from all samples in the training region, a final model weight file is obtained, which is then used for prediction in the test region. Next, the trained glacier and glacial lake classification deep learning model is transferred to the target domain for domain adaptation. First, the model weights trained in the source domain are loaded, and then further adjustments to the model weight parameters are made in the target domain. In specific applications, model training and testing were first conducted on a total sample database covering the Qilian Mountains, West Kunlun Mountains, and Tianshan Mountains to obtain a weight parameter file. Glaciers and glacial lakes in southeastern Tibet were selected as the application area for direct transfer. Domain adaptation is a technique used to address the differences between the source and target domains. In the glacier identification example, the source domain is glacier and glacial lake data from one region, while the target domain is glacier and glacial lake data from another region. There is a domain shift between the source and target domains, meaning that the glacier and glacial lake features, geographical environment, or visual appearance differ significantly between the two regions. Therefore, directly applying a model from the source domain may lead to performance degradation. Domain adaptation aims to adjust the existing model weights through relearning in unknown regions, enabling it to adapt to the data differences in the target domain, thereby further improving the accuracy of the model's extraction results in the target domain.

[0041] S107: Input the remote sensing image to be predicted into the domain-adapted deep learning model for glacier and lake classification, output the glacier and lake extraction results, perform restoration and noise reduction processing on the glacier and lake extraction results, and obtain the glacier and lake identification results.

[0042] In this embodiment, the remote sensing image to be predicted is input into a domain-adapted deep learning model for glacier and glacier lake classification, and the output is a glacier / glacier lake extraction result. This result is a binary classification mask of glaciers / glaciers of the same size as the input image. Since deep learning training requires significant GPU memory and computing power, the image was previously segmented into numerous 224×224×11 images according to row and column numbers to meet the requirements of model training and prediction. The prediction results from small-scale image slices cannot describe the state of the entire large-scale glacier and glacier lake. Therefore, the glacier and glacier lake extraction results are subjected to corresponding merging, restoration, and noise reduction processing to obtain the glacier and glacier lake identification result.

[0043] Specifically, in one embodiment of this application, the noise reduction processing of the glacier-lake extraction results includes: S201: Image stitching and coordinate system restoration of the glacier and glacial lake extraction results.

[0044] S203: Remove noise from the restored image, fill in the internal holes in the extracted results, and remove non-glacial lake water.

[0045] In one embodiment of this application, data is sliced ​​and merged, and coordinate system restoration is performed. Image stitching and coordinate system restoration are reverse engineering of image sliding window slicing, implemented using the gdal library in Python. During image slicing and prediction, the geographic transformation information of the image is preserved, allowing direct calculation of the affine transformation matrix and offset of the input image for image stitching and coordinate system restoration based on this information. Because ground features share similar spectral and textural characteristics, discrete and independent pixels are generated in the classification result image. To address this issue, a method combining morphological processing and Gaussian blur is employed, implemented using the OpenCV library in Python.

[0046] Furthermore, surface moraines often exist on glaciers. In remote sensing imagery, surface moraines typically exhibit characteristics completely different from glaciers, leading to misclassification by models as non-glacial areas and resulting in holes within the perceived glacier. Similarly, the presence of ice fragments or other floating debris on the surface of glacial lakes can also cause models to classify them as non-glacial lake areas, creating holes within the lake. To repair these holes, a hole-filling technique is employed to enhance the accuracy of remote sensing image interpretation. This method is as follows: Figure 5 As shown, this implementation uses the NumPy library in Python. First, the predicted image is binary-inverted, transforming the glacier region (originally a pixel value of 1) into the background (originally a pixel value of 0), and vice versa. Then, connected component analysis is used, employing 8-connectivity (connecting a pixel to its eight adjacent directions: top, bottom, left, right, and four diagonal directions). This identifies independent regions in the inverted image, labels each region, and calculates its area. If a region's area is less than a set threshold, it is considered a hole and its value is set to 0. Finally, the image is binary-inverted again to ensure the glacier region remains intact while removing internal holes.

[0047] In addition, by integrating high-precision river system data and verifying and correcting the glacial lake extraction results through manual visual interpretation, water bodies unrelated to glaciers were effectively eliminated, ensuring the reliability of the glacial lake extraction results and providing a reliable data foundation for hydrological research in glacial areas.

[0048] In one embodiment of this application, the image stitching and coordinate system restoration of the glacier-lake extraction results includes: S301: Determine the total geographic range of the stitched area and the image geographic transformation matrix based on the geographic transformation information extracted from glaciers and glacial lakes.

[0049] S303: Image stitching is performed based on the total geographic range of the stitching area and the image geographic transformation matrix, and the overlapping area data is fused using a weighted average method.

[0050] S305: Restore the projection, coordinate system, and spatial reference information of the image stitching result.

[0051] In one embodiment of this application, such as Figure 6 As shown, firstly, the geographic transformation information of all slice images is iterated to find the maximum and minimum values ​​of the horizontal and vertical coordinates, thus determining the total geographic range of the stitching area. Then, the relative position of each slice image within this total geographic range is calculated. Next, the geographic transformation matrix of each image is calculated in reverse to ensure that its pixel data is accurately aligned to the stitching area. The images are then converted into arrays and written into the stitching area. During the stitching process, to ensure that no stitching lines are created at the image edges, the data in overlapping areas need to be fused using a weighted average method. Finally, the output image is reassigned to the projection and coordinate system of the overall stitching area, restoring its spatial reference information and thus achieving coordinate system restoration. This ensures that the stitched image is spatially consistent with the original data, making it suitable for subsequent analysis and applications.

[0052] In one embodiment of this application, the noise removal of the restored image includes: S401: Employs morphology-based erosion and dilation operations to remove discrete noise points and restore the overall outline of ground features.

[0053] S403: Use Gaussian blur to smooth the overall contour boundary.

[0054] In one embodiment of this application, two basic morphological operations, erosion and dilation, are first employed to initially remove discrete noise points from the image and restore the integrity of the overall contour of the terrain features. The main function of erosion is to weaken or remove isolated small regions or noise points, especially those randomly misdetected regions caused by model errors or image noise. Through erosion, noise points at the boundaries of glaciers and glacial lakes are weakened, and the remaining areas more closely resemble the actual glacier distribution. However, erosion may cause overall shrinkage of the glacier and glacial lake regions. To compensate for this problem, dilation is used after erosion to expand the glacier region back to its original shape, restoring boundary connectivity and maintaining overall shape characteristics. By expanding the neighborhood range of each glacier and glacial lake pixel, the segmentation result is smoother, eliminating the excessive shrinkage caused by erosion, thereby further improving the connectivity and integrity of the glacier region. Subsequently, Gaussian blurring is used to further smooth the overall contour boundaries of glaciers and glacial lakes and eliminate boundary sharpening problems caused by pixel-level irregularities. The basic idea of ​​this technique is to perform a weighted average of each pixel in the image within its neighborhood according to a Gaussian distribution, thereby reducing noise and smoothing details in the image. Figure 7 As shown, through Gaussian blurring, the glacier and glacial lake boundaries in the identification results are transformed from their original jagged or overly sharp shapes into more natural and smooth shapes, thus better reflecting the actual glacier and glacial lake boundaries.

[0055] In one embodiment of this application, such as Figure 8 The image shown is a rendering of the glacier extraction results, where (a) is the Qilian Mountains, (g) is the West Kunlun Mountains, and (m) is the Tianshan Mountains. UNet(b,h,n), Swin-UNet(c,i,o), Deeplabv3+(d,j,p), TransUNet(e,k,q), and Trans-CAUNet(f,l,r) are used. Figure 9 As shown, the model's performance was evaluated using confusion matrix, OA, F1, and MIoU. The model was tested on test datasets from the Qilian Mountains, West Kunlun Mountains, and Tianshan Mountains. The results show that the Trans-CAUnet model outperforms Swin-UNet, UNet, Deeplabv3+, and TransUNet models. Figure 10 As shown, the highest accuracy of the evaluation indices OA, F1, and MIoU in the total glacier sample database is 0.9685, 0.9619, and 0.9270, respectively. Figure 11As shown, the highest accuracy of the evaluation indicators OA, F1, and MIoU on the total ice lake sample database is 0.9995, 0.9509, and 0.910, respectively. Trans-CAUnet has a stronger perception of the spectral features of ground objects in the image, more complete overall outlines of ground objects, more refined edge details, fewer fragmented patches, and better elimination of interference from similar ground objects.

[0056] In one embodiment of this application, after field applicability, such as Figure 12 As shown, in the glacier recognition results, the Trans-CAUNet model achieved accuracies of 0.96, 0.85, and 0.91 in OA, F1, and MIoU, respectively, representing improvements of 4%, 3%, and 11% compared to direct transfer applications. In the glacial lake recognition results, the Trans-CAUNet model achieved accuracies of 0.99, 0.95, and 0.90 in OA, F1, and MIoU, respectively, representing improvements of 0.5%, 3%, and 3% compared to direct transfer applications. The experiments demonstrate that even in different regions, the model maintains a high recognition accuracy, although the accuracy may be slightly lower than in the source region. In this case, domain adaptation is an effective solution.

[0057] The aforementioned automatic glacier and glacial lake identification method based on multi-source remote sensing and domain adaptation first acquires and preprocesses remote sensing image data. Then, a deep learning model for glacier and glacial lake classification is constructed, comprising an encoder and a decoder. The encoder consists of convolutional layers, a Swing Transformer module, and a channel attention mechanism, while the decoder uses transposed convolution and multi-scale skip connections to fuse feature maps. Next, the preprocessed remote sensing image data is used to train the deep learning model, which is then adapted to the target domain. Finally, the remote sensing image to be predicted is input into the domain-adapted deep learning model, which outputs the extracted glacier and glacial lake data. This extracted data is then subjected to denoising and restoration processing to obtain the glacier and glacial lake identification result. In other words, by constructing a novel deep learning model, effective extraction of glacier and glacial lake features from remote sensing images is achieved. Combined with domain adaptation technology, this improves the model's generalization ability and recognition accuracy under cross-regional and multi-source data conditions.

[0058] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0059] Based on the same inventive concept, this application also provides an automatic glacier and glacial lake identification device based on multi-source remote sensing and domain adaptation for implementing the aforementioned automatic glacier and glacial lake identification method based on multi-source remote sensing and domain adaptation. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the automatic glacier and glacial lake identification device based on multi-source remote sensing and domain adaptation provided below can be found in the limitations of the automatic glacier and glacial lake identification method based on multi-source remote sensing and domain adaptation described above, and will not be repeated here.

[0060] In one embodiment, such as Figure 13 As shown, an automatic glacier and glacial lake identification device 1300 based on multi-source remote sensing and domain adaptation is provided, including: a data acquisition and preprocessing module 1301, a model building module 1303, a model training and transfer module 1305, and a glacier and glacial lake identification module 1307, wherein: The data acquisition and preprocessing module 1301 is used to acquire remote sensing image data and perform preprocessing.

[0061] Model building module 1303 is used to build a deep learning model for glacier and lake classification. The model includes an encoder and a decoder. The encoder consists of convolutional layers, a Swing Transformer module and a channel attention mechanism. The decoder uses transposed convolution and multi-scale skip connections to fuse feature maps.

[0062] The model training and transfer module 1305 is used to train the deep learning model for glacier and lake classification using preprocessed remote sensing image data, and to transfer the trained deep learning model for glacier and lake classification to the target domain for domain adaptation.

[0063] The glacier and glacier lake identification module 1307 is used to input the remote sensing image to be predicted into the domain-adapted deep learning model for glacier and glacier lake classification, output the glacier and glacier lake extraction results, and perform restoration and noise reduction processing on the glacier and glacier lake extraction results to obtain the glacier and glacier lake identification results.

[0064] In one embodiment of this application, the data acquisition and preprocessing module is further configured to: The remote sensing image data is subjected to radiometric calibration, atmospheric correction, geometric correction, and resampling operations. The remote sensing image data is normalized and sliced ​​using a sliding window.

[0065] In one embodiment of this application, the Swin Transformer module includes a sliding window multi-head self-attention mechanism and a window self-attention mechanism, which calculate multi-head attention both inside and between windows to obtain global and local semantic information of the feature image.

[0066] In one embodiment of this application, the channel attention mechanism includes an average pooling layer, a max pooling layer, a multilayer perceptron, and an activation function.

[0067] In one embodiment of this application, the glacier-lake identification module is further used for: Image stitching and coordinate system restoration were performed on the extracted glacier and glacial lake results; The restored image is then subjected to noise removal, internal holes in the extracted results are filled, and non-glacial lake water is removed.

[0068] In one embodiment of this application, the glacier-lake identification module is further used for: The total geographic extent of the stitched area and the image geographic transformation matrix are determined based on the geographic transformation information extracted from glaciers and glacial lakes. Image stitching is performed based on the total geographic range of the stitching area and the image geographic transformation matrix, and the overlapping area data is fused using a weighted average method. Restore the projection, coordinate system, and spatial reference information of the image stitching result.

[0069] In one embodiment of this application, the glacier-lake identification module is further configured to: Morphology-based erosion and dilation operations are used to remove discrete noise points and restore the overall outline of ground features; The overall contour boundary is smoothed using the Gaussian blur method.

[0070] The modules in the aforementioned automatic glacier and glacial lake identification device based on multi-source remote sensing and domain adaptation can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0071] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 14 As shown, the computer device includes a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements an automatic glacier and glacial lake identification method based on multi-source remote sensing and domain adaptation. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0072] Those skilled in the art will understand that Figure 14 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0073] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0074] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0075] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0076] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0077] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0078] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0079] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for automatic identification of glacier and lake based on multi-source remote sensing and domain adaptation, characterized in that, The method comprises: acquiring remote sensing image data and performing preprocessing; constructing a glacier and ice lake classification deep learning model, the model comprising an encoder part and a decoder part, wherein the encoder part is composed of a convolutional layer, a Swin Transformer module and a channel attention mechanism, and the decoder part adopts a transposed convolution and a multi-scale skip connection to fuse feature maps; training the glacier and ice lake classification deep learning model using the remote sensing image data after preprocessing, and migrating the trained glacier and ice lake classification deep learning model to a target field for domain adaptation adjustment; inputting remote sensing images to be predicted into the glacier and ice lake classification deep learning model after domain adaptation adjustment, outputting glacier and ice lake extraction results, and performing restoration and noise reduction processing on the glacier and ice lake extraction results to obtain glacier and ice lake recognition results.

2. The method of claim 1, wherein the method is characterized by, The preprocessing comprises: performing radiation calibration, atmospheric correction, geometric correction and resampling operations on the remote sensing image data; performing normalization and sliding window data slicing on the remote sensing image data.

3. The method of claim 1, wherein the method is characterized by, The Swin Transformer module comprises a sliding window multi-head self-attention mechanism and a window self-attention mechanism, which respectively calculate multi-head attention within and between windows to obtain global and local semantic information of feature images.

4. The method of claim 1, wherein the method is characterized by, The channel attention mechanism comprises an average pooling layer, a maximum pooling layer, a multi-layer perceptron and an activation function.

5. The method of claim 1, wherein the method is based on multi-source remote sensing and domain adaptation. The restoration and noise reduction processing on the glacier and ice lake extraction results comprises: performing image stitching and coordinate system restoration on the glacier and ice lake extraction results; performing noise removal on the restored images, filling internal holes of the extraction results, and removing non-ice lake water bodies.

6. The method of claim 5, wherein the method is characterized by, The image stitching and coordinate system restoration on the glacier and ice lake extraction results comprise: determining the total geographical range of the stitching area and the image geographical transformation matrix based on the geographical transformation information of the glacier and ice lake extraction results; performing image stitching based on the total geographical range of the stitching area and the image geographical transformation matrix, and fusing the overlapping area data using a weighted average method; restoring the projection, coordinate system and spatial reference information of the image stitching result.

7. The method of claim 5, wherein the method further comprises: The noise removal on the restored images comprises: removing discrete noise points using morphological erosion and dilation operations to restore the overall outline of the ground features; smoothing the overall outline boundary using a Gaussian blur method.

8. An automatic glacier-lake identification device based on multi-source remote sensing and domain adaptation, characterized in that, The device comprises: a data acquisition and preprocessing module for acquiring remote sensing image data and performing preprocessing; a model construction module for constructing a glacier and ice lake classification deep learning model, the model comprising an encoder part and a decoder part, wherein the encoder part is composed of a convolutional layer, a Swin Transformer module and a channel attention mechanism, and the decoder part adopts a transposed convolution and a multi-scale skip connection to fuse feature maps; a model training and migration module for training the glacier and ice lake classification deep learning model using the remote sensing image data after preprocessing, and migrating the trained glacier and ice lake classification deep learning model to a target field for domain adaptation adjustment; The glacier lake identification module is configured to input the remote sensing image to be predicted into the glacier lake classification deep learning model adjusted according to the field, and output a glacier lake extraction result. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The computer program, when executed by the processor, implements the steps of the method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by the processor, implements the steps of the method of any one of claims 1 to 7.

Images