Ice lake intelligent identification method and system based on multi-source data fusion, and electronic equipment

Abstract

The invention provides a multi-source data fusion ice lake intelligent identification method and system and electronic equipment, and relates to the technical field of terrain monitoring. The method comprises the following steps: acquiring Sentinel-2 data, gradient data extracted by an ALOS-DEM and NDWI data, performing preprocessing, dividing the data into a training set and a test set, taking ResNet-50 as a backbone network, introducing a CBAM module, constructing a Mask R-CNN model, performing target detection and segmentation on a glacial lake region, performing model training by adopting a multi-task joint loss function, performing preprocessing on to-be-detected data, and then performing target detection and segmentation on the to-be-detected data. And inputting the optimal Mask R-CNN model, and finally obtaining a recognition result of the ice lake. The gradient data and the NDWI generated by combining the high-resolution ALOS-DEM elevation data effectively reduce the influence and misjudgment of the water body turbidity and the mountain shadow on the ice lake recognition, significantly improve the accuracy of the ice lake recognition, realize the automatic and efficient recognition of the ice lake area of the complex plateau terrain area, greatly improve the efficiency of the ice lake recognition, and reduce the cost of the ice lake recognition. The method is of great significance in predicting and preventing disasters such as glacier collapse and glacial lake outburst.

Description

Technical Field

[0001] The present invention relates to the field of terrain monitoring technology, and in particular to a multi-source data fusion glacial lake intelligent identification method, system and electronic equipment. Background Art

[0002] Glacial lakes in the plateau region are of great significance to climate change and environmental evolution. Accurate monitoring of these lakes is crucial for climate change assessment and water resource management. However, the harsh climate and inaccessible transportation in the plateau make traditional field surveys challenging. Satellite remote sensing technology has become a key tool for glacial lake monitoring. Current research on glacial lake outburst hazards indicates that the formation and evolution of small glacial lakes (<0.05 km²) are important precursors to glacial disasters. Timely monitoring and analysis of changes in small glacial lakes can help predict and prevent glacier collapse, glacial lake outbursts, and other disasters. However, identifying glacial lakes faces multiple challenges, including their small size and dispersed distribution, which makes them difficult to capture using traditional remote sensing methods. 1) When glacial lakes are connected to debris-covered glaciers, glacial debris and turbid water often blend with the lake's background. 2) During the melt season, the color of glacial lakes is significantly affected by factors such as dissolved organic matter and suspended particulate matter in the water, as well as weather, seasonal lighting conditions, and other factors. Shadows and reflections can alter the lake's characteristics, making identification more difficult. 3) Due to the influence of lighting conditions, the shadow areas on the shady side of the mountain have smooth texture and uniform color. They are also difficult to distinguish from glacial lakes in multispectral images. Even in the manual digitization process, it is difficult to delineate the boundaries of the lakes.

[0003] Existing glacial lake identification methods based on remote sensing images have defects. The single spectral band threshold method is easily affected by other surface features, such as mountain shadows, rivers, dissolved organic matter and suspended particulate matter in water bodies, resulting in low identification accuracy and inability to meet actual needs, and urgently needs to be improved. Summary of the Invention

[0004] The main purpose of the present invention is to provide a multi-source data fusion glacial lake intelligent identification method, system and electronic equipment to solve the technical problems in the prior art that glacial lake monitoring is susceptible to interference, lake boundary demarcation is not accurate enough, glacial lake feature identification is difficult and the identification accuracy is not high.

[0005] To solve the above technical problems, the present invention adopts a technical solution: a multi-source data fusion ice lake intelligent identification method, comprising the following steps: S1: Acquire Sentinel-2 data, slope data extracted from ALOS-DEM, and NDWI data, perform data preprocessing, and then synthesize the bands to form a multi-channel dataset containing red, green, blue, near-infrared, NDWI, and slope data; S2: After the multi-channel dataset undergoes the second preprocessing, it is divided into a training dataset and a test dataset according to a preset ratio; S3: Build a Mask R-CNN model based on ResNet-50. Introduce the CBAM module into the high-level layers of the backbone network (Conv4 / Conv5), each FPN layer, and the detection head. After inputting the training dataset for multi-task joint loss function training, save the optimal Mask R-CNN model. S4: Automatic Identification of Ice Lakes: After preprocessing the test data, input it into the optimal Mask R-CNN model to obtain the ice lake identification results.

[0006] The preferred solution also includes: Evaluation and optimization: The secondary evaluation indicators precision, recall rate and accuracy are used as the accuracy evaluation of the Mask R-CNN model. The formula is: ; ; ; Among them, TP is true positive, indicating the accurately described glacial lake area; FP is false positive, defined as the non-glacial lake area detected by the applied method as the glacial lake area in the image; FN is false negative, indicating the real glacial lake slope area that was not detected; TN is true negative, indicating the detected non-glacial lake area.

[0007] In a preferred embodiment, the data preprocessing includes: After atmospheric correction, Sentinel-2 data were converted to ENVI format; ALOS-DEM data were resampled to the preset resolution, and slope data were extracted using the slope tool; The calculation formula for NDWI data is: ; Where B3 is the Green band data and B8 is the NIR band data.

[0008] In the preferred solution, Sentinel-2 data is high-resolution remote sensing imagery with clear, moderate contrast, and rich color gradation, and cloud cover of less than 20%. Specifically: A pre-trained U-Net cloud detection model was used to retain images with cloud coverage less than 20%; For clarity evaluation, the average gradient (AG) is used to reflect edge sharpness: Standard deviation (STD) + dynamic range (DR) is used for contrast evaluation; The HSV spatial saturation mean is used to evaluate color richness.

[0009] In the preferred solution, introducing the CBAM module into the high-level features of the backbone network includes: introducing the CBAM module into the Conv4 and Conv5 layers; The CBAM module combines channel attention and spatial attention mechanisms, where: The channel attention module performs global average pooling and maximum pooling on the extracted features to generate channel weights, and applies these weights to the feature map to enhance the highly relevant features of the ice lake target; The spatial attention module performs spatial attention operations on the feature map, weights local features through convolution kernels, highlights the spatial location information of the ice lake, and suppresses background interference.

[0010] In the preferred solution, in step S2, during model training, a multi-task joint loss function training strategy is adopted, comprehensively considering the target classification loss , bounding box regression loss and segmentation mask loss , the formula is: ; Target classification loss Cross Entropy Loss is used to measure the difference between the predicted category and the actual category. The formula is: ; Where: is the true category, The probability predicted by the model; Bounding box regression loss A smooth L1 loss function is used to optimize the gap between the predicted bounding box position and the true bounding box. The formula is: ; Where: is the true bounding box position, To predict the bounding box location, the smooth L1 loss is defined as: ; Segmentation mask loss Use the binary cross entropy loss for each pixel and combine it with the attention weight brought by the CBAM module to improve the segmentation accuracy. The loss on each pixel is calculated as follows: ; Where: is the true mask pixel label, is the predicted mask pixel, is the output of the Sigmoid function applied to each pixel, that is, the predicted probability.

[0011] In the preferred solution, the loss weight is dynamically adjusted based on the performance of each task of the multi-task loss function, namely classification, detection and segmentation, on the validation set. The task weight is inversely proportional to the error. The higher the error, the greater the weight. The formula is: ; Where, is the classification error rate, for Detection error, is the segmentation error; The weighted total loss expression is: ; Where, and are the weight coefficients for classification, detection and segmentation respectively, and are the loss functions for classification, detection and segmentation respectively.

[0012] In a preferred embodiment, the second preprocessing in step S2 includes: Considering the area, elevation, color and shape characteristics, multiple images are color-homogenized, and multi-channel datasets are labeled and processed by slicing, random cropping, rotation and normalization.

[0013] A multi-source data fusion intelligent identification system for glacial lakes in complex plateau terrain areas, including: The data acquisition module is used to acquire Sentinel-2 data, slope data extracted from ALOS-DEM, and NDWI data. After data preprocessing, it synthesizes the bands to form a multi-channel dataset containing red, green, blue, near-infrared, NDWI, and slope data. A data partitioning module is used to partition the multi-channel data set into a training data set and a test data set according to a preset ratio after the second preprocessing; The model building module is used to build the Mask R-CNN model based on ResNet-50. The CBAM module is introduced in the high-level backbone network (Conv4 / Conv5), each FPN layer, and the detection head. After inputting the training dataset and performing multi-task joint loss function training, the optimal Mask R-CNN model is saved. The recognition module is used for automatic ice lake identification: After data preprocessing, the test data is input into the optimal Mask R-CNN model to obtain ice lake identification results.

[0014] An electronic device comprising a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the ice lake intelligent identification method based on multi-source data fusion as described in any one of claims 1 to 8 when executing the computer program.

[0015] The present invention provides a multi-source data fusion method, system, and electronic device for intelligent glacial lake identification. This method acquires Sentinel-2 data, slope data extracted from ALOS-DEM, and NDWI data, preprocesses them into training and test sets, and uses a ResNet-50 backbone network with the CBAM module to construct a Mask R-CNN model for target detection and segmentation in glacial lake areas. This model is trained using a multi-task joint loss function. After preprocessing the test data, the optimal Mask R-CNN model is input to ultimately obtain glacial lake identification results. By combining slope data and NDWI generated from high-resolution ALOS-DEM elevation data, the system effectively reduces the impact of water turbidity and mountain shadows on glacial lake identification and prevents misjudgment. This significantly improves the accuracy of glacial lake identification, enabling automated and efficient identification of glacial lake areas in complex plateau terrain. This significantly increases the efficiency of glacial lake identification, which is of great significance for predicting and preventing disasters such as glacier collapse and glacial lake outbursts. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The present invention will be further described below with reference to the accompanying drawings and examples: Figure 1 It is a flow chart of the identification method of the present invention; Figure 2 Schematic diagram of the Mask R-CNN model structure of the present invention; Figure 3 is a graph of the loss function of the model training of the present invention; Figure 4 This is a structural diagram of the automatic identification of ice lakes according to the present invention. DETAILED DESCRIPTION

[0017] Example 1 like Figure 1-4 As shown in FIG, a multi-source data fusion ice lake intelligent identification method includes the following steps: S1: Acquire Sentinel-2 data, slope data extracted from ALOS-DEM, and Normalized Difference Water Index (NDWI) data, perform data preprocessing, and then form a multi-channel dataset through band synthesis. Sentinel-2 data includes four bands of high-resolution remote sensing imagery: red, green, blue, and near-infrared. The multi-channel dataset is a six-channel dataset containing red, green, blue, near-infrared, NDWI, and slope data.

[0018] S2: After the second preprocessing, the multi-channel dataset is divided into a training dataset and a test dataset according to a preset ratio.

[0019] The second preprocessing is to label the multi-channel dataset considering area, altitude, color and shape features, slice it according to the preset pixel size, randomly crop it, rotate it 90°, and then normalize it.

[0020] S3: Build a Mask R-CNN model based on ResNet-50. After inputting the training dataset for model training, save the optimal Mask R-CNN model. The Mask R-CNN model includes the backbone network, region proposal network, ROI alignment layer, and detection head network. The CBAM module is introduced into the high-level features of the backbone network, the feature maps of each layer of FPN, and the detection head.

[0021] Furthermore, it also includes evaluation and optimization: the secondary evaluation indicators precision, recall rate and accuracy are used as the accuracy evaluation of the Mask R-CNN model. The formula is: ; ; ; Among them, TP is true positive, indicating the accurately described glacial lake area; FP is false positive, defined as the non-glacial lake area detected by the applied method as the glacial lake area in the image; FN is false negative, indicating the real glacial lake slope area that was not detected; TN is true negative, indicating the detected non-glacial lake area.

[0022] S4: Ice Lake automatic recognition: After data preprocessing, the test data is input into the optimal Mask R-CNN model to obtain the recognition results.

[0023] This example obtains Sentinel-2 data, slope data extracted from ALOS-DEM, and NDWI data, preprocesses them into training and test sets, and then introduces the CBAM module based on ResNet-50 as the backbone network to construct a Mask R-CNN model for target detection and segmentation in glacial lake areas. A multi-task joint loss function is used for model training. After preprocessing the test data, the optimal Mask R-CNN model is input to ultimately obtain glacial lake identification results. Combining slope data and NDWI generated from high-resolution ALOS-DEM elevation data effectively reduces the impact of water turbidity and mountain shadows on glacial lake identification and can significantly improve the accuracy of glacial lake identification. This enables automated and efficient identification of glacial lake areas in complex plateau terrain, greatly improving the efficiency of glacial lake identification and having significant implications for predicting and preventing disasters such as glacier collapse and glacial lake outbursts.

[0024] This example fully considers the characteristics of glacial lakes in complex plateau terrain. It utilizes the red, green, blue, and near-infrared bands of Sentinel-2 high-resolution remote sensing imagery, integrates slope data extracted from the ALOS-DEM, and NDWI data to form a multi-band dataset. Band synthesis creates a six-channel dataset. Based on this, the Mask R-CNN deep learning model is used to manually label glacial lake learning samples. Sample optimization and processing are then performed to create a sample dataset suitable for the Mask R-CNN deep learning model. The deep learning model is then trained using the sample dataset to obtain an automatic glacial lake recognition model based on the Mask R-CNN deep learning model, and the accuracy of this model is evaluated.

[0025] In step S1, data collection involves downloading high-resolution Sentinel-2 remote sensing imagery for the identified area using the Copernicus Explorer (https: / / dataspace.copernicus.eu / ). High-quality imagery should include cloud cover less than 20%, clarity, moderate contrast, and rich tonal (color) gradations. ALOS-DEM data (https: / / search.asf.alaska.edu / ) should also be obtained for the identified area. Field survey data should include the distribution of some of the glacial lakes surveyed.

[0026] In the preferred scheme, Sentinel-2 data should be high-resolution remote sensing images with clear, moderate contrast, rich color gradation, and cloud cover of less than 20%. Specifically: A pre-trained U-Net cloud detection model was used to retain images with cloud coverage less than 20%; For clarity evaluation, the average gradient (AG) is used to reflect edge sharpness: Standard deviation (STD) + dynamic range (DR) is used for contrast evaluation; The HSV spatial saturation mean is used to evaluate color richness.

[0027] In this embodiment, cloud cover detection is performed on Sentinel-2 images using a pre-trained U-Net cloud detection model, effectively removing most of the images affected by cloud cover, improving data availability and enabling the image data to more realistically reflect surface information.

[0028] The average gradient is used to measure the clarity of the edges in the image: when the average gradient value is high, it means that the image has rich edge information, clear details, and clear boundaries between objects, and different object categories can be accurately identified and distinguished.

[0029] Contrast assessment using standard deviation (STD) + dynamic range (DR): helps enhance the visual quality of images and the distinguishability of ground objects.

[0030] The saturation mean in HSV space is used to evaluate color richness, which improves the accuracy of object classification and recognition based on color information.

[0031] In the preferred embodiment, data preprocessing includes: After atmospheric correction, Sentinel-2 data were converted to ENVI format; ALOS-DEM data were resampled to the preset resolution, and slope data were extracted using the slope tool; The calculation formula for NDWI data is: ; Where, B3 is the Green band data, and B8 is the NIR band data; In this example, atmospheric correction is performed on the Level-1C data product based on the acquired Sentinel-2 data to produce Level-2A data. This process can be performed using the Sen2Cor tool by inputting L2A_Process.bat followed by the "L1C file location."

[0032] The four bands, B2 (Blue), B3 (Green), B4 (Red), and B8 (NIR), were then converted to ENVI format using the Data Conversion function in SNAP (Sentinel Application Platform) software. NDWI was then calculated using the Raster Calculator tool in ArcGIS Pro. NDWI data can also be downloaded through the Copernicus Browser (https: / / dataspace.copernicus.eu / ) using the calculation formula shown above.

[0033] The ALOS-DEM data were resampled to a 10m resolution using the Resample tool in ArcGIS Pro, and slope data were extracted using the Slope tool. The processed data were then composited using the Band Composite tool in ArcGIS Pro to produce a six-channel fused feature dataset: B2 (Blue), B3 (Green), B4 (Red), B8 (NIR), NDWI, and slope.

[0034] In step S2, the data is first labeled.

[0035] Deep learning sample labeling was performed based on collected data from an on-site glacial lake extent survey. Sample labeling selected representative glacial lakes, taking into account features such as area, elevation, color, and shape. All glacial lake features were required to be uniform in number and accurately defined. To meet the input requirements of the Mask R-CNN model, each sample was sliced into image slices of equal size using a 256×256 pixel window. The overlap between image slices was set to 16 pixels. To increase the sample size, the sample data was randomly divided into a training dataset (90%) and a test dataset (10%) using random cropping and 90° rotation. Input image slices were normalized to significantly improve model convergence.

[0036] In this embodiment, the mean standard deviation method is used to normalize all input image slices. The normalization formula is as follows: ; Where, Indicates the i The original data points, is the minimum value in the original data set, is the maximum value in the original data set, The corresponding i Normalized values of the original data points.

[0037] After obtaining the data set, step S3 builds and trains the model.

[0038] In a preferred embodiment, introducing the CBAM module into the high-level features of the backbone network includes: introducing the CBAM module into the Conv4 and Conv5 layers.

[0039] The CBAM module combines channel attention and spatial attention mechanisms, where: The channel attention module performs global average pooling and maximum pooling operations on the extracted features, generates channel weights, and applies these weights to the feature map to enhance the highly relevant features of the ice lake target.

[0040] The spatial attention module performs spatial attention operations on the feature map, weights local features through convolution kernels, highlights the spatial location information of the ice lake, and reduces other background interference.

[0041] like Figure 2 As shown in the figure, it is a schematic diagram of the Mask R-CNN model structure. The specific analysis is as follows: 1. Build a Mask R-CNN model based on ResNet-50, which includes the following main modules: Backbone Network: Combines ResNet-50 with Feature Pyramid Network (FPN) to extract features from the input image through a multi-scale feature pyramid.

[0042] Region Proposal Network (RPN): Generates candidate regions of different sizes and scales on the feature pyramid to capture potential ice lake targets.

[0043] ROI Align layer: maintains the spatial correspondence between the image and the mask to achieve pixel-level accurate mask prediction.

[0044] Detection Head: performs object classification, bounding box regression, and mask prediction for each region.

[0045] 2. Introduce the CBAM module (Convolutional Block Attention Module) into the high-level features of the backbone network (Conv4 and Conv5 layers). The CBAM module enhances the recognition ability of ice lake targets by combining channel attention and spatial attention mechanisms. The specific steps are as follows: Channel Attention Module: Performs global average pooling and maximum pooling on the extracted features to generate channel weights, and applies these weights to the feature map to enhance the highly relevant features of the Ice Lake target.

[0046] Spatial Attention Module: Performs spatial attention on the feature map, weighting local features through convolution kernels to highlight the spatial location information of the ice lake and reduce other background interference.

[0047] 3. Introducing the CBAM module into each layer of the feature pyramid network (FPN) feature map. The different scale feature maps of FPN represent different spatial details and global semantic information. By introducing the CBAM module, the expressive power of these feature maps can be effectively improved: After introducing the CBAM module at each feature layer, the channel attention module strengthens the weight of key feature channels, and the spatial attention module further focuses on the boundaries and shape details of the ice lake at different scales.

[0048] The multiple introduction of CBAM ensures that feature maps at different scales can effectively capture the characteristics of glacial lakes, especially those with small areas and irregular shapes.

[0049] 4. Introducing the CBAM module into the Detection Head to further enhance the mask branch’s ability to process pixel-level details: Through the channel attention mechanism of the CBAM module, the model can pay more attention to the high-confidence pixels in the ice lake area.

[0050] The spatial attention module helps the model better distinguish the boundaries and complex shapes of glacial lakes at the pixel level.

[0051] 5. The model adopts a multi-task training strategy.

[0052] like Figure 3 As shown in Figure 2, it is a graph of the loss function for model training.

[0053] In the preferred solution, in step S2, during model training, a multi-task joint loss function training strategy is adopted, comprehensively considering the target classification loss , bounding box regression loss and segmentation mask loss , the formula is: ; Among them, the mask loss is calculated by the Sigmoid function of each pixel, and the weight information brought by the CBAM module is combined in the loss function to further improve the segmentation accuracy.

[0054] Target classification loss Cross Entropy Loss is used to measure the difference between the predicted category and the actual category. The formula is: ; Where: is the true category, The probability predicted by the model; Bounding box regression loss A smooth L1 loss function is used to optimize the gap between the predicted bounding box position and the true bounding box. The formula is: ; Where: is the true bounding box position, To predict the bounding box location, the smooth L1 loss is defined as: ; Segmentation mask loss Use the binary cross entropy loss for each pixel and combine it with the attention weight brought by the CBAM module to improve the segmentation accuracy. The loss on each pixel is calculated as follows: ; Where: is the true mask pixel label, is the predicted mask pixel, is the output of the Sigmoid function applied to each pixel, that is, the predicted probability.

[0055] 6. The improved Mask R-CNN model was trained end-to-end using processed and fused multi-channel data (blue, green, red, NIR, NDWI, and slope). After introducing the CBAM module, the model, based on the fused feature data, can more effectively extract the boundary features and internal structure of glacial lakes, significantly improving the accuracy of glacial lake identification and segmentation.

[0056] Based on the collected field survey data and high-resolution image data as reference standards, the method of this embodiment was evaluated and optimized before being put into practical use.

[0057] In the preferred solution, the loss weight is dynamically adjusted based on the performance of each task of the multi-task loss function, namely classification, detection and segmentation, on the validation set. The task weight is inversely proportional to the error. The higher the error, the greater the weight. The formula is: ; Where, is the classification error rate, for Detection error, is the segmentation error; The weighted total loss expression is: ; Where, and are the weight coefficients for classification, detection and segmentation respectively, and are the loss functions for classification, detection and segmentation respectively.

[0058] Through the above methods, the model's ability to recognize the boundaries of glacial lakes is improved, and the stability of multi-task learning is improved.

[0059] This embodiment takes a certain area in southeastern Tibet as an example and explains it in combination with actual data.

[0060] Step 1: Collect data.

[0061] Using the Copernicus browser, we downloaded eight high-resolution Sentinel-2 remote sensing images covering the identified area from June to October 2023. A data overview is shown in Table 1. We also acquired the ALOS-DEM data for the identified area. We also obtained a 30-meter-resolution glacial lake dataset for the Asian high mountain region in 2017 for deep learning labeling and accuracy verification.

[0062] Table 1. Sentinel-2 data overview

[0063] The second step is data preprocessing to obtain a six-channel dataset of B2 (Blue), B3 (Green), B4 (Red), B8 (NIR), NDWI, and slope.

[0064] Step 3: Label the sample data.

[0065] Based on the 30-meter-resolution glacial lake dataset in Asia's high mountains in 2017 and compared with image data, typical glacial lake samples were selected according to characteristics such as area, altitude, color, and shape, and deep learning sample labeling was performed. At the same time, the sample labeling of small glacial lakes (<0.05 km2) was also taken into consideration, and a total of 225 samples were labeled.

[0066] Each sample uses a 256×256 pixel slice window to partition the sample data and corresponding labels into image slices of equal size. The overlap between image slices is set to 16 pixels. To increase the number of samples, random cropping and 90° rotation are used to randomly divide the data into a training dataset (90%) and a test dataset (10%). The input image slices are normalized.

[0067] Step 4: Build and train the improved Mask R-CNN model.

[0068] Step 5: Accuracy Assessment and Optimization: Using the 2017 Asian High Mountain Area 30-meter resolution glacial lake dataset and high-resolution image data as reference standards, the second-level evaluation indicators of precision, recall, and accuracy are used to evaluate the accuracy of the recognition model. The accuracy evaluation of this embodiment is shown in Table 2: Table 2 Comparison of model accuracy

[0069] Step 6: Automatic identification of ice lakes: Use the Detect Objects Using DeepingLearning module on the ArcGIS Pro platform to automatically identify ice lakes in the image. Figure 4 FIG. 1 shows a structural diagram of the automatic identification of ice lakes according to this embodiment.

[0070] The Mask R-CNN deep learning model for identifying glacial lakes proposed in this embodiment is generated using high-resolution ALOS-DEM elevation data. The Sentinel-2 data used has high spatial resolution, making it more practical for identifying and extracting small glacial lakes. Furthermore, the glacial lake annotation data of the present invention is based on on-site glacial lake range survey data combined with manual annotation. The dataset includes a large number of small glacial lake samples, making the trained deep learning model more user-friendly for small glacial lakes and extracting more precise glacial lake boundaries, which helps improve the overall accuracy of lake extraction.

[0071] Example 2 In conjunction with Example 1, a multi-source data fusion intelligent identification system for glacial lakes in complex plateau terrain areas is provided, which is characterized by comprising: The data acquisition module is used to acquire Sentinel-2 data, slope data extracted from ALOS-DEM, and NDWI data. After data preprocessing, it forms a multi-channel dataset containing red, green, blue, near-infrared, NDWI, and slope data through band synthesis.

[0072] The data partitioning module is used to partition the multi-channel data set into a training data set and a test data set according to a preset ratio after the second preprocessing.

[0073] The model building module is used to build a Mask R-CNN model based on ResNet-50. The CBAM module is introduced in the high-level backbone network (Conv4 / Conv5), each FPN layer, and the detection head. After inputting the training dataset for multi-task joint loss function training, the optimal Mask R-CNN model is saved.

[0074] The recognition module is used for automatic ice lake identification: After data preprocessing, the test data is input into the optimal Mask R-CNN model to obtain ice lake identification results.

[0075] An electronic device, comprising a memory and a processor; memory for storing computer programs; The processor is configured to implement the ice lake intelligent identification method using multi-source data fusion as described in Example 1 when executing the computer program.

[0076] The above embodiments are merely preferred technical solutions of the present invention and should not be construed as limiting the present invention. The scope of protection of the present invention shall be the technical solutions set forth in the claims, including equivalent alternatives to the technical features of the technical solutions set forth in the claims. In other words, equivalent alternatives and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A multi-source data fusion method for intelligent identification of icy lakes, characterized by: The following steps are involved: S1: Acquire Sentinel-2 data, slope data extracted from ALOS-DEM, and NDWI data, perform data preprocessing, and then synthesize the bands to form a multi-channel dataset containing red, green, blue, near-infrared, NDWI, and slope data; S2: After the multi-channel dataset undergoes the second preprocessing, it is divided into a training dataset and a test dataset according to a preset ratio; S3: Build a Mask R-CNN model based on ResNet-50. Introduce the CBAM module into the high-level layers of the backbone network (Conv4 / Conv5), each FPN layer, and the detection head. After inputting the training dataset for multi-task joint loss function training, save the optimal Mask R-CNN model. S4: Automatic Identification of Ice Lakes: After preprocessing the test data, input it into the optimal Mask R-CNN model to obtain the ice lake identification results.

2. The multi-source data fusion ice lake intelligent identification method according to claim 1 is characterized in that: Also includes: Evaluation and optimization: The secondary evaluation indicators precision, recall rate and accuracy are used as the accuracy evaluation of the Mask R-CNN model. The formula is: ； ； ； Among them, TP is true positive, which means the ice lake area is accurately described; FP is false positive, which is defined as the non-ice lake area detected by the application method as the ice lake area in the image; FN is a false negative, indicating that the actual ice lake slope area was not detected; TN is a true negative, indicating that the non-glacial lake area was detected.

3. The multi-source data fusion ice lake intelligent identification method according to claim 1 is characterized in that: The data preprocessing includes: After atmospheric correction, Sentinel-2 data were converted to ENVI format; ALOS-DEM data were resampled to the preset resolution, and slope data were extracted using the slope tool; The calculation formula for NDWI data is: ； Where B3 is the Green band data and B8 is the NIR band data.

4. The multi-source data fusion ice lake intelligent identification method according to claim 1 is characterized in that: Sentinel-2 data is a high-resolution remote sensing image. The image should be clear, with moderate contrast and rich color gradation, and the cloud cover should be less than 20%. Specifically: A pre-trained U-Net cloud detection model was used to retain images with cloud coverage less than 20%; For clarity evaluation, the average gradient (AG) is used to reflect edge sharpness: Standard deviation (STD) + dynamic range (DR) is used for contrast evaluation; The HSV spatial saturation mean is used to evaluate color richness.

5. The multi-source data fusion ice lake intelligent identification method according to claim 1 is characterized in that: The introduction of CBAM modules into high-level features of the backbone network includes: the introduction of CBAM modules into Conv4 and Conv5 layers; The CBAM module combines channel attention and spatial attention mechanisms, where: The channel attention module performs global average pooling and maximum pooling on the extracted features to generate channel weights, and applies these weights to the feature map to enhance the highly relevant features of the ice lake target; The spatial attention module performs spatial attention operations on the feature map, weights local features through convolution kernels, highlights the spatial location information of the ice lake, and suppresses background interference.

6. The multi-source data fusion ice lake intelligent identification method according to claim 1 is characterized in that: In S2, during model training, a multi-task joint loss function training strategy is adopted, comprehensively considering the target classification loss , bounding box regression loss and segmentation mask loss , the formula is: ； Target classification loss Cross Entropy Loss is used to measure the difference between the predicted category and the actual category. The formula is: ； Where: is the true category, The probability predicted by the model; Bounding box regression loss A smooth L1 loss function is used to optimize the gap between the predicted bounding box position and the true bounding box. The formula is: ； Where: is the true bounding box position, To predict the bounding box location, the smooth L1 loss is defined as: ； Segmentation mask loss Use the binary cross entropy loss for each pixel and combine it with the attention weight brought by the CBAM module to improve the segmentation accuracy. The loss on each pixel is calculated as follows: ； Where: is the true mask pixel label, is the predicted mask pixel, is the output of the Sigmoid function applied to each pixel, that is, the predicted probability.

7. The multi-source data fusion ice lake intelligent identification method according to claim 1 is characterized in that: According to the performance of each task of the multi-task loss function, namely classification, detection and segmentation, on the validation set, the loss weight is dynamically adjusted. The task weight is inversely proportional to the error. The higher the error, the greater the weight. The formula is: ； Where, is the classification error rate, for Detection error, is the segmentation error; The weighted total loss expression is: ； Where, and are the weight coefficients for classification, detection and segmentation respectively, and are the loss functions for classification, detection and segmentation respectively.

8. The multi-source data fusion ice lake intelligent identification method according to claim 1 is characterized in that: S2: The second preprocessing includes: Considering the area, elevation, color and shape characteristics, multiple images are color-homogenized, and multi-channel datasets are labeled and processed by slicing, random cropping, rotation and normalization.

9. An intelligent ice lake identification system based on multi-source data fusion, characterized by: include: The data acquisition module is used to acquire Sentinel-2 data, slope data extracted from ALOS-DEM, and NDWI data. After data preprocessing, it synthesizes the bands to form a multi-channel dataset containing red, green, blue, near-infrared, NDWI, and slope data. A data partitioning module is used to partition the multi-channel data set into a training data set and a test data set according to a preset ratio after the second preprocessing; The model building module is used to build the Mask R-CNN model based on ResNet-50. The CBAM module is introduced in the high-level backbone network (Conv4 / Conv5), each FPN layer, and the detection head. After inputting the training dataset and performing multi-task joint loss function training, the optimal Mask R-CNN model is saved. The recognition module is used for automatic ice lake identification: After data preprocessing, the test data is input into the optimal Mask R-CNN model to obtain ice lake identification results.

10. An electronic device, characterized in that: including memory and processor; The memory is used to store computer programs; The processor is configured to implement the ice lake intelligent identification method based on multi-source data fusion as described in any one of claims 1 to 8 when executing the computer program.

Images