A deep learning-based glacial area surface water resource distribution image extraction method

Abstract

The application discloses a kind of based on deep learning's ice glacier area ground surface water resource distribution image extraction method, belong to ground surface water resource image extraction field, including steps: obtaining the multispectral remote sensing image and digital elevation model data covering ice glacier area, data pre-processing is carried out;YOLOv12 instance segmentation model including main network, neck network and detection head is built, and improved YOLOv12 segmentation model is constructed;Model training and optimization are carried out;ground surface water resource is segmented and extracted by the segmented model after optimization;after segmentation is completed, finally the distribution map of the optimized multi-class ground surface water resource is generated.The application is based on the improved YOLOv12 network architecture, through the fusion vision, convolution block attention module and space to depth convolution technology, the ground surface water resource in ice glacier area in remote sensing image is comprehensively and automatically extracted, effectively deal with ice and snow high reflection, ice crack, stone ice glacier complex texture and edge fuzzy problem, improve segmentation precision.

Description

Technical Field

[0001] This invention relates to the field of surface water resource image extraction technology, and in particular to a method for extracting surface water resource distribution images in glacial areas based on deep learning. Background Technology

[0002] Modern glacial regions refer to high-altitude, arid areas where modern glacial activity has impacted the landform, soil, and rock environment. These regions are characterized by scarce water resources, fragile ecosystems, and sparse human presence, making the acquisition of surface water resource information crucial. However, effective monitoring methods for the current status and dynamic changes of surface water resources in glacial regions are currently lacking. Given the poor accessibility of glacial regions, satellite remote sensing interpretation has become a key means of detecting water resource changes. Traditionally, researchers have used optical imagery combined with empirical thresholds: for example, extracting glacier information based on the Normalized Difference Snow Index (NDSI), extracting glacial lake information based on the Normalized Difference Water Index (NDWI), and collaboratively extracting stony glacier information using the Soil Adjusted Vegetation Index (SAVI) and the Enhanced Vegetation Index (EVI). While these methods can segment different types of surface water resources by artificially setting thresholds, they typically only target a single objective and lack adaptability to complex terrain and variable meteorological conditions. For example, in scenarios where the lake surface is frozen or covered with snow, the boundary of the frozen lake is easily misjudged, leading to reduced segmentation accuracy and stability.

[0003] With the rapid development of computer vision, deep learning methods have been widely used in water resource extraction in glacial areas due to their advantages of high segmentation accuracy, high speed, and minimal human intervention. However, related research is mainly based on U-Net or its improved networks, which typically suffer from high computational complexity and limited generalization ability. Recent studies have shown that by improving the DeepLabV3+ network and the Mask R-CNN model, high-precision extraction of glaciers or glacial lakes can be achieved in complex high-altitude areas; models that integrate convolutional neural networks (CNN) and Transformers have also achieved high overall accuracy on multi-temporal satellite imagery. While these methods demonstrate the powerful capabilities of deep learning in glacier and related water body extraction, they often target single objects or specific data sources, lacking a unified framework for extracting multiple types of surface water resources in glacial areas. Furthermore, the application of deep learning methods in the cryosphere still faces many challenges. For example, the spatiotemporal generalization ability of large-scale regions is insufficient, making it difficult to adapt to the differences in data from different sensors; the extraction of the boundaries of partially frozen ice lakes is inaccurate, and mixed ice, debris, or wet ice areas are prone to misclassification; and there is no complete technical framework that can comprehensively extract multiple targets such as glaciers, ice lakes, snow, and stony glaciers.

[0004] The existing surface water resource segmentation technology in glacial areas faces the following major bottlenecks:

[0005] (1) The threshold method has poor adaptability: Due to limitations in sensor performance and data type, traditional remote sensing segmentation methods often rely on manually set thresholds, which are difficult to adapt to complex terrain and variable climate conditions. In extreme scenarios such as frozen lakes and snow cover, ice lake extraction based on empirical thresholds often results in boundary misjudgment, and the segmentation accuracy and stability are low.

[0006] (2) Limitations of deep learning methods: Existing deep learning segmentation models are often designed for single-type targets and cannot handle multiple types of land cover at the same time. The generalization ability of the models on multi-source remote sensing images is insufficient, and it is difficult to separate different targets that are stuck together during instance segmentation; the calculation of multi-temporal data fusion is complex and the real-time processing efficiency is not high, making it difficult to meet the needs of simultaneous extraction of multiple targets.

[0007] (3) Limited data representativeness: The terrain of the glacier area is complex and varied, and the spectral characteristics of ground objects are blurred. The training data is difficult to fully cover all situations, resulting in poor performance of the model when processing low-quality and blurry images. It is impossible to fully obtain the overall picture and dynamic change information of surface water resources in the glacier area.

[0008] In summary, existing technologies are insufficient to simultaneously meet the comprehensive, real-time, and high-precision requirements of surface water resource surveys in glacial areas, and significant technical bottlenecks exist that urgently need to be overcome. Summary of the Invention

[0009] The purpose of this invention is to provide a deep learning-based method for extracting surface water resources distribution images in glacier areas. Based on an improved YOLOv12 network architecture, this method integrates visual processing, convolutional block attention modules, and spatial-to-depth convolution techniques to comprehensively and automatically extract glacier surface water resources from remote sensing images. This method effectively addresses challenges such as high reflectivity of snow and ice, ice fissures, complex textures of rocky glaciers, and blurred edges, thereby improving segmentation accuracy.

[0010] To achieve the above objectives, this invention provides a method for extracting surface water resource distribution images in glacial areas based on deep learning, comprising the following steps:

[0011] S1, Multi-source data preprocessing:

[0012] S11, Band Synthesis: Acquire multispectral remote sensing images and digital elevation model (DEM) data covering the glacier area, synthesize the multispectral remote sensing images into bands, use the DEM data to perform topographic correction on the multispectral remote sensing images, unify the multispectral remote sensing images to the same resolution, and calculate the Soil Adjusted Vegetation Index (SAVI) and Enhanced Vegetation Index (EVI). The glacier area includes glaciers, glacial lakes, snow cover, and stony glaciers.

[0013] S12, Normalized Difference Index Calculation: Based on the multispectral remote sensing image corrected by S11, the normalized difference index is calculated, and the normalized differential snow index (NDSI) and normalized differential water index (NDWI) are calculated.

[0014] S13, Tag Format Conversion: Obtain vector data of water resources in glacier areas and convert the vector boundary labels of the water resources vector data in glacier areas into YOLO format labels;

[0015] S14, Image Slicing and Pairing: Slice multispectral remote sensing images with their corresponding converted labels;

[0016] S15, Data Augmentation: Using the cropped multispectral remote sensing images as samples in the training set, data augmentation is performed based on the spatial distribution and characteristics of the samples in the training set.

[0017] S16, Sample screening: Filter the cropped multispectral remote sensing images according to the target area threshold;

[0018] S17, Radiometric Normalization and Correction: Radiometric correction is performed on multispectral remote sensing images from different periods, and the band images of the multispectral remote sensing images and the calculated index images are linearly normalized.

[0019] S2, Build an improved YOLOv12 segmentation model including a backbone network, a neck network, and a detection head:

[0020] S21, Backbone Network Optimization: CSPDarknet53 is adopted as the backbone. While retaining the original network structure in the first three stages, the ViT global context module is introduced in the fourth stage.

[0021] S22, Feature Recalibration: Convolutional Block Attention (CBAM) modules are connected in series at the output end of each stage of the backbone network to perform feature recalibration;

[0022] S23, Spatial-Depth Transform Downsampling: Replaces the downsampling convolution operation in the original YOLOv12 model network with a combination of spatial-to-depth convolution and standard convolution;

[0023] S24, Multi-scale segmentation head: Three detection segmentation heads are set at the output of the improved YOLOv12 segmentation model, corresponding to feature layers with different downsampling factors;

[0024] S3, Model Training and Optimization: Multispectral remote sensing images obtained from multi-source data preprocessing are used as input, and the converted labels corresponding to the multispectral remote sensing images are used as supervision signals to train the improved YOLOv12 segmentation model. A weighted sum of bounding box regression loss, classification loss, and segmentation loss is used as the total loss function, and the stochastic gradient descent algorithm is employed to optimize the parameters of the improved YOLOv12 segmentation model. During training, the F1-score evaluation metric is introduced to balance precision and recall, and mean AP is used to monitor and improve the performance of the YOLOv12 segmentation model. The learning rate and positive / negative sample weights are adjusted to improve training stability. Training continues until the F1-score on the validation set converges without significant improvement, at which point the final segmentation model is output.

[0025] S4, Surface Water Resource Segmentation and Extraction: After preprocessing, the multispectral remote sensing image of the glacier area to be analyzed is input into the final segmentation model trained in S3. The segmentation model is used to detect and segment surface water resource targets in the multispectral remote sensing image, and segmentation mask maps of each target including glaciers, glacial lakes, snow cover, and stony glaciers are obtained. Combining the bounding box coordinates and masks output by the segmentation model, each target instance is located and classified.

[0026] S5, Post-processing of results: Boundary optimization is performed on the initial segmentation results output by the segmentation model. The segmentation mask of each type of target is converted into a vector polygon, and the contour is smoothed to remove jagged noise. Segmentation polygons that are adjacent and belong to the same target are merged to eliminate fragmented segmentation. Isolated false detection areas with an area lower than the preset threshold are deleted. Finally, an optimized multi-category surface water resource distribution map is generated.

[0027] Preferably, in S11, the acquired multispectral remote sensing images include true-color blue, green, and red bands, near-infrared (NIR) bands, and short-wave infrared (SWIR) bands. The acquired multispectral remote sensing images are synthesized to form an image containing five spectral bands. The Soil-Adjusted Vegetation Index (SAVI) and Enhanced Vegetation Index (EVI) are calculated as additional channels to characterize the spectral features of glacial lakes and glaciers, obtaining rich spectral information input for surface water resource instances in glacial areas. The calculation formulas for SAVI and EVI are as follows:

[0028]

[0029] In the formula, NIR and RED are the reflectance in the near-infrared band and the red band, respectively; BLUE is the reflectance in the blue band; C1 is the atmospheric correction parameter for red light, which is taken as 6.0; C2 is the atmospheric correction parameter for blue light, which is taken as 7.5; and L is the soil adjustment parameter, which is taken as 1.

[0030] Preferably, in S12, binary maps of Normalized Difference Water Index (NDWI) and Normalized Difference Snow Index (NDSI) are calculated using the corrected multispectral remote sensing imagery. These maps highlight the boundary and extent changes of liquid water and solid ice and snow, respectively. Preliminary results of the glacier-lake boundary and the distribution of glacier / snow cover areas are obtained through threshold segmentation. The calculation methods for Normalized Difference Water Index (NDWI) and Normalized Difference Snow Index (NDSI) are as follows:

[0031]

[0032] In the formula, Green represents the green band, NIR represents the near-infrared band, and SWIR represents the short-wave infrared band. Based on the image in the above formula, a corresponding binary map is generated to obtain the spectral information of the corresponding ground features.

[0033] Preferably, in S13, the vector polygon labels (such as Shapefile format) of the ice lake, glacier, and stony glacier targets are converted into COCO format text label files required by YOLO-Seg; specifically, the vertex coordinates of each target polygon are extracted, normalized according to YOLO-Seg format, and output as TXT labels for use in the training of the instance segmentation model.

[0034] In step S14, the multispectral remote sensing image and its corresponding converted label are cropped into small patches of a fixed size (e.g., 640×640 pixels); ensuring that each patch contains a complete target polygon label. This patch pairing process facilitates increasing the number of training samples and avoids memory overhead caused by excessively large images.

[0035] In S15, random rotation, mirror flipping, and salt-and-pepper noise image enhancement operations are performed based on the spatial distribution and features of the samples in the training set to expand the diversity of the training data and reduce the overfitting of the segmentation model to specific scenes.

[0036] In S16, the cropped multispectral remote sensing image patch is filtered according to the target area threshold, and irrelevant samples without valid targets are removed to ensure that each sample in the training set contains surface water resource instances, thereby improving the training efficiency of the segmentation model.

[0037] Preferably, in S21, the backbone network is optimized as follows: the ViT global context module is a visual Transformer module, which is inserted between the third and fourth stages of the backbone network to fuse local convolutional features with long-range dependence information of glacier surface, thereby enhancing the segmentation model's ability to understand large-scale spatial structure and connectivity.

[0038] In S22, CBAM comprises two parts: Channel Attention Module and Spatial Attention Module. These are executed sequentially on the output features of each stage of the backbone network, recalibrating the feature maps to enhance their focus on surface water resource targets. Channel Attention generates channel weights through global pooling and a multilayer perceptron (MLP), while Spatial Attention generates positional weights through convolution. The two are then multiplied element-wise to obtain the enhanced features. Introducing CBAM can preserve key information while suppressing noise interference, thereby improving the segmentation model's ability to identify small target instances.

[0039] Preferably, in S23, the space-depth transformation downsampling involves replacing all downsampling convolution operations with a stride of 2 in the original YOLOv12 network with a combination of Space-to-Depth Convolution (SPD-Conv) and standard convolutions. SPD-Conv achieves a halving of spatial resolution and an increase in the number of channels by dividing the input feature map into sub-regions and rearranging them to the channel dimension. Compared to direct downsampling, SPD-Conv effectively preserves fine-grained edge and texture information, which helps improve the discrimination ability for small-area targets (such as narrow glacial lakes and thin snow cover).

[0040] Preferably, in S24, the three detection and segmentation heads correspond to feature layers with different downsampling factors (40×40, 20×20, 10×10). Each detection and segmentation head simultaneously predicts the target category, bounding box, and segmentation mask. Prediction on multi-scale feature maps can take into account ground objects of different sizes and improve the detection and segmentation effect of the segmentation model on water resources and water bodies at different scales.

[0041] Preferably, in step S3, after each training iteration, the precision and recall of the improved YOLOv12 segmentation model on the validation set are calculated, and the F1-score is calculated accordingly. Training stops when the F1-score gradually levels off with each training iteration. By monitoring the changes in F1-score and mean AP, the false negative rate and false positive rate of the segmentation model are balanced, enabling the trained segmentation model to have higher overall accuracy and robustness in multi-object segmentation tasks.

[0042] Therefore, the present invention employs the above-mentioned deep learning-based image extraction method for surface water resource distribution in glacial areas, which has the following beneficial effects:

[0043] 1. Eliminate manual intervention and improve segmentation stability: An adaptive multimodal feature fusion fully automatic segmentation method is proposed to replace manual threshold adjustment, enabling the segmentation model to adapt to complex scene changes, effectively distinguish adjacent water body instances in the same image, reduce the phenomenon of misjudgment due to adhesion, and improve the stability of segmentation results.

[0044] 2. Enhance generalization ability to meet real-time requirements: Adopt lightweight design of segmentation model and transfer learning technology to reduce the computational complexity of multi-temporal and multi-source remote sensing data processing. By improving the inference speed and generalization performance of segmentation model, real-time response of segmentation results can be achieved, enabling it to adapt to data from different times and regions.

[0045] 3. Construct a comprehensive extraction framework to improve computing efficiency: Integrate remote sensing information processing methods, digital image processing technology and deep learning architecture to construct a comprehensive automatic extraction framework that can simultaneously extract multiple types of targets such as glaciers, glacial lakes, snow cover and stony glaciers. This method completes the entire chain of processing from data preprocessing to target segmentation on a case-by-case basis, and can output the spatial distribution of surface water resources in glacier areas in an integrated manner.

[0046] 4. Improve segmentation accuracy and meet multiple needs: Through innovative network structure improvements, the segmentation model enhances the accuracy of identifying complex water targets in glacial areas, thereby improving segmentation accuracy. The segmentation model balances accuracy and speed, meeting the dual requirements of high precision and real-time performance for water resource monitoring in high-altitude and cold mountainous areas, and improving the performance of result services.

[0047] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the technical process of an embodiment of the deep learning-based image extraction method for surface water resource distribution in glacier areas according to the present invention.

[0049] Figure 2This is a schematic diagram of the improved YOLOv12 segmentation model network structure according to an embodiment of the present invention;

[0050] Figure 3 This is a schematic diagram of the convolutional block attention module structure according to an embodiment of the present invention;

[0051] Figure 4 This is a schematic diagram of the spatial-to-depth convolutional downsampling process in an embodiment of the present invention;

[0052] Figure 5 This is a graph showing the performance index changes during the training process of the improved YOLOv12 segmentation model according to an embodiment of the present invention.

[0053] Figure 6 This is a graph showing the change in evaluation metrics during the training process of the improved YOLOv12 segmentation model according to an embodiment of the present invention.

[0054] Figure 7 This is a visual comparison of the segmentation results of the improved YOLOv12 segmentation model in this embodiment of the invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments in this application without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0056] It should be noted that the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, such that a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0057] Similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0058] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the product of this invention is usually placed when in use. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0059] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," and "connect" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0060] Example

[0061] The method of this invention mainly includes three modules: data preprocessing, an improved YOLOv12 instance segmentation network, and segmentation result boundary optimization. The overall technical route and the relationship between each module are as follows: Figure 1 As shown.

[0062] The development and operating environment for the method of this invention is as follows:

[0063] ① Hardware and software environment configuration

[0064] hardware platform

[0065] Processor (CPU): Core TM i7-14600KF processor;

[0066] Graphics Processing Unit (GPU): NVIDIA GeForce RTX 5070 Ti graphics card;

[0067] Memory: 32GB DDR5 memory;

[0068] Storage: 2TB solid-state drive.

[0069] Software environment:

[0070] Operating system: Windows 11 64-bit;

[0071] Deep learning frameworks: PyTorch 2.0.0, CUDA (Compute Unified Device Architecture) 11.8;

[0072] Auxiliary tools: GDAL 3.4.1 (Geographic Data Abstraction Library), OpenCV 4.5.5 (Open Source Computer Vision Library, used for image processing and enhancement).

[0073] ② Environment configuration process

[0074] To ensure the normal operation of the above-mentioned hardware and software environment for the method of this invention, the following configuration steps are required:

[0075] Install the Anaconda platform: Download the Anaconda installer package compatible with Python 3.8 and complete the installation. After installation, add the Anaconda installation path to the system environment variable PATH.

[0076] Installing CUDA and cuDNN: Depending on your GPU model, download the corresponding version of CUDA Toolkit (e.g., CUDA 11.8) and cuDNN library from the NVIDIA official website and install and configure them according to the instructions. This will provide the necessary drivers and deep learning operator support for GPU acceleration.

[0077] To create a Python virtual environment: Open the command line and execute `conda create -n yolo12python=3.8` to create a virtual environment named "yolo12", then activate it with `conda activate yolo12`. Virtual environments help isolate project dependencies.

[0078] Install the PyTorch framework: In the active virtual environment, select the installation command based on your system's GPU. If using a GPU and the CUDA acceleration library version is 11.8, you can execute: `pip install torch torchvision torchaudio --index-url https: / / download.pytorch.org / whl / cu118` to install PyTorch 2.0.0 with CUDA support and related components.

[0079] Install project dependencies: Execute `pip install ultralytics` in the virtual environment. This library contains the code implementations and commonly used utility functions required for the YOLO model. After installation, you can use the YOLO model structure and interface described in this invention.

[0080] The present invention discloses a deep learning-based method for extracting surface water resource distribution images in glacial areas, comprising the following steps:

[0081] S1, Multi-source data preprocessing:

[0082] S11, Band Combining: Acquire multispectral remote sensing images and digital elevation model (DEM) data covering glacial areas (glaciers, glacial lakes, snow cover, stony glaciers, etc.). Combine the true-color blue, green, and red bands of the acquired multispectral remote sensing images with the near-infrared (NIR) and short-wave infrared (SWIR) bands to form an image containing 5 spectral bands. Perform topographic correction on the image using the DEM data, and resample the short-wave infrared bands from 20m resolution to 10m resolution, consistent with other bands, using bilinear interpolation. Calculate the Enhanced Vegetation Index (EVI) and Soil-Regulated Vegetation Index (SAVI), which are input into the segmentation model along with the optical bands as additional channels to highlight the spectral characteristics of moraine-covered glaciers and stony glaciers. Obtain rich spectral information on surface water resources in glacial areas. The calculation formulas for Soil-Regulated Vegetation Index (SAVI) and Enhanced Vegetation Index (EVI) are as follows:

[0083]

[0084] In the formula, NIR and RED are the reflectance in the near-infrared band and the red band, respectively; BLUE is the reflectance in the blue band; C1 is the atmospheric correction parameter for red light, which is taken as 6.0; C2 is the atmospheric correction parameter for blue light, which is taken as 7.5; and L is the soil adjustment parameter, which is taken as 1.

[0085] S12, Normalized Difference Index Calculation: Using the multispectral remote sensing image corrected in S11, binary maps of the Normalized Water Index (NDWI) and Normalized Differential Snow Index (NDSI) are calculated to highlight the boundary and extent changes of liquid water and solid ice and snow, respectively. Preliminary results of glacier-lake boundaries and the distribution of glacier / snow cover areas are obtained through threshold segmentation. The calculation methods of the Normalized Water Index (NDWI) and Normalized Differential Snow Index (NDSI) are as follows:

[0086]

[0087]

[0088] In the formula, Green represents the green band, NIR represents the near-infrared band, and SWIR represents the short-wave infrared band. Based on the image in the above formula, a corresponding binary map is generated to obtain the spectral information of the corresponding ground features.

[0089] S13, Label Format Conversion: Convert the vector polygon labels (such as Shapefile format) of ice lakes, glaciers, and stony glaciers into COCO format text label files required by YOLO-Seg; specifically, extract the vertex coordinates of each target polygon, normalize them according to YOLO-Seg format, and output them as TXT labels for use in the training of the instance segmentation model.

[0090] S14, Image Tackling and Pairing: Large-scale original multispectral remote sensing images and their corresponding labels are cropped into small patches of a fixed size (e.g., 640×640 pixels); ensuring that each patch contains a complete target polygon label. This patch pairing process facilitates increasing the number of training samples and avoids memory overhead caused by excessively large images.

[0091] S15, Data Augmentation: Based on the spatial distribution and features of the samples in the training set, perform random rotation, mirror flipping, and salt-and-pepper noise image augmentation operations to expand the diversity of the training data and reduce the overfitting of the segmentation model to specific scenes.

[0092] S16, Sample Selection: Filter the cropped image patches according to the target area threshold, remove irrelevant samples that do not contain valid targets, and ensure that each sample in the training set contains surface water resource instances, thereby improving the training efficiency of the segmentation model.

[0093] S17, Radiometric Normalization and Correction: Radiometric correction is performed on multispectral remote sensing images from different periods. The band images of the multispectral remote sensing images and the calculated index images are linearly normalized. This makes the data acquired by different sensors and at different times comparable in spectrum. By eliminating differences in atmospheric and observation conditions, the segmentation model can ensure consistent performance of cross-scene data.

[0094] This embodiment selects multispectral remote sensing data and digital elevation model (DEM) data provided by the European Space Agency's Sentinel-2 satellite as the main data sources, which contain rich surface information of the glacier area. The processing flow is as follows:

[0095] Input data preparation: Acquire Sentinel-2 Level-2A optical imagery covering the target glacier area, along with corresponding DEM elevation data. Sentinel-2 provides multispectral images in 13 bands, of which this invention selects bands with higher resolution and close relevance to water body identification: blue B2 (center wavelength approximately 490 nm), green B3 (560 nm), red B4 (665 nm), near-infrared B8 (842 nm), and short-wave infrared B11 (1610 nm). The above spectral data, after radiometric calibration and atmospheric correction, can be used for subsequent analysis.

[0096] Data calibration and alignment: Due to the different original resolutions of different bands (10m for Sentinel-2's B2 / B3 / B4 / B8, and 20m for B11), the nearest neighbor interpolation method was first used to resample all bands to a uniform 10m spatial resolution. Simultaneously, the DEM data was cropped and resampled to align with the same spatial range and resolution as the optical imagery, ensuring a one-to-one correspondence between data sources at the pixel level. Geometric correction was performed as necessary to correct distortion and offset in the satellite imagery. After this step, multi-band optical imagery and DEM data aligned under the same grid were obtained.

[0097] Remote sensing index calculation: Based on the calibrated and aligned Sentinel-2 optical bands, several commonly used remote sensing indices are calculated to highlight the characteristics of different types of surface water resources: Normalized Differential Snow Index (NDSI), Normalized Differential Water Index (NDWI), Enhanced Vegetation Index (EVI), and Soil-Regulated Vegetation Index (SAVI). The calculated index images are thresholded to generate binary masks to indicate possible snow and water body distribution areas, providing prior information for subsequent segmentation.

[0098] Data normalization: Multi-band images and calculated exponential images are linearly normalized according to the 0-1 interval. This eliminates dimensional differences between different bands and accelerates the convergence of the neural network. The normalized spectral and exponential images can be stacked as needed to form a multi-channel input tensor (e.g., stacking 5 optical bands with several exponential layers to form a 7 or 8-channel image).

[0099] Training Label Preparation: Historical monitoring data or high-resolution image interpretation results of the study area were collected to obtain vector contours of targets such as glaciers, glacial lakes, snow cover, and stony glaciers. These vector labels were converted into YOLO-readable label files according to the COCO dataset format: for each target polygon, its normalized vertex coordinates relative to the image's width and height were calculated and stored as polygon annotations in a TXT file. Each training image corresponds to a TXT label file, containing the polygon coordinates and category identifiers of one or more targets. After the above processing, multi-source fused image data and standardized labels were finally obtained for use in training the segmentation model.

[0100] S2, Improved YOLOv12 Segmentation Model Construction: A YOLOv12 instance segmentation model was built, comprising a backbone network, a neck network, and a detection head. Based on the YOLO series of object detection and segmentation models, in-depth improvements were made for multi-object segmentation tasks in glacier regions. Figure 2 The overall structure of the improved YOLOv12 network is shown. The main improvements are as follows:

[0101] S21, Backbone Network Optimization: CSPDarknet53 is used as the backbone. While retaining the structure of the first three stages of the original network, a ViT global context module is introduced in the fourth stage. Specifically, a visual Transformer module is embedded after the third stage of the backbone to fuse local convolutional features with long-range dependency information of the glacier surface, enhancing the segmentation model's ability to understand large-scale spatial structure and connectivity. The visual Transformer module is inserted between the third and fourth stages of the backbone network, with a patch size of 8×8 pixels, and employs a 3-head self-attention mechanism to fuse long-range dependency features from glacier imagery.

[0102] S22, Feature Recalibration: Convolutional Block Attention Modules (CBAMs) are cascaded at the output ends of each stage of the backbone network (e.g., ...). Figure 3 As shown, (a) represents the structure of the channel attention submodule, and (b) represents the structure of the spatial attention submodule (this module is used to enhance the effectiveness of feature representation). CBAM consists of two parts: Channel Attention Module and Spatial Attention Module. These are executed sequentially on the output features of each stage of the backbone network, recalibrating the feature maps to improve attention to glaciers and water bodies. Channel attention generates channel weights through global pooling and a multilayer perceptron (MLP), while spatial attention generates positional weights through convolution. The two are then multiplied element-wise to obtain the enhanced features. Introducing CBAM can preserve key information while suppressing noise interference, thereby improving the segmentation model's ability to recognize small target instances.

[0103] S23, Space-Depth Transform Downsampling: All downsampling convolutions with a stride of 2 in the original YOLOv12 network are replaced with a combination of Space-to-Depth Convolution (SPD-Conv) and standard convolutions. SPD-Conv achieves a halving of spatial resolution and an increase in the number of channels by dividing the input feature map into sub-regions and rearranging them to the channel dimension (see...). Figure 4When the scaling factor is 2, the spatial dimensions of the input feature map are divided and reorganized to the channel dimension, achieving downsampling while preserving edge details. The figure shows (a) the input feature map, (b) and (c) the spatial-to-channel transformation process, (d) the reorganized feature map, and (e) the output result after feature extraction via convolution. Compared to direct downsampling, SPD-Conv effectively preserves fine-grained edge and texture information, which helps improve the discrimination ability of small-area targets (such as narrow glacial lakes and thin snow cover). The input feature map is divided into several sub-regions with a scaling factor of 2 and rearranged to the channel dimension, resulting in a reorganized feature map with half the original size and four times the number of channels. Then, a convolution with a stride of 1 is performed on the reorganized feature map to extract detailed features and compress the channel dimension. By replacing the original convolutional layer with a stride of 2 in the YOLOv12 network with the above SPD-Conv structure, the edge texture information of glacial targets is preserved while reducing the resolution of the feature map, thereby improving the segmentation accuracy of small-area glacial lakes and narrow glacial fissures.

[0104] S24, Multi-scale Segmentation Head: Three detection segmentation heads are set at the output of the improved network, corresponding to feature layers with different downsampling factors (40×40, 20×20, 10×10). Each detection segmentation head simultaneously predicts the target category, bounding box, and segmentation mask. Prediction on multi-scale feature maps can take into account ground objects of different sizes, improving the segmentation model's detection and segmentation performance for water bodies at different scales.

[0105] In this specific implementation, the key structure of the YOLOv12 segmentation model was improved as follows to meet the segmentation requirements of glacier region instances:

[0106] Backbone Improvement: The CSPDarknet53 backbone structure from the YOLOv5 series is adopted. The first three stages retain the original CSP (Cross Stage Partial) modules, each containing several 3×3 convolutional and residual modules, with output channels of 64, 128, and 256 respectively. In the fourth stage, a visual Transformer module is inserted to enhance global feature representation: the visual Transformer divides the input features into 8×8 patches, and global context is calculated through 4-head self-attention. The visual Transformer output is fused with the convolutional features, enriching the segmentation model's perception of large-scale connected ice and snow regions. The output channel dimension of this stage is set to 512 to provide higher feature representation capabilities.

[0107] Attention mechanism fusion: A CBAM module is introduced immediately after the output of each stage of the backbone network. Taking the final stage as an example, the output features first pass through the channel attention submodule to calculate the importance weight of each channel; then through the spatial attention submodule to assign weights to each position in the feature map. The attention mask generated by the combination of the two is multiplied element-wise with the original features to obtain a feature map that emphasizes key information. By performing the above processing on the feature maps of each stage, the segmentation model can pay more attention to the salient feature regions of glaciers and water bodies, filter out background interference, and improve the detection ability of small targets (such as scattered snow) and details (such as the winding edges of glacial lakes).

[0108] Downsampling layer modification: All existing downsampling convolutions with a stride of 2 in the network are replaced with a combination of "Space-to-Depth transformation + ordinary convolution" (i.e., SPD-Conv structure). Taking a feature map of size S×S and number of channels C1 as an example, the Space-to-Depth operation (scale 2) rearranges it into a tensor of size (S / 2×S / 2) and number of channels C1. Then, a 3×3 convolution (stride 1) is applied to reduce the number of channels to 2C1, while extracting more complex local features. This SPD-Conv replaces the original stride 2 convolution, achieving downsampling while preserving as much detail information as possible, especially ensuring that small-scale structural features such as ice fragments and fissures are not completely discarded due to downsampling. Experiments show that SPD-Conv significantly improves the segmentation effect of the segmentation model on the slender boundaries of glacial lakes and the texture of stony glaciers.

[0109] Output Header Configuration: The improved YOLOv12 uses three detector heads in the output layer, receiving feature maps of different scales: Layer 1 is 40×40 (downsampled by 16x), Layer 2 is 20×20 (downsampled by 32x), and Layer 3 is 10×10 (downsampled by 64x). Each detector head simultaneously regresses the target's bounding box, classification confidence, and segmentation mask for the corresponding scale feature map. When the detector head predicts the segmentation mask, it outputs the pixel mask contour for each target using a method similar to YOLO-Seg. The use of multi-scale detector heads enables the segmentation model to simultaneously handle large glaciers (identified on high-order feature maps) and small glacial lakes (identified on low-order detail features), improving the detection rate for targets of different sizes.

[0110] S3, Model Training and Optimization: Multispectral remote sensing images obtained from multi-source data preprocessing are used as input data, and the converted labels corresponding to the multispectral remote sensing images are used as supervision signals to train the improved YOLOv12 segmentation model. The weighted sum of the bounding box regression loss, classification loss, and segmentation loss in the improved YOLOv12 model framework is used as the total loss function, and the stochastic gradient descent algorithm is used to optimize the parameters of the improved YOLOv12 model. During training, the F1-score evaluation metric is introduced to balance precision and recall, and the mean AP is used to monitor the performance of the improved YOLOv12 model. The learning rate and positive / negative sample weights are adjusted to improve training stability. Training continues until the F1-score on the validation set converges without significant improvement, at which point the final segmentation model is output.

[0111] In S3, the improved YOLOv12 model training process incorporates the F1-score as a performance metric. After each training iteration, the precision and recall of the improved YOLOv12 model on the validation set are calculated, and the F1-score is calculated accordingly. Training stops when the F1-score plateaus with increasing training iterations. By monitoring the changes in F1-score and mean AP, the false negative and false positive rates of the segmentation model are balanced, resulting in a segmentation model with higher overall accuracy and robustness in multi-object segmentation tasks.

[0112] In this embodiment, F1-score is introduced as an evaluation metric during the segmentation model training process to balance the focus on precision and recall, while mean accuracy (mAP) is combined to evaluate the overall performance of the segmentation model. Specific implementation details are as follows:

[0113] Loss Function and Training Strategy: The multi-task loss function from the YOLOv5-Seg framework is adopted, including a weighted sum of bounding box regression loss (localization error), object classification loss, and mask segmentation loss. The Adam optimizer is used to train the segmentation model parameters, with an initial learning rate of 0.001. The learning rate is gradually decayed using cosine annealing based on the validation set performance. During training, the weights of positive and negative samples are dynamically adjusted in each epoch (e.g., using FocalLoss to reduce the weight of easily classified samples) to alleviate class imbalance. The entire segmentation model converges after approximately 500 iterations on the training set, and the loss function steadily decreases to a low level (see...). Figure 5 The errors of the training set and the validation set are close, and no obvious overfitting is observed.

[0114] Evaluation Metrics Calculation: The segmentation model is evaluated using a validation set after each training round. Precision, Recall, and F1-score are primarily used to measure segmentation performance. Precision = TP / (TP+FP), Recall = TP / (TP+FN), and F1-score = 2*Precision*Recall / (Precision+Recall). Here, TP, FP, and FN represent the number of true positives, false positives, and false negatives detected, respectively. Precision measures the accuracy of the prediction, Recall measures the completeness of target detection, and F1-score serves as a harmonic mean, balancing Precision and Recall. When either Precision or Recall is too high or too low, the F1-score is appropriately reduced to avoid bias in the segmentation model due to an overemphasis on one metric. Therefore, F1-score is used as one of the key metrics, combined with mAP (mean Average Precision) to comprehensively evaluate the performance of the segmentation model in multi-class, small-target scenarios. In the validation set, the Recall and Precision of the segmentation model of this invention remained above approximately 0.82, and the final F1-score stabilized at around 0.83 (see [link to validation set]). Figure 6 The results show that the segmentation model has a low false alarm rate while ensuring detection completeness, and has high reliability.

[0115] S4, Surface Water Resource Segmentation and Extraction: After preprocessing, the multispectral remote sensing image of the glacier area to be analyzed is input into the final segmentation model trained in S3. The segmentation model is used to detect and segment surface water resource targets in the multispectral remote sensing image, and segmentation mask maps of each target including glaciers, glacial lakes, snow cover, and stony glaciers are obtained. Combining the bounding box coordinates and mask output by the segmentation model, each target instance is located and classified.

[0116] S5, Post-processing of results: Boundary optimization is performed on the initial segmentation results output by the segmentation model. The segmentation mask of each type of target is converted into a vector polygon, and the contour is smoothed to remove jagged noise. Segmentation polygons that are adjacent and belong to the same target are merged to eliminate fragmented segmentation. Isolated false detection areas with an area lower than the preset threshold are deleted. Finally, an optimized multi-category surface water resource distribution map is generated.

[0117] As can be seen from the detailed description of the above embodiments, the method provided by the present invention achieves excellent results in the extraction of surface water resources in modern glacial areas. Based on a high-performance deep learning model and sufficient data processing, the present invention realizes automatic, accurate and efficient segmentation of multiple types of surface water bodies, and can be used for various practical applications such as water resource surveys in high-altitude and cold mountainous areas, glacial lake outburst risk monitoring and climate change impact assessment.

[0118] This invention combines the ViT, CBAM, and SPD-Conv modules to specifically optimize the YOLOv12 model, demonstrating significant performance improvements in remote sensing image segmentation tasks in glacier areas under complex high-altitude terrain conditions. Specific technical effects are reflected in the following aspects:

[0119] ① Strong network recognition capability

[0120] The improved YOLOv12 model (the model of this invention) significantly outperforms existing mainstream semantic segmentation models in terms of segmentation accuracy and comprehensive recognition ability. As shown in the comparative experiments, the model of this invention achieves an F1-score of 0.833, which is about 5.1 percentage points higher than the DeepLabv3+ model, while also achieving a recall of 0.8567, demonstrating optimal balanced performance. In comparison, DeepLab v3+ suffers from insufficient differentiation between glacier boundaries and background, resulting in an F1 score of only about 0.782, while the U-Net model exhibits edge blurring issues in some regions, achieving an F1 score of approximately 0.807 (Table 1). Figure 5 The improved YOLOv12 of this invention, by introducing Transformer and attention mechanisms, can more accurately distinguish adjacent glaciers from the background, especially in the segmentation of glacier edges, narrow glacial lakes, and irregular glacier structures. Figure 2 The results are close to those of finely annotated manually. The above evaluation demonstrates the superior performance of the model of this invention in identifying complex glacial features.

[0121] Table 1. Results of the improved YOLO12 alignment experiment.

[0122]

[0123]

[0124] ② Fast segmentation operation speed

[0125] This invention significantly improves the inference speed of the model while ensuring high accuracy. Under the same hardware environment, the improved YOLOv12 has an average inference time of approximately 19.85 ms for a single image, which is about 4.9 times faster than the U-Net model's 96.45 ms; it is also about twice faster than the DeepLab v3+ model's 39.43 ms (Table 2). This means that the method of this invention has a significant efficiency advantage when processing massive amounts of remote sensing data and can meet the requirements of near real-time monitoring. The high inference speed combined with excellent accuracy makes the model of this invention highly valuable for practical deployment.

[0126] Table 2 Comparison of inference speed between the improved model and mainstream segmentation methods

[0127] Method Name Inference speed / ms U-net 96.45 DeepLabv3+ 39.43 This invention model 19.85

[0128] ③Simultaneous segmentation of multiple targets

[0129] Traditional methods for glacier area extraction often target a single object, requiring separate model training for different objects such as glaciers, glacial lakes, and snow cover, resulting in a cumbersome and inefficient process. This invention significantly improves the model's ability to distinguish between multiple object categories by improving the network structure and introducing modules such as ViT global features and CBAM attention enhancement. Experiments demonstrate that the improved YOLOv12 model can simultaneously segment multiple surface water bodies, including glaciers, glacial lakes, snow cover, and stony glaciers, achieving "one network for multiple uses" (see...). Figure 7 This technology enables the identification and extraction of multiple target categories in a single forward inference, simplifying the processing flow and saving significant computational resources and time. In practical applications, this multi-target joint segmentation capability greatly improves the efficiency and completeness of regional water resource surveys.

[0130] This invention can simultaneously identify and extract multiple surface water resource targets within glacial areas in a single model inference, including but not limited to the glacier body, glacial lakes, snow-covered areas, and stony glacier regions. The segmentation results for each target category can be differentiated and output according to preset classification labels, achieving integrated multi-target joint monitoring. Based on the optimized segmentation results, the spatial attribute information such as area and perimeter of each type of water resource target within the glacial area is statistically analyzed, and the segmentation results from different time phases are compared and analyzed to obtain dynamic change information such as glacier retreat, glacial lake expansion, and snow cover changes, thereby providing a scientific basis for water resource assessment and disaster early warning in high-altitude and cold mountainous areas.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for extracting surface water resource distribution images in glacial areas based on deep learning, characterized in that, Includes the following steps: S1, Multi-source data preprocessing: S11, Band synthesis: Acquire multispectral remote sensing images and digital elevation model data covering the glacier area, synthesize the multispectral remote sensing images into bands, use digital elevation model data to perform topographic correction on the multispectral remote sensing images, unify the multispectral remote sensing images to the same resolution, and calculate the enhanced vegetation index and soil-regulated vegetation index. The glacier area includes glaciers, glacial lakes, snow cover, and stony glaciers. S12, Normalized Difference Index Calculation: Based on the multispectral remote sensing image corrected by S11, the normalized difference index is calculated, and the normalized differential snow index and normalized water index are calculated. S13, Tag Format Conversion: Obtain vector data of surface water resources in glacier areas and convert the vector boundary labels of the glacier water resource vector data into COCO format labels; S14, Image Slicing and Pairing: Cropping multispectral remote sensing images with their corresponding converted labels; S15, Data Augmentation: Using the cropped multispectral remote sensing image as a training set, data augmentation is performed based on the spatial distribution and characteristics of the samples in the training set; S16, Sample screening: Filter the cropped multispectral remote sensing images according to the target area threshold; S17, Radiometric Correction: Radiometric correction is performed on multispectral remote sensing images from different periods, and the band images of the multispectral remote sensing images and the calculated index images are linearly normalized. S2, Build an improved YOLOv12 segmentation model including a backbone network, a neck network, and a detection head: S21, Backbone Network Optimization: CSPDarknet53 is adopted as the backbone. While retaining the original network structure in the first three stages, the ViT global context module is introduced in the fourth stage. S22, Feature Recalibration: Convolutional Block Attention (CBAM) modules are connected in series at the output ends of each stage of the backbone network to perform feature recalibration; S23, Spatial-Depth Transform Downsampling: Replaces the downsampling convolution operation in the original YOLOv12 model network with a combination of spatial-to-depth convolution and standard convolution; S24, Multi-scale segmentation head: Three detection segmentation heads are set at the output of the improved YOLOv12, corresponding to feature layers with different downsampling factors; S3, Model Training and Optimization: Multispectral remote sensing images obtained from multi-source data preprocessing are used as input data, and the converted labels corresponding to the multispectral remote sensing images are used as supervision signals to train the improved YOLOv12 segmentation model. The weighted sum of the bounding box regression loss, classification loss, and segmentation loss in the improved YOLOv12 segmentation model framework is used as the total loss function, and the parameters of the improved YOLOv12 segmentation model are optimized using the stochastic gradient descent algorithm. During the training process, the F1-score evaluation index is introduced to balance precision and recall, and the average precision mAP is combined to monitor the performance of the improved YOLOv12 segmentation model, and the learning rate and positive and negative sample weights are adjusted accordingly. The final segmentation model is output when the F1-score on the validation set no longer shows significant improvement after training. S4, Surface Water Resource Segmentation and Extraction: After preprocessing, the multispectral remote sensing image of the glacier area to be analyzed is input into the final segmentation model trained in S3. The segmentation model is used to detect and segment surface water resource targets in the multispectral remote sensing image, and segmentation mask maps of each target including glaciers, glacial lakes, snow cover, and stony glaciers are obtained. Combining the bounding box coordinates and masks output by the segmentation model, each target instance is located and classified. S5, Post-processing of results: Boundary optimization is performed on the initial segmentation results output by the segmentation model. The segmentation mask of each type of target is converted into a vector polygon, and the contour is smoothed to remove jagged noise. For adjacent segmentation polygons belonging to the same target, fragmented segmentation is eliminated by merging. Isolated false detection areas with an area lower than the preset threshold are deleted. Finally, an optimized multi-category surface water resource distribution map is generated.

2. The method for extracting surface water resource distribution images in glacial areas based on deep learning according to claim 1, characterized in that: In S11, the acquired multispectral remote sensing images include true-color blue, green, and red bands, near-infrared (NIR) bands, and short-wave infrared (SWIR) bands. These multispectral images are then synthesized to form an image containing five spectral bands. The soil-regulated vegetation index (SRV) and enhanced vegetation index (EVI) are calculated as additional channels to characterize the spectral features of glacial lakes and glaciers, providing spectral information input for surface water resources in the glacial region. The formulas for calculating the SRV and EVI are as follows: In the formula, NIR and RED are the near-infrared band reflectance and red band reflectance, respectively; BLUE is the blue band reflectance; C1 is the atmospheric correction parameter for red light; C2 is the atmospheric correction parameter for blue light; and L is the soil adjustment parameter.

3. The method for extracting surface water resource distribution images in glacial areas based on deep learning according to claim 1, characterized in that: In S12, binary maps of the Normalized Water Index and Normalized Differential Snow Index were calculated using corrected multispectral remote sensing images. These maps highlight the boundary and extent changes of liquid water and solid ice and snow, respectively. Preliminary results of the glacial lake boundary and the distribution of glacier / snow cover areas were obtained through threshold segmentation. The calculation methods for the Normalized Water Index and Normalized Differential Snow Index are as follows: In the formula, Green represents the green band, NIR represents the near-infrared band, and SWIR represents the short-wave infrared band. Based on the image in the above formula, a corresponding binary map is generated to obtain the spectral information of the corresponding ground features.

4. The method for extracting surface water resource distribution images in glacial areas based on deep learning according to claim 1, characterized in that: In S13, the vector polygon labels of the targets of glacial lakes, glaciers, and stony glaciers are converted into COCO format text label files required by YOLO-Seg; In S14, the multispectral remote sensing image and its corresponding converted label are cut into small patches according to a fixed size. In S15, random rotation, mirror flip, and salt-and-pepper noise image enhancement operations are performed based on the spatial distribution and characteristics of the samples in the training set to expand the diversity of the training data. In S16, the cropped multispectral remote sensing image patch is filtered based on the target area threshold, and irrelevant samples are removed.

5. The method for extracting surface water resource distribution images in glacial areas based on deep learning according to claim 1, characterized in that: In S21, the ViT global context module is a visual Transformer module. The visual Transformer module is inserted between the third and fourth stages of the backbone network to fuse local convolutional features with long-range dependence information of glacier surface. In S22, the Convolutional Block Attention (CBAM) module includes channel attention and spatial attention, which are executed sequentially on the output features of each stage of the backbone network to recalibrate the feature maps and improve the attention to glaciers and water resources targets.

6. The method for extracting surface water resource distribution images in glacial areas based on deep learning according to claim 1, characterized in that: In S23, all downsampling convolution operations with a stride of 2 in the original YOLOv12 network are replaced with a combination of spatial-to-depth convolution SPD-Conv and standard convolution. Spatial-to-depth convolution SPD-Conv achieves a transformation of halving the spatial resolution and increasing the number of channels by dividing the input feature map into sub-regions and rearranging them to the channel dimension.

7. The method for extracting surface water resource distribution images in glacial areas based on deep learning according to claim 1, characterized in that: In S24, each detection segmentation head simultaneously predicts the target's category, bounding box, and segmentation mask.

8. The method for extracting surface water resource distribution images in glacial areas based on deep learning according to claim 1, characterized in that: In S3, after each training iteration, the precision and recall of the improved YOLOv12 segmentation model on the validation set are calculated, and the F1-score is calculated accordingly. Stop training when the F1-score begins to improve gradually with each training run.

Images