Glacier identification method and device, electronic equipment and storage medium

Abstract

The application provides a glacier recognition method and device, electronic equipment and a storage medium, and relates to the technical field of image processing. The method comprises the following steps: acquiring a glacier image to be recognized; inputting the glacier image into a glacier recognition model to obtain a recognition result output by the glacier segmentation model, wherein the glacier recognition model is obtained by training based on sample glacier images and label data. Through the glacier recognition model, the recognition result can be accurately obtained, and the precision and efficiency of glacier recognition are improved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a glacier recognition method, apparatus, electronic device, and storage medium. Background Technology

[0002] In the field of automated glacier extraction, researchers have explored various computer vision and remote sensing analysis methods, including supervised and unsupervised classification methods, ratio thresholding, object-oriented image analysis (OBIA), and deep learning methods that have been widely used in recent years. Among these, traditional classification and thresholding methods perform well in clean glacier areas, but in glacier areas covered by debris, the reliability of conventional classifiers decreases significantly because terminal moraine deposits obscure the spectral characteristic signals of the glacier.

[0003] With the advancement of deep learning technology, Convolutional Neural Networks (CNNs) are increasingly being applied to glacier segmentation and boundary recognition tasks. For example, U-Net is widely used for glacier calving front identification and glacier boundary extraction, but its limited receptive field leads to insufficient modeling of long-range dependencies in complex backgrounds, easily resulting in incomplete edges and local omissions. Furthermore, glaciers in debris-covered areas often exhibit similar spectral characteristics to the surrounding mountains, increasing segmentation errors and hindering model generalization, ultimately leading to low accuracy in glacier recognition. Summary of the Invention

[0004] This invention provides a glacier identification method, device, electronic device, and storage medium to solve the problem of low accuracy in glacier identification.

[0005] This invention provides a glacier identification method, comprising: Acquire images of the glacier to be identified; The glacier image is input into the glacier recognition model to obtain the recognition result output by the glacier segmentation model. The glacier recognition model is trained based on sample glacier images and label data.

[0006] According to the glacier identification method provided by the present invention, the glacier identification model includes a first feature extraction module, a context enhancement module, a multi-channel attention module, and a second feature extraction module; The glacier image is input into the glacier recognition model to obtain the recognition result output by the glacier segmentation model, including: The glacier image is input into the first feature extraction module to obtain feature maps of different dimensions output by the first feature extraction module. The feature maps of multiple first target dimensions in the feature maps of different dimensions are respectively input to each of the channel attention modules to obtain the local channel feature maps output by each of the channel attention modules; The feature map of the second target dimension in the feature maps of different dimensions is input into the context enhancement module to obtain the global context feature map output by the context enhancement module; The channel feature maps and the global context feature map are input into the second feature extraction module to obtain the recognition result output by the second feature extraction module.

[0007] According to the glacier identification method provided by the present invention, the channel attention module includes a global average pooling unit and a first convolutional unit; The feature maps of multiple first target dimensions from the feature maps of different dimensions are respectively input to each of the channel attention modules to obtain the local channel feature maps output by each of the channel attention modules, including: For each feature map of the first target dimension, the feature map of the first target dimension is input into the global pooling unit in the channel attention module corresponding to the feature map of the first target dimension, so as to obtain the first global description information of each channel output by the global pooling unit; The first global description information of all channels is input into the first convolutional unit to obtain the local channel attention weights of each channel output by the first convolutional unit. Based on the local channel attention weights, the feature map of the first target dimension is channel-weighted to obtain the local channel feature map.

[0008] According to the glacier identification method provided by the present invention, the context enhancement module includes a global pooling unit, a normalization layer and multiple convolutional units; The feature map of the second target dimension from the feature maps of different dimensions is input into the context enhancement module to obtain the global context feature map output by the context enhancement module, including: The feature map of the second target dimension is input into the global pooling unit to obtain the second global description information of each channel output by the global pooling unit; The second global description information of all channels is input into the second convolutional unit to obtain the first convolutional feature map output by the second convolutional unit; The first convolutional feature map is input into the third convolutional unit to obtain the second convolutional feature map output by the third convolutional unit; The feature map of the second target dimension is input into the fourth convolutional unit to obtain the third convolutional feature map output by the fourth convolutional unit; The third convolutional feature map is input into the batch normalization layer to obtain the normalized information output by the batch normalization layer; The normalized information is processed using an activation function to obtain the processed result; The processed result is input into the fifth convolutional unit to obtain the fourth convolutional feature map output by the fifth convolutional unit; Based on the second convolutional feature map and the fourth convolutional feature map, the channel attention weight tensor is determined; Based on the channel attention weight tensor, the feature map of the second target dimension is channel-weighted to obtain the global context feature map.

[0009] According to a glacier identification method provided by the present invention, the step of inputting the feature maps of each channel and the global context feature map into a second feature extraction module to obtain the identification result output by the second feature extraction module includes: The global context feature map is upsampled to obtain a first upsampled feature map; The first channel feature map and the first upsampled feature map in each of the channel feature maps are concatenated to obtain the first concatenated feature map; Perform a convolution operation on the first concatenated feature map to obtain a fifth convolutional feature map; The fifth convolutional feature map is upsampled to obtain a second upsampled feature map; The second channel feature map and the second upsampled feature map in each of the channel feature maps are concatenated to obtain the second concatenated feature map; Perform a convolution operation on the second concatenated feature map to obtain the sixth convolutional feature map; The sixth convolutional feature map is upsampled to obtain the third upsampled feature map; The third channel feature map and the third upsampled feature map in each of the channel feature maps are concatenated to obtain the third concatenated feature map; The third spliced ​​feature map is convolved to obtain the recognition result.

[0010] According to the glacier identification method provided by the present invention, the loss function L during the training process of the glacier identification model is expressed as: (1) in, Indicates the weighting coefficient. Let represent the cross-entropy loss, p represent the predicted class probability, and y represent the class probability of the label data in one-hot format. Indicates mutual information loss; (2) (3) Where N represents the total number of sample glacier images in each batch, H represents the height of the sample glacier image, and W represents the width of the sample glacier image. Let represent the probability that the pixel position (i,j) corresponding to the nth sample glacier image, expressed in one-hot format, belongs to class m. Let represent the predicted class probability of the nth sample glacier image belonging to class m at pixel position (i,j). This represents the smallest constant used for numerical stability.

[0011] The present invention also provides a glacier identification device, comprising: The acquisition module is used to acquire images of the glacier to be identified. The recognition module is used to input the glacier image into the glacier recognition model and obtain the recognition result output by the glacier segmentation model. The glacier recognition model is trained based on sample glacier images and label data.

[0012] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any of the glacier recognition methods described above.

[0013] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the glacier identification method as described above.

[0014] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements any of the glacier recognition methods described above.

[0015] The glacier identification method, apparatus, electronic device, and storage medium provided by this invention acquire glacier images to be identified; input the glacier images into a glacier identification model to obtain the identification result output by the glacier segmentation model, wherein the glacier identification model is trained based on sample glacier images and label data. Through the glacier identification model, accurate identification results can be obtained, improving the accuracy and efficiency of glacier identification. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the glacier identification method provided by the present invention.

[0018] Figure 2 This is a schematic diagram of the channel attention module.

[0019] Figure 3 This is a schematic diagram of the context enhancement module provided by the present invention.

[0020] Figure 4 This is a schematic diagram of the structure of the glacier identification model provided by the present invention.

[0021] Figure 5 This is a schematic diagram of the structure of the glacier identification device provided by the present invention.

[0022] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0024] The following is combined with Figures 1-4 The glacier identification method of the present invention is described.

[0025] Figure 1 This is a flowchart illustrating the glacier identification method provided by the present invention, as shown below. Figure 1 As shown, the method includes the following steps 101-102.

[0026] Step 101: Obtain the image of the glacier to be identified.

[0027] Step 102: Input the glacier image into the glacier recognition model to obtain the recognition result output by the glacier segmentation model. The glacier recognition model is trained based on sample glacier images and label data.

[0028] It should be noted that the glacier identification method provided by this invention is applicable to glacier identification scenarios. The executing entity of the glacier identification method provided by this invention can be the glacier identification device provided by this invention, such as an electronic device, or a control module in the glacier identification device for executing the glacier identification method.

[0029] Specifically, the glacier identification model is G-LEAU-Net, which is an improvement upon the U-Net basic network. The identification results are seasonal snow-covered glaciers and / or moraine glaciers. Inputting glacier images into the glacier identification model yields the identification results output by the glacier segmentation model; the glacier identification model is trained based on sample glacier images and label data.

[0030] The glacier identification method provided by this invention involves acquiring an image of a glacier to be identified; inputting the glacier image into a glacier identification model to obtain the identification result output by the glacier segmentation model, wherein the glacier identification model is trained based on sample glacier images and label data. Through the glacier identification model, accurate identification results can be obtained, improving the accuracy and efficiency of glacier identification.

[0031] Optionally, the loss function L during the training process of the glacier recognition model is expressed as: (1) in, Indicates the weighting coefficient. Let represent the cross-entropy loss, p represent the predicted class probability, and y represent the class probability of the label data in one-hot format. Indicates mutual information loss; (2) (3) Where N represents the total number of sample glacier images in each batch, H represents the height of the sample glacier image, and W represents the width of the sample glacier image. Let represent the probability that the pixel position (i,j) corresponding to the nth sample glacier image, expressed in one-hot format, belongs to class m. Let represent the predicted class probability of the nth sample glacier image belonging to class m at pixel position (i,j). This represents the smallest constant used for numerical stability.

[0032] Specifically, in the training process of the glacier recognition model, this invention proposes a Class-Weighted Mutual Information (CWMI) loss function to improve the model's discrimination performance in scenarios involving spectral aliasing and class imbalance. This invention introduces CWMI on top of the basic cross-entropy loss to enhance the consistency between the predicted results and the real label data. First, the recognition results output by the model are softmaxed to obtain the class probability p, and the real label data is converted into one-hot encoded form y. Based on this, the consistency constraint term is calculated, i.e. Cross-entropy Loss It is used to measure the difference between the model's predicted probability distribution and the real label data, and is one of the most commonly used supervised losses in classification and semantic segmentation tasks.

[0033] The CWMI loss effectively addresses the identification bias caused by uneven class distribution and spectral aliasing. Through mutual information constraints, the model enhances class discriminative power in regions of high spectral similarity, resulting in more continuous and smoother output in glacier boundary regions. This loss function significantly improves the model's segmentation accuracy and generalization performance under complex terrain, lighting, and moraine cover conditions.

[0034] Optionally, the glacier recognition model includes a first feature extraction module, a context enhancement module, a multi-channel attention module, and a second feature extraction module; the specific implementation of step 102 includes: The glacier image is input into the first feature extraction module to obtain feature maps of different dimensions output by the first feature extraction module; multiple first target dimension feature maps from the different dimension feature maps are respectively input into each of the channel attention modules to obtain local channel feature maps output by each of the channel attention modules; the second target dimension feature maps from the different dimension feature maps are input into the context enhancement module to obtain a global context feature map output by the context enhancement module; the channel feature maps and the global context feature map are input into the second feature extraction module to obtain the recognition result output by the second feature extraction module.

[0035] Specifically, the first feature extraction module extracts feature maps of different dimensions. The LightCE (Light Context Enhancement) module is a lightweight context enhancement module based on the Squeeze-and-Excitation (SE) mechanism, used to enhance the capture of global context information and feature representation capabilities across multiple bands. The Efficient Channel Attention (ECA) module enhances channel selection and fine-grained feature modeling capabilities during the feature fusion stage. The second feature extraction module upsamples and convolves the feature maps to obtain the final recognition result. During downsampling, LightCE enhances the semantic consistency and discriminative ability of high-level features by fusing global context and local information; ECA performs channel-level fine-tuning of high-resolution features at the connection layer, highlighting details related to the target boundary and suppressing redundant noise. The two modules work together to enable the model to maintain accurate boundary detail representation while taking into account global semantic understanding, thereby jointly improving feature representation and boundary recognition capabilities.

[0036] Glacier images ( The image is input to the first feature extraction module, which performs two convolutions and an activation function (ReLU) on the glacier image to obtain the feature map extracted in the first stage. ); and then the first feature map of the first stage ( Max pooling is performed to obtain the second feature map of the first stage. ); and then the second feature map ( The second stage involves two convolutions and the ReLU activation function to obtain the third feature map extracted in the second stage. ); and then the third feature map of the second stage ( Max pooling is performed to obtain the fourth feature map in the second stage. ); and then the fourth feature map ( The third stage extracts the fifth feature map by performing two convolutions and the ReLU activation function. Then, the fifth feature map extracted in the third stage ( Max pooling is performed to obtain the sixth feature map in the third stage. ).

[0037] Multiple feature maps of the first target dimension from different dimension feature maps (i.e. , , The features are input into each channel attention module to obtain the local channel feature maps output by each channel attention module; the feature maps of the second target dimension in the feature maps of different dimensions are then compared. The input is fed into the context enhancement module, and the global context feature map output by the context enhancement module is obtained. The feature maps of each channel and the global context feature map are input into the second feature extraction module to obtain the recognition results output by the second feature extraction module.

[0038] Optionally, the channel attention module includes a global average pooling unit and a first convolutional unit; multiple feature maps of the first target dimension from the feature maps of different dimensions are respectively input to each of the channel attention modules to obtain the local channel feature maps output by each of the channel attention modules, including: For each feature map of the first target dimension, the feature map of the first target dimension is input into the global pooling unit in the channel attention module corresponding to the feature map of the first target dimension to obtain the first global description information of each channel output by the global pooling unit; the first global description information of all channels is input into the first convolution unit to obtain the local channel attention weights output by the first convolution unit; based on the local channel attention weights, the feature map of the first target dimension is channel-weighted to obtain the local channel feature map.

[0039] Specifically, Figure 2 This is a schematic diagram of the channel attention module, such as... Figure 2 As shown, the Channel Attention Module (ECA) includes a Global Average Pooling Unit (GAP) and a first convolutional unit.

[0040] For each feature map of the first target dimension, the feature map X of the first target dimension is input into the global pooling unit in the channel attention module corresponding to the feature map of the first target dimension, so as to obtain the first global description information of each channel output by the global pooling unit. , represented as , H represents the pixel value corresponding to channel c of pixel (i,j) in the feature map, H represents the height of the feature map, and W represents the width of the feature map.

[0041] First global description information for all channels ( The input is fed into the first convolutional unit to obtain the local channel attention weights output by the first convolutional unit. , represented as , This represents a one-dimensional convolution kernel of length k for the first convolutional unit. Related to the adaptive number of channels C, for example , and b are hyperparameters, and the local channel attention weights for each channel are . .

[0042] Based on local channel attention weights Channel weighting is applied to the feature map of the first target dimension to obtain the local channel feature map. , represented as , where the local channel feature map of each channel c .

[0043] The ECA module achieves efficient channel dependency modeling through local cross-channel one-dimensional convolutions. Specifically, ECA models only the dependencies between adjacent channels using local one-dimensional convolutions, achieving fine-grained characterization of channel importance without introducing additional dimensionality compression. The ECA module does not perform channel dimensionality reduction, fully preserving channel feature information and efficiently learning dependencies between adjacent channels through local cross-channel convolutions. This mechanism significantly enhances the model's response to glacier boundary and texture features, effectively mitigating spectral aliasing, boundary blurring, and misclassification problems in debris-covered regions. Furthermore, ECA has extremely low parameter counts (only about 80 additional parameters on ResNet-50), improving segmentation accuracy while maintaining high inference speed. When used in conjunction with LightCE, the two complement each other; LightCE provides global contextual information, while ECA achieves fine-grained channel-level attention, jointly enhancing the model's feature representation and boundary recognition capabilities.

[0044] Optionally, the context enhancement module includes a global pooling unit, a normalization layer, and multiple convolutional units; the feature map of the second target dimension in the feature maps of different dimensions is input into the context enhancement module to obtain the global context feature map output by the context enhancement module, including: The feature map of the second target dimension is input into the global pooling unit to obtain the second global description information of each channel output by the global pooling unit; the second global description information of all channels is input into the second convolutional unit to obtain the first convolutional feature map output by the second convolutional unit; the first convolutional feature map is input into the third convolutional unit to obtain the second convolutional feature map output by the third convolutional unit; the feature map of the second target dimension is input into the fourth convolutional unit to obtain the third convolutional feature map output by the fourth convolutional unit; the third convolutional feature map is input into the batch normalization layer to obtain the normalized information output by the batch normalization layer; the normalized information is processed using an activation function to obtain the processed result; the processed result is input into the fifth convolutional unit to obtain the fourth convolutional feature map output by the fifth convolutional unit; based on the second convolutional feature map and the fourth convolutional feature map, the channel attention weight tensor is determined; based on the channel attention weight tensor, the feature map of the second target dimension is channel-weighted to obtain the global context feature map.

[0045] Specifically, Figure 3 This is a schematic diagram of the structure of the context enhancement module provided by the present invention, as shown below. Figure 3 As shown, the Context Enhancement Module (LightCE) includes a global pooling unit, a normalization layer (BN), and multiple convolutional units (Conv), each of which includes an activation function.

[0046] The feature map of the second target dimension ( ) The input is fed into the global pooling unit to obtain the second global description information of each channel output by the global pooling unit. , represented as ;Second global description information for all channels ( The input is fed into the second convolutional unit (the convolutional kernel is W1). for The first convolutional unit (with convolution kernel W2) performs convolution and activation function operations on the second global description information of all channels to obtain the first convolutional feature map output by the second convolutional unit. This first convolutional feature map is then input into the third convolutional unit (with kernel W2). The third convolutional unit performs convolution and activation function operations on the first convolutional feature map to obtain the second convolutional feature map output by the third convolutional unit, denoted as... , Represents the ReLU activation function. This represents the Sigmoid function.

[0047] The feature map of the second target dimension ( ) The input is fed into the fourth convolutional unit (the convolutional kernel is W3, and W3 is a depthwise convolutional kernel (groups=C)). The fourth convolutional unit processes the feature map of the second target dimension. ) Convolution is performed to obtain the third convolutional feature map output by the fourth convolutional unit; this third convolutional feature map is then input into the batch normalization layer (BN) to obtain the normalized information output by the batch normalization layer; the normalized information is processed using an activation function to obtain the processed result; the processed result is then input into the fifth convolutional unit (with kernel W4, W4 being...). (using the convolution kernel), to obtain the fourth convolution feature map output by the fifth convolution unit. , represented as .

[0048] The second convolutional feature map and the fourth convolutional feature map Adding them together yields the channel attention weight tensor. , represented as Based on channel attention weight tensor Channel weighting is applied to the feature map of the second target dimension to obtain the global context feature map. , represented as Among them, the global context feature map of each channel c. .

[0049] The LightCE module obtains full-image-level statistical information through global pooling and performs fusion modeling at the channel level, allowing weights to reflect the importance of different channels at the overall semantic level, thus providing global contextual constraints. By capturing global contextual information across multiple bands, the LightCE module significantly expands the network's receptive field and enhances the model's ability to identify the main water feature region. Simultaneously, the LightCE module maintains a lightweight structure and low computational cost, improving feature representation and discriminative capabilities without increasing the number of parameters, effectively alleviating the problem of insufficient context awareness in traditional U-Net models.

[0050] Optionally, the step of inputting each of the channel feature maps and the global context feature map into the second feature extraction module to obtain the recognition result output by the second feature extraction module includes: The global context feature map is upsampled to obtain a first upsampled feature map; the first channel feature map and the first upsampled feature map in each channel feature map are concatenated to obtain a first concatenated feature map; a convolution operation is performed on the first concatenated feature map to obtain a fifth convolutional feature map; the fifth convolutional feature map is upsampled to obtain a second upsampled feature map; the second channel feature map and the second upsampled feature map in each channel feature map are concatenated to obtain a second concatenated feature map; a convolution operation is performed on the second concatenated feature map to obtain a sixth convolutional feature map; the sixth convolutional feature map is upsampled to obtain a third upsampled feature map; the third channel feature map in each channel feature map and the third upsampled feature map are concatenated to obtain a third concatenated feature map; a convolution operation is performed on the third concatenated feature map to obtain the recognition result.

[0051] Specifically, for the global context feature map ( ) Upsampling is performed to obtain the first upsampled feature map. ); the first channel feature map in each channel feature map ( ) and the first upsampled feature map ( The first concatenated feature map is obtained by concatenating the first concatenated feature map with the second concatenated feature map. A convolution operation is then performed on the first concatenated feature map to obtain the fifth convolutional feature map, where the convolution operation includes two convolutions and an activation function (ReLU). Finally, the fifth convolutional feature map is upsampled to obtain the second upsampled feature map. ); and the second channel feature map of each channel feature map ( ) and the second upsampled feature map ( The two concatenated feature maps are concatenated to obtain the second concatenated feature map; a convolution operation is performed on the second concatenated feature map to obtain the sixth convolutional feature map, where the convolution operation includes two convolutions and an activation function (ReLU); the sixth convolutional feature map is upsampled to obtain the third upsampled feature map. ); and the third channel feature map in each channel feature map ( ) and the third upsampled feature map ( The features are concatenated to obtain a third concatenated feature map; a convolution operation is then performed on the third concatenated feature map to obtain the recognition result. The convolution operation includes two convolutions and an activation function (ReLU).

[0052] Figure 4 This is a schematic diagram of the structure of the glacier identification model provided by the present invention, as shown below. Figure 4 As shown, the glacier recognition model includes a first feature extraction module, a context enhancement module (LightCE), a multi-channel attention module (ECA), and a second feature extraction module; wherein, the glacier image ( The image is input to the first feature extraction module, which performs two convolutions and an activation function (ReLU) on the glacier image to obtain the feature map extracted in the first stage. ); and then the first feature map of the first stage ( Max pooling is performed to obtain the second feature map of the first stage. ); and then the second feature map ( The second stage involves two convolutions and the ReLU activation function to obtain the third feature map extracted in the second stage. ); and then the third feature map of the second stage ( Max pooling is performed to obtain the fourth feature map in the second stage. ); and then the fourth feature map ( The third stage extracts the fifth feature map by performing two convolutions and the ReLU activation function. Then, the fifth feature map extracted in the third stage ( Max pooling is performed to obtain the sixth feature map in the third stage. ). Multiple feature maps of the first target dimension from different dimension feature maps (i.e. , , The features are input into each channel attention module to obtain the local channel feature maps output by each channel attention module; the feature maps of the second target dimension in the feature maps of different dimensions are then compared. The input is fed into the context enhancement module, and the global context feature map output by the context enhancement module is obtained. ).

[0053] For global context feature maps ( Upsampling is performed to obtain the first upsampled feature map. ); the first channel feature map in each channel feature map ( ) and the first upsampled feature map ( The first concatenated feature map is obtained by concatenating the first concatenated feature map with the second concatenated feature map. A convolution operation is then performed on the first concatenated feature map to obtain the fifth convolutional feature map, where the convolution operation includes two convolutions and an activation function (ReLU). Finally, the fifth convolutional feature map is upsampled to obtain the second upsampled feature map. ); and the second channel feature map of each channel feature map ( ) and the second upsampled feature map ( The two concatenated feature maps are concatenated to obtain the second concatenated feature map; a convolution operation is performed on the second concatenated feature map to obtain the sixth convolutional feature map, where the convolution operation includes two convolutions and an activation function (ReLU); the sixth convolutional feature map is upsampled to obtain the third upsampled feature map. ); and the third channel feature map in each channel feature map ( ) and the third upsampled feature map ( The features are concatenated to obtain a third concatenated feature map; a convolution operation is then performed on the third concatenated feature map to obtain the recognition result. The convolution operation includes two convolutions and an activation function (ReLU).

[0054] This invention systematically improves the performance of a glacier recognition model from three levels: context enhancement, channel attention optimization, and loss function, by introducing LightCE, ECA, and the loss function (CWMI) used in the training process. This model not only possesses strong feature representation capabilities and boundary discrimination accuracy, but also features lightweight design, high efficiency, and good cross-regional generalization ability, providing a reliable and scalable technical solution for the automatic identification of debris-covered glaciers.

[0055] The G-LEAU-Net model proposed in this application integrates a lightweight context enhancement module (LightCE), an efficient channel attention mechanism (ECA), and a loss function (CWMI) on the basis of the traditional U-Net framework, demonstrating significant technical advantages in the task of automatic glacier recognition and boundary extraction.

[0056] (i) Improving Glacier Identification Accuracy. Experiments on Landsat and Gaofen-6PMS remote sensing images demonstrate that G-LEAU-Net achieves a mean intersection over union (MIOU) of 93.18% and 98.69%, respectively, significantly outperforming existing mainstream models (such as U-Net and DeepLabV3+). The model can accurately identify glacier boundaries in debris-covered areas and complex terrains, with smoother, more continuous identification results and significantly improved boundary consistency.

[0057] (ii) Strong robustness and environmental adaptability. G-LEAU-Net demonstrates good stability and generalization on multi-source and multi-temporal remote sensing data. Regardless of different resolutions, lighting conditions, or areas with high terrain complexity, the model can maintain high-precision segmentation results.

[0058] (III) Lightweight Design and Efficient Deployment Capabilities. Both the LightCE and ECA modules adopt a lightweight structural design, with significantly lower parameter and computational costs compared to traditional attention mechanisms. The overall model boasts fast inference speed and low computational cost, enabling deployment and operation on ordinary GPUs or edge computing devices, making it suitable for large-scale, batch processing of glacier remote sensing images and automated monitoring tasks.

[0059] As can be seen, the glacier identification model provided in this application achieves an average intersection-union ratio of over 93% in multi-source image data, significantly improving the accuracy of glacier boundary identification; it maintains stable performance and strong robustness even in complex terrain, lighting variations, and spectral aliasing scenarios; the network structure is lightweight, with high training and inference efficiency and high computational efficiency; it is applicable to multi-source remote sensing images and different glacier environments, and has strong generalization ability.

[0060] The glacier identification device provided by the present invention is described below. The glacier identification device described below and the glacier identification method described above can be referred to in correspondence.

[0061] Figure 5 This is a schematic diagram of the structure of the glacier identification device provided by the present invention, as shown below. Figure 5 As shown, the glacier identification device 500 includes: an acquisition module 501 and an identification module 502; wherein, The acquisition module 501 is used to acquire images of the glacier to be identified; The recognition module 502 is used to input the glacier image into the glacier recognition model and obtain the recognition result output by the glacier segmentation model. The glacier recognition model is trained based on sample glacier images and label data.

[0062] The glacier recognition device provided by this invention acquires an image of a glacier to be recognized; the glacier image is input into a glacier recognition model to obtain the recognition result output by the glacier segmentation model, which is trained based on sample glacier images and label data. Through the glacier recognition model, accurate recognition results can be obtained, improving the accuracy and efficiency of glacier recognition.

[0063] Optionally, the glacier recognition model includes a first feature extraction module, a context enhancement module, a multi-channel attention module, and a second feature extraction module; the recognition module 502 is specifically used for: The glacier image is input into the first feature extraction module to obtain feature maps of different dimensions output by the first feature extraction module. The feature maps of multiple first target dimensions in the feature maps of different dimensions are respectively input to each of the channel attention modules to obtain the local channel feature maps output by each of the channel attention modules; The feature map of the second target dimension in the feature maps of different dimensions is input into the context enhancement module to obtain the global context feature map output by the context enhancement module; The channel feature maps and the global context feature map are input into the second feature extraction module to obtain the recognition result output by the second feature extraction module.

[0064] Optionally, the channel attention module includes a global average pooling unit and a first convolutional unit; the recognition module 502 is further configured to: For each feature map of the first target dimension, the feature map of the first target dimension is input into the global pooling unit in the channel attention module corresponding to the feature map of the first target dimension, so as to obtain the first global description information of each channel output by the global pooling unit; The first global description information of all channels is input into the first convolutional unit to obtain the local channel attention weights output by the first convolutional unit; Based on the local channel attention weights, the feature map of the first target dimension is channel-weighted to obtain the local channel feature map.

[0065] Optionally, the context enhancement module includes a global pooling unit, a normalization layer, and multiple convolutional units; the recognition module 502 is further configured to: The feature map of the second target dimension is input into the global pooling unit to obtain the second global description information of each channel output by the global pooling unit; The second global description information of all channels is input into the second convolutional unit to obtain the first convolutional feature map output by the second convolutional unit; The first convolutional feature map is input into the third convolutional unit to obtain the second convolutional feature map output by the third convolutional unit; The feature map of the second target dimension is input into the fourth convolutional unit to obtain the third convolutional feature map output by the fourth convolutional unit; The third convolutional feature map is input into the batch normalization layer to obtain the normalized information output by the batch normalization layer; The normalized information is processed using an activation function to obtain the processed result; The processed result is input into the fifth convolutional unit to obtain the fourth convolutional feature map output by the fifth convolutional unit; Based on the second convolutional feature map and the fourth convolutional feature map, the channel attention weight tensor is determined; Based on the channel attention weight tensor, the feature map of the second target dimension is channel-weighted to obtain the global context feature map.

[0066] Optionally, the identification module 502 is further configured to: The global context feature map is upsampled to obtain a first upsampled feature map; The first channel feature map and the first upsampled feature map in each of the channel feature maps are concatenated to obtain the first concatenated feature map; Perform a convolution operation on the first concatenated feature map to obtain a fifth convolutional feature map; The fifth convolutional feature map is upsampled to obtain a second upsampled feature map; The second channel feature map and the second upsampled feature map in each of the channel feature maps are concatenated to obtain the second concatenated feature map; Perform a convolution operation on the second concatenated feature map to obtain the sixth convolutional feature map; The sixth convolutional feature map is upsampled to obtain the third upsampled feature map; The third channel feature map and the third upsampled feature map in each of the channel feature maps are concatenated to obtain the third concatenated feature map; The third spliced ​​feature map is convolved to obtain the recognition result.

[0067] Optionally, the loss function L during the training process of the glacier recognition model is expressed as: (1) in, Indicates the weighting coefficient. Let represent the cross-entropy loss, p represent the predicted class probability, and y represent the class probability of the label data in one-hot format. Indicates mutual information loss; (2) (3) Where N represents the total number of sample glacier images in each batch, H represents the height of the sample glacier image, and W represents the width of the sample glacier image. Let represent the probability that the pixel position (i,j) corresponding to the nth sample glacier image, expressed in one-hot format, belongs to class m. Let represent the predicted class probability of the nth sample glacier image belonging to class m at pixel position (i,j). This represents the smallest constant used for numerical stability.

[0068] Figure 6 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 6 As shown, the electronic device 600 may include a processor 610, a communications interface 620, a memory 630, and a communication bus 640. The processor 610, communications interface 620, and memory 630 communicate with each other via the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute a glacier recognition method. This method includes: acquiring a glacier image to be recognized; inputting the glacier image into a glacier recognition model to obtain the recognition result output by the glacier segmentation model, wherein the glacier recognition model is trained based on sample glacier images and label data.

[0069] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0070] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the glacier recognition method provided by the above methods. The method includes: acquiring a glacier image to be recognized; inputting the glacier image into a glacier recognition model to obtain the recognition result output by the glacier segmentation model, wherein the glacier recognition model is trained based on sample glacier images and label data.

[0071] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When executed by a processor, the computer program is implemented to perform the glacier recognition method provided by the above methods. The method includes: acquiring a glacier image to be recognized; inputting the glacier image into a glacier recognition model to obtain a recognition result output by the glacier segmentation model, wherein the glacier recognition model is trained based on sample glacier images and label data.

[0072] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0073] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0074] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying glaciers, characterized in that, include: Acquire images of the glacier to be identified; The glacier image is input into the glacier recognition model to obtain the recognition result output by the glacier segmentation model. The glacier recognition model is trained based on sample glacier images and label data.

2. The glacier identification method according to claim 1, characterized in that, The glacier recognition model includes a first feature extraction module, a context enhancement module, a multi-channel attention module, and a second feature extraction module; The glacier image is input into the glacier recognition model to obtain the recognition result output by the glacier segmentation model, including: The glacier image is input into the first feature extraction module to obtain feature maps of different dimensions output by the first feature extraction module. The feature maps of multiple first target dimensions in the feature maps of different dimensions are respectively input to each of the channel attention modules to obtain the local channel feature maps output by each of the channel attention modules; The feature map of the second target dimension in the feature maps of different dimensions is input into the context enhancement module to obtain the global context feature map output by the context enhancement module; The channel feature maps and the global context feature map are input into the second feature extraction module to obtain the recognition result output by the second feature extraction module.

3. The glacier identification method according to claim 2, characterized in that, The channel attention module includes a global average pooling unit and a first convolutional unit; The feature maps of multiple first target dimensions from the feature maps of different dimensions are respectively input to each of the channel attention modules to obtain the local channel feature maps output by each of the channel attention modules, including: For each feature map of the first target dimension, the feature map of the first target dimension is input into the global pooling unit in the channel attention module corresponding to the feature map of the first target dimension, so as to obtain the first global description information of each channel output by the global pooling unit; The first global description information of all channels is input into the first convolutional unit to obtain the local channel attention weights output by the first convolutional unit; Based on the local channel attention weights, the feature map of the first target dimension is channel-weighted to obtain the local channel feature map.

4. The glacier identification method according to claim 2, characterized in that, The context enhancement module includes a global pooling unit, a normalization layer, and multiple convolutional units; The feature map of the second target dimension from the feature maps of different dimensions is input into the context enhancement module to obtain the global context feature map output by the context enhancement module, including: The feature map of the second target dimension is input into the global pooling unit to obtain the second global description information of each channel output by the global pooling unit; The second global description information of all channels is input into the second convolutional unit to obtain the first convolutional feature map output by the second convolutional unit; The first convolutional feature map is input into the third convolutional unit to obtain the second convolutional feature map output by the third convolutional unit; The feature map of the second target dimension is input into the fourth convolutional unit to obtain the third convolutional feature map output by the fourth convolutional unit; The third convolutional feature map is input into the batch normalization layer to obtain the normalized information output by the batch normalization layer; The normalized information is processed using an activation function to obtain the processed result; The processed result is input into the fifth convolutional unit to obtain the fourth convolutional feature map output by the fifth convolutional unit; Based on the second convolutional feature map and the fourth convolutional feature map, the channel attention weight tensor is determined; Based on the channel attention weight tensor, the feature map of the second target dimension is channel-weighted to obtain the global context feature map.

5. The glacier identification method according to claim 2, characterized in that, The step of inputting each of the channel feature maps and the global context feature map into the second feature extraction module to obtain the recognition result output by the second feature extraction module includes: The global context feature map is upsampled to obtain a first upsampled feature map; The first channel feature map and the first upsampled feature map in each of the channel feature maps are concatenated to obtain the first concatenated feature map; Perform a convolution operation on the first concatenated feature map to obtain a fifth convolutional feature map; The fifth convolutional feature map is upsampled to obtain a second upsampled feature map; The second channel feature map and the second upsampled feature map in each of the channel feature maps are concatenated to obtain the second concatenated feature map; Perform a convolution operation on the second concatenated feature map to obtain the sixth convolutional feature map; The sixth convolutional feature map is upsampled to obtain the third upsampled feature map; The third channel feature map and the third upsampled feature map in each of the channel feature maps are concatenated to obtain the third concatenated feature map; The third spliced ​​feature map is convolved to obtain the recognition result.

6. The glacier identification method according to any one of claims 1 to 5, characterized in that, The loss function L during the training process of the glacier recognition model is expressed as: （1） in, Indicates the weighting coefficient. Let represent the cross-entropy loss, p represent the predicted class probability, and y represent the class probability of the label data in one-hot encoding. Indicates mutual information loss; （2） （3） Where N represents the total number of sample glacier images in each batch, H represents the height of the sample glacier image, and W represents the width of the sample glacier image. Let represent the probability that the pixel position (i,j) of the nth sample glacier image, expressed in one-hot format, belongs to class m. Let represent the predicted class probability of the nth sample glacier image belonging to class m at pixel position (i,j). This represents the smallest constant used for numerical stability.

7. A glacier identification device, characterized in that, include: The acquisition module is used to acquire images of the glacier to be identified. The recognition module is used to input the glacier image into the glacier recognition model and obtain the recognition result output by the glacier segmentation model. The glacier recognition model is trained based on sample glacier images and label data.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the glacier identification method as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the glacier identification method as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the glacier identification method as described in any one of claims 1 to 6.

Images