A glacier recognition model modeling method based on a deep learning network and semi-supervised learning
By using a glacier identification model based on deep learning networks and semi-supervised learning, and leveraging high-resolution remote sensing images and digital elevation models, glacier changes can be automatically identified. This solves the problems of traditional methods being labor-intensive, resource-intensive, and having limited data, and enables efficient and accurate glacier monitoring and global climate change analysis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SOUTHWEST JIAOTONG UNIV
- Filing Date
- 2022-12-10
- Publication Date
- 2026-05-29
AI Technical Summary
Existing technologies are insufficient for efficiently and accurately monitoring global glacier changes through remote sensing imagery. Traditional field survey methods are labor-intensive, resource-intensive, and limited in data, failing to fully reflect the trend of glacier retreat.
A glacier identification model employing deep learning networks and semi-supervised learning is developed. Utilizing high-resolution remote sensing imagery and digital elevation model data, combined with deep residual networks and semi-supervised learning algorithms, the model automatically identifies glacier changes and constructs a glacier identification model.
It has enabled efficient and accurate monitoring of glacier area changes, reduced labor costs, improved the timeliness and objectivity of identification and detection, and provided data support for global climate change.
Smart Images
Figure CN115796045B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of surveying and mapping technology, specifically to a glacier identification modeling method based on deep learning networks and semi-supervised learning. Background Technology
[0002] The stable state of the Earth's environment refers to the fact that all five spheres maintain their relatively normal state. The cryosphere is one of the five spheres of the Earth's climate system, composed of various glaciers, cirques, glacial erosion, and permafrost, with glaciers being the most abundant component. However, due to the continuous accumulation of the greenhouse effect in recent years leading to global warming, the loss of global glaciers and glacial material has accelerated dramatically, bringing many potential crises to human life in the future. This phenomenon is also reflected in our daily lives, such as the significant year-on-year increase in summer temperatures. According to relevant statistics from the Meteorological Bureau, from 1951 to 2021, the average annual surface temperature in my country showed a significant upward trend, with a warming rate of 0.26℃ / 10 years. The biggest impact of rising temperatures on humans is the change in perceived temperature and the resulting symptoms such as heatstroke. The impact of rising temperatures on glaciers is to accelerate their melting, leading to a shrinkage and reduction in their area. Except for the relatively stable glaciers in the Karakoram Mountains, glaciers in other regions of China are showing a state of retreat and thinning. It is evident that global warming is accelerating the shrinking of global glaciers, posing threats to human health and well-being, such as short-term glacial melting, long-term glacial lake rise, and sea-level rise. Historically, glacier observation and research began relatively early; over a century ago, several countries were already conducting glacier monitoring. However, before the widespread adoption of remote sensing imagery, traditional glacier measurement methods primarily relied on field surveys. This method not only consumed significant human and material resources but also failed to provide comprehensive, dynamic monitoring of global glaciers based solely on limited field data. The lack of completely objective and accurate data to concretely reflect the retreat trends, speed, and extent of glaciers limited the applicability of the obtained glacier change trends and analytical conclusions in assessing global climate change.
[0003] With the development and maturation of remote sensing images and deep learning technology, it has become a trend to entrust the work of human research, identification and search to computers trained by deep learning models to complete it in batches and systematically. This can effectively save manpower and reduce costs, and improve the timeliness and objectivity of identification and detection results. However, there is currently no reliable method for identifying glaciers based on deep learning technology. Summary of the Invention
[0004] To address the problems existing in the prior art, this invention provides a glacier identification modeling method based on deep learning networks and semi-supervised learning. The purpose is to automatically identify and detect changes in glacier area from high-resolution remote sensing images, and to entrust the work of manual research, identification, and search to computers trained with deep learning models to complete the work in batches and systematically, thus providing a basis for analyzing the rate of glacier melting and monitoring global climate change.
[0005] The technical solution adopted in this invention is as follows:
[0006] A glacier identification modeling method based on deep learning networks and semi-supervised learning includes the following steps:
[0007] S1: Select high-resolution remote sensing images of multiple glaciers before their melting period and corresponding digital elevation model data to establish a glacier identification remote sensing image dataset. Segment the glacier identification remote sensing image dataset to obtain labeled and unlabeled data.
[0008] S2: Construct a deep learning convolutional neural network model based on a deep residual network using a deep residual network, and pre-train the deep learning convolutional neural network model based on the deep residual network using the labeled data obtained in S1.
[0009] S3: A semi-supervised learning model is used to perform consistency regularization on labeled and unlabeled data to obtain a semi-supervised learning dataset. The semi-supervised learning dataset is then combined with the deep learning convolutional neural network model based on deep residual network obtained in S2 to construct a glacier recognition model.
[0010] Preferably, the process of establishing the glacier identification remote sensing image dataset in S1 is as follows:
[0011] The selected high-resolution remote sensing images of multiple glaciers before their melting period were registered and resampled with the corresponding digital elevation model data to ensure that the high-resolution remote sensing images of multiple glaciers before their melting period and the corresponding digital elevation model data have the same number of pixels.
[0012] Preferably, the process of segmenting the glacier identification remote sensing image dataset in S1 to obtain labeled and unlabeled data is as follows:
[0013] Using ArcGIS software, glacier target areas in high-resolution remote sensing images of multiple glaciers before their melting period were labeled, resulting in labeled and unlabeled data. The labeled and unlabeled data were saved separately, with a ratio of 2:8 between labeled and unlabeled data.
[0014] Preferably, the process of constructing a deep learning convolutional neural network model based on a deep residual network in step S2 is as follows:
[0015] The RGB information of the labeled data obtained in S1 is input into the first convolutional layer of the deep residual network. Convolution, regularization, activation function and max pooling are performed on the input to change the number of channels and size of the original image, so as to obtain a deep learning convolutional neural network model based on the deep residual network.
[0016] Furthermore, the process of pre-training the deep learning convolutional neural network model based on the deep residual network using the labeled data obtained in S1 is as follows:
[0017] The deep residual network comprises four residual blocks, each with three convolutional layers. Labeled data is input into the residual block, first undergoing dimensionality reduction through 1x1 convolutions, then 3x3 pooling convolutions, and finally 1x1 convolutions to restore its dimensionality. After each convolution operation in the residual block, the output feature map is superimposed with the feature map of the previous layer. After the corresponding convolution operations in the residual block, a global average pooling operation is performed to complete the pre-training of the deep learning convolutional neural network model based on the deep residual network.
[0018] Furthermore, the deep residual network is ResNet50, which contains forty-nine convolutional layers and one global average pooling layer.
[0019] Preferably, the process in S3 of using a semi-supervised learning model to perform consistency regularization on labeled and unlabeled data to obtain a semi-supervised learning dataset is as follows:
[0020] Two data augmentations are performed on unlabeled data, and one data augmentation is performed on labeled data. The consistency regularization loss term is expressed as follows:
[0021] (1)
[0022] in, It is the probability distribution of the model's predicted categories for the augmented input samples. It is unlabeled data. and They are respectively for The new data generated after the first and second data augmentations, These are model parameters. It is the model's prediction result;
[0023] The model classifier outputs low-entropy predictions for unlabeled data and calculates the average value, which is then used as the model's artificial label. The average value is then subjected to entropy minimization using the Sharpening function, and this output is used as the pseudo-label for the unlabeled data. The Sharpening function is expressed as follows:
[0024] (2)
[0025] in, This is the average value calculated by the model classifier to output low-entropy predictions for unlabeled data. The temperature parameter is used to adjust the classification entropy, and L is the loss function value.
[0026] Finally, use hyperparameters. pass Weighting factors are obtained by function sampling As shown in the following formula:
[0027] (3)
[0028] Pick The maximum value is obtained by mixing and overlaying labeled and unlabeled data using a MixUp method to generate a semi-supervised learning dataset. The expression for the semi-supervised learning dataset is:
[0029] (4)
[0030] The labels for labeled and unlabeled data are respectively and .
[0031] Preferably, the semi-supervised learning model is MixMatch.
[0032] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:
[0033] 1. The ResNet50 algorithm model used in this scheme can be pre-trained using features from a large number of remote sensing image samples to obtain preliminary glacier identification results. This method can effectively perform calculations on high-resolution remote sensing images and effectively identify and extract vector information of glacier boundaries from them;
[0034] 2. The original dataset was optimized using the MixMatch algorithm, which enabled the dataset to still obtain good prediction results even with minor perturbations, effectively improving the glacier recognition performance.
[0035] 3. A glacier intelligent recognition scheme based on semi-supervised learning convolutional neural networks was constructed. This scheme can be used for object detection without building a large amount of labeled data in advance. Only a portion of labeled data is needed for model pre-training. Then, by combining semi-supervised learning to optimize the dataset, the accuracy equivalent to R-CNN can be achieved while saving the time and manpower costs of labeling samples, and the speed is also improved to a certain extent. Attached Figure Description
[0036] The present invention will be described by way of example and with reference to the accompanying drawings, wherein:
[0037] Figure 1 This is a schematic diagram of the method flow of the present invention;
[0038] Figure 2 This is a schematic diagram of the sample label of the present invention;
[0039] Figure 3 This is a schematic diagram of the algorithm model structure of the present invention;
[0040] Figure 4 This is a diagram showing the recognition results of the present invention. Detailed Implementation
[0041] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.
[0042] In the description of the embodiments of this application, it should be noted that the terms "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "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 commonly used when the product of the invention is in use. They are only for the convenience of describing this application 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 limitations on this application. In addition, the terms "first," "second," and "third," etc., are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.
[0043] The following is combined with Figures 1-4The present invention will be described in detail below.
[0044] Figure 1 This demonstrates the overall architecture design of the invention, which is divided into three parts: S1, S2, and S3, performed sequentially. Each part has a separate output and serves as the input for the next part.
[0045] Figure 2 The image shown is a sample label diagram of the present invention. The left image is high-resolution remote sensing image data, and the right image is sample label data. The unlabeled sample is a high-resolution image, and there is no need to construct the label shown in the right image.
[0046] Figure 3 The diagram shows the algorithm model structure of the present invention. Labeled data is first input into ResNet50 for model training, then a MixMatch semi-supervised model is constructed to optimize the dataset, and finally the optimized dataset is input into a deep learning convolutional neural network glacier intelligent recognition model to obtain a glacier intelligent recognition model with excellent performance.
[0047] Figure 4 The image shown is a prediction result diagram of the present invention. The left image is the original image, and the right image is the prediction result diagram of the model proposed in this invention. The present invention effectively predicts glacier regions.
[0048] Example
[0049] A glacier identification modeling method based on deep learning networks and semi-supervised learning includes the following steps:
[0050] S1: Select high-resolution remote sensing images of multiple glaciers before their melting period and corresponding digital elevation model data to establish a glacier identification remote sensing image dataset. Segment the glacier identification remote sensing image dataset to obtain labeled and unlabeled data.
[0051] S2: Construct a deep learning convolutional neural network model based on a deep residual network using a deep residual network, and pre-train the deep learning convolutional neural network model based on the deep residual network using the labeled data obtained in S1.
[0052] S3: A semi-supervised learning model is used to perform consistency regularization on labeled and unlabeled data to obtain a semi-supervised learning dataset. The semi-supervised learning dataset is then combined with the deep learning convolutional neural network model based on deep residual network obtained in S2 to construct a glacier recognition model.
[0053] The process of establishing the glacier identification remote sensing image dataset in S1 is as follows:
[0054] The selected high-resolution remote sensing images of multiple glaciers before their melting period were registered and resampled with the corresponding digital elevation model data to ensure that the high-resolution remote sensing images of multiple glaciers before their melting period and the corresponding digital elevation model data have the same number of pixels.
[0055] The process of segmenting the glacier identification remote sensing image dataset in S1 to obtain labeled and unlabeled data is as follows:
[0056] Using ArcGIS software, glacier target areas in high-resolution remote sensing images of multiple glaciers before their melting period were labeled, resulting in labeled and unlabeled data. The labeled and unlabeled data were saved separately, with a ratio of 2:8 between labeled and unlabeled data.
[0057] The process of constructing a deep learning convolutional neural network model based on a deep residual network in S2 is as follows:
[0058] The RGB information of the labeled data obtained in S1 is input into the first convolutional layer of the deep residual network. Convolution, regularization, activation function and max pooling are performed on the input to change the number of channels and size of the original image, so as to obtain a deep learning convolutional neural network model based on the deep residual network.
[0059] The process of pre-training a deep learning convolutional neural network model based on a deep residual network using the labeled data obtained in S1 is as follows:
[0060] The deep residual network comprises four residual blocks, each with three convolutional layers. Labeled data is input into the residual block, first undergoing dimensionality reduction through 1x1 convolutions, then 3x3 pooling convolutions, and finally 1x1 convolutions to restore its dimensionality. The residual network addresses the training error caused by increasing the number of layers in a convolutional neural network. After each convolution operation in the residual block, the output feature map is superimposed with the feature map of the previous layer to improve model accuracy. After the convolution operations in the residual blocks, a global average pooling operation is performed to complete the pre-training of the deep learning convolutional neural network model based on the deep residual network.
[0061] The deep residual network is ResNet50, which contains forty-nine convolutional layers and one global average pooling layer.
[0062] In S3, a semi-supervised learning model is used to perform consistency regularization on labeled and unlabeled data, resulting in the semi-supervised learning dataset.
[0063] Two data augmentations are performed on unlabeled data, and one data augmentation is performed on labeled data. The consistency regularization loss term is expressed as follows:
[0064] (1)
[0065] in, It is the probability distribution of the model's predicted categories for the augmented input samples. It is unlabeled data. and They are respectively for The new data generated after the first and second data augmentations, These are model parameters. It is the model's prediction result;
[0066] The model classifier outputs low-entropy predictions for unlabeled data and calculates the average value, which is then used as the model's artificial label. The average value is then subjected to entropy minimization using the Sharpening function, and this output is used as the pseudo-label for the unlabeled data. The Sharpening function is expressed as follows:
[0067] (2)
[0068] in, This is the average value calculated by the model classifier to output low-entropy predictions for unlabeled data. The temperature parameter is used to adjust the classification entropy, and L is the loss function value.
[0069] Finally, use hyperparameters. pass Weighting factors are obtained by function sampling As shown in the following formula:
[0070] (3)
[0071] Pick The maximum value is obtained by mixing and overlaying labeled and unlabeled data using a MixUp method to generate a semi-supervised learning dataset. The expression for the semi-supervised learning dataset is:
[0072] (4)
[0073] The labels for labeled and unlabeled data are respectively and The semi-supervised learning model is MixMatch. The idea behind the MixMatch algorithm is to optimize the dataset and enhance the training accuracy of the model by performing data augmentation on the samples, generating pseudo-labels, and performing MixUp stacking operations.
[0074] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed and specific, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the technical solution of this application, and these modifications and improvements all fall within the scope of protection of this application.
Claims
1. A glacier identification modeling method based on deep learning networks and semi-supervised learning, characterized in that, Includes the following steps: S1: Select high-resolution remote sensing images of multiple glaciers before their melting period and corresponding digital elevation model data to establish a glacier identification remote sensing image dataset. Segment the glacier identification remote sensing image dataset to obtain labeled and unlabeled data. S2: Construct a deep learning convolutional neural network model based on a deep residual network using a deep residual network, and pre-train the deep learning convolutional neural network model based on the deep residual network using the labeled data obtained in S1. S3: A semi-supervised learning model is used to perform consistency regularization on labeled and unlabeled data to obtain a semi-supervised learning dataset. This dataset is then combined with the deep learning convolutional neural network model based on deep residual networks obtained in S2 to construct a glacier identification model. The process of establishing the glacier identification remote sensing image dataset in S1 is as follows: The selected high-resolution remote sensing images of multiple glaciers before the melting period were registered and resampled with the corresponding digital elevation model data to ensure that the high-resolution remote sensing images of multiple glaciers before the melting period and the corresponding digital elevation model data have the same number of pixels. The process of pre-training the deep learning convolutional neural network model based on deep residual networks using the labeled data obtained in S1 is as follows: The deep residual network comprises four residual blocks, each with three convolutional layers. Labeled data is input into the residual block, first undergoing dimensionality reduction through 1x1 convolution, then 3x3 pooling convolution, and finally 1x1 convolution to restore its dimension. After each convolution operation in the residual block, the output feature map is superimposed with the feature map of the previous layer. After the corresponding convolution operations in the residual block, global average pooling is performed to complete the pre-training of the deep learning convolutional neural network model based on the deep residual network. In step S3, a semi-supervised learning model is used to perform consistency regularization on labeled and unlabeled data to obtain the semi-supervised learning dataset. Two data augmentations are performed on unlabeled data, and one data augmentation is performed on labeled data. The consistency regularization loss term is expressed as follows: (1)； in, It is the probability distribution of the model's predicted categories for the augmented input samples. It is unlabeled data. and They are respectively for The new data generated after the first and second data augmentations, These are model parameters. It is the model's prediction result; The model classifier outputs low-entropy predictions for unlabeled data and calculates the average value, which is then used as the model's artificial label. The average value is then subjected to entropy minimization using the Sharpening function, and this output is used as the pseudo-label for the unlabeled data. The Sharpening function is expressed as follows: (2)； in, This is the average value calculated by the model classifier to output low-entropy predictions for unlabeled data. The temperature parameter is used to adjust the classification entropy, and L is the loss function value. Finally, use hyperparameters. pass Weighting factors are obtained by function sampling As shown in the following formula: (3)； Pick The maximum value is obtained by mixing and overlaying labeled and unlabeled data using a MixUp method to generate a semi-supervised learning dataset. The expression for the semi-supervised learning dataset is: (4)； The labels for labeled and unlabeled data are respectively and .
2. The glacier identification modeling method based on deep learning networks and semi-supervised learning according to claim 1, characterized in that, The process of segmenting the glacier identification remote sensing image dataset in S1 to obtain labeled and unlabeled data is as follows: Using ArcGIS software, glacier target areas in high-resolution remote sensing images of multiple glaciers before their melting period were labeled, resulting in labeled and unlabeled data. The labeled and unlabeled data were saved separately, with a ratio of 2:8 between them.
3. The glacier identification modeling method based on deep learning networks and semi-supervised learning according to claim 1, characterized in that, The process of constructing a deep learning convolutional neural network model based on a deep residual network in S2 is as follows: The RGB information of the labeled data obtained in S1 is input into the first convolutional layer of the deep residual network. Convolution, regularization, activation function and max pooling are performed on the input to change the number of channels and size of the original image, so as to obtain a deep learning convolutional neural network model based on the deep residual network.
4. A glacier identification modeling method based on deep learning networks and semi-supervised learning according to claim 1 or 3, characterized in that, The deep residual network is ResNet50, which contains forty-nine convolutional layers and one global average pooling layer.
5. The glacier identification modeling method based on deep learning networks and semi-supervised learning according to claim 1, characterized in that, The semi-supervised learning model is MixMatch.