An Unsupervised Contrastive Learning Method for Ice Lake Extraction

Abstract

The present invention provides an unsupervised contrast learning method for ice lake extraction, which mainly solves the technical problems in the existing ice lake extraction methods that the production of training sample labels is complex, time-consuming and laborious, and the model is difficult to be directly transferred to other data. Based on a convolutional neural network, this method performs transformation processing on the original image of the ice lake remote sensing image to obtain a transformed image, and forms a sample pair with two branches by combining the original image and the transformed image; then, downsampling processing and mapping processing are respectively performed on the sample pair to obtain a contrast learning module for the ice lake; at the same time, the water body index NDWI spectral feature map is used as the pseudo label for contrast learning, and the position loss is calculated to obtain a position learning module for the ice lake. Finally, an ice lake extraction model is obtained, and the ice lake information can be automatically extracted by inputting any ice lake remote sensing image into this model. The extraction process using this method is more convenient, time-saving and labor-saving, and the extraction efficiency of the ice lake is greatly improved.

Description

Technical Field

[0001] The present invention belongs to the fields of remote sensing image processing and cryosphere, and particularly relates to an unsupervised contrast learning method for ice lake extraction. Background Art

[0002] Ice lakes, with glacial meltwater as the main material source, mostly develop in plateau areas. In recent years, with the global climate warming, the state of ice lakes has also changed accordingly. When the area of an ice lake gradually expands and the moraine dam cannot bear its pressure, the state of the ice lake will change, resulting in the breach of the ice lake. Therefore, accurately monitoring the state of ice lakes is of great significance for preventing the occurrence of ice lake breach disasters and reducing property losses in downstream areas. In recent decades, with the development of remote sensing technology, remote sensing images have recorded rich ground object information, and the types of remote sensing images have become more diverse, and the spatial and temporal resolutions have also been gradually improved, making it gradually possible to continuously monitor ice lakes on a large scale. Therefore, there is an urgent need for a method that can quickly and accurately extract ice lake information.

[0003] With the development of machine learning and deep learning in the field of computer vision, great progress has also been made in object detection methods for glacial lake extraction. For example, the glacial lake extraction method based on random forest (Wangchuk S.; Bolch T. Mapping of glacial lakes using Sentinel-1 and Sentinel-2 data and a random forest classifier: Strengths and challenges. Science of Remote Sensing, 2020, 2), which collects glacial lake and non-glacial lake pixel samples in remote sensing images, uses the reflectance of pixels in each band as spectral features to construct a training set, learns the high-dimensional spectral features of glacial lakes, and then realizes the classification of glacial lake pixels and non-glacial lake pixels. Another example is the application of the deep learning segmentation model U-net to glacial lake extraction, referring to (Qayyum N.; Ghuffar S.; Ahmad H.M.; Yousaf A.; Shahid I. Glacial Lakes Mapping Using Multi Satellite PlanetScope Imagery and Deep Learning. ISPRS International Journal of Geo-Information, 2020, 9) and (Wu R.; Liu G.; Zhang R.; Wang X.; Li Y.; Zhang B.; Cai J.; Xiang W. A Deep Learning Method for Mapping Glacial Lakes from the Combined Use of Synthetic-Aperture Radar and Optical Satellite Images. Remote Sensing, 2020, 12), which also achieved good results. Another example is the combination of generative adversarial networks (Zhao H.; Zhang M.; Chen F. GAN-GL: Generative Adversarial Networks for Glacial Lake Mapping. Remote Sensing, 2021, 13) to obtain the area of glacial lakes in a generative manner.However, existing glacial lake extraction methods, such as NDWI (Li, J.; Sheng, Y. An automated scheme for glacial lake dynamics mapping using Landsat imagery and Digital Elevation Models: A Case Study in the Himalayas. Int. J. Remote Sens. 2012, 33, 5194–5213), C-V model (Zhao H.; Chen F.; Zhang M. A Systematic Extraction Approach for Mapping Glacial lakes in High Mountain Regions of Asia. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2018, 11, 2788-2799), object-oriented method (Mitkari K. V.; Arora M. K.; Tiwari R. K. Extraction of Glacial Lakes in Gangotri Glacier Using Object-Based Image Analysis. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2017, 10, 5275-5283), spectral relationship method, etc., are all difficult to remove the influence of some interference factors, such as shadows, melting glaciers, clouds, etc. At the same time, these methods are all data learning-based methods. The test results of such method models mostly depend on the completeness of the training data. However, obtaining training data with labels often requires a lot of effort. Once facing new remote sensing data, the training samples must be remade. Such methods greatly limit the applicability of the model in different data. Summary of the Invention

[0004] The present invention provides an unsupervised contrast learning glacial lake extraction method, which mainly solves the technical problems in existing glacial lake extraction methods that the production of training sample labels is complex, time-consuming and laborious, and the model is difficult to be directly transferred to other data.

[0005] To solve the above technical problems, the technical solutions provided by the present invention are as follows:

[0006] An unsupervised contrastive learning method for extracting ice lakes, which is characterized in that it includes the following steps:

[0007] Step 1, perform transformation processing on the original image of the ice lake remote sensing image to obtain a transformed image, and form a sample pair containing two branches with the original image and the transformed image;

[0008] Step 2, input the sample pair into a network with shared weights for downsampling processing, and extract the corresponding feature maps of the two branches at different scales respectively;

[0009] Step 3, input the corresponding feature maps of the two branches obtained in Step 2 at different scales into the projection layer for mapping respectively, obtain the mapped feature vectors of the two at the corresponding scales, and use the similarity loss to measure the similarity degree of the two. Among them, the similarity loss is calculated using cosine similarity, and then the contrastive learning module of the ice lake is obtained;

[0010] Step 4, calculate the water body index NDWI spectral feature map of the original image, set the water body index threshold to T, and then obtain the rough ice lake distribution area map of the original image. Use this rough result map as the pseudo-label for learning by the contrastive learning module obtained in Step 3 to guide the recognition of ice lake features;

[0011] Step 5, perform upsampling processing on the ice lake feature maps corresponding to the original image of the two branches obtained in Step 2 in the network with shared weights, and output the ice lake extraction result corresponding to the original image;

[0012] Step 6, calculate the position loss by combining the pseudo-label obtained in Step 4 and the ice lake extraction result obtained in Step 5, and then obtain the position learning module of the ice lake. Combine the position learning module and the contrastive learning module obtained in Step 3 to obtain the ice lake extraction model;

[0013] Step 7, input any ice lake remote sensing image into the ice lake extraction model obtained in Step 6, and the ice lake extraction result can be output.

[0014] Furthermore, in Step 1, the transformation processing is performed by any one of color mapping transformation, flipping transformation, grayscale transformation, and blur transformation, and noise is randomly added, and at the same time, the image size is processed to 448×448.

[0015] Further, in step 2, the scales include 4 scales with sizes of 224×224, 112×112, 56×56, and 28×28 respectively. One scale is a downsampling block, and each downsampling block includes two convolutional layers and one downsampling layer. Each convolutional layer is activated using ReLU, and each downsampling layer uses a convolutional operation with a stride of 2; the ice lake feature maps of the original image and the transformed image are respectively the results activated by the last convolutional layer before each downsampling layer at the corresponding scale.

[0016] Further, in step 3, the projection layer includes three fully connected layers. The mapping means that ReLU is respectively used to activate the feature maps entering the two fully connected layers, so as to obtain the mapping result in the third fully connected layer. The mapping result is the feature map vector corresponding to different scales, and this vector is used to calculate the similarity loss of the mapping result. The similarity loss is calculated using cosine similarity, and the calculation formula is as follows:

[0017]

[0018] where q represents the ice lake feature map of the original image, q′ represents the ice lake feature map of the transformed image, i represents different scales, i is a positive integer, and 1≤i≤4; p(q i ) represents the feature vector after mapping of the ice lake feature map of the original image; p(q′ i ) represents the feature vector after mapping of the ice lake feature map of the transformed image; ||p(q i )||2 represents the L2 norm of the feature vector after mapping of the ice lake feature map of the original image; ||p(q′ i )||2 represents the L2 norm of the feature vector after mapping of the ice lake feature map of the transformed image.

[0019] Further, in step 4, the calculation formula for the water body index NDWI spectral feature map is:

[0020]

[0021] where ρ Green represents the top-of-atmosphere apparent reflectance in the green band, 0<ρ Green <1, ρ NIR represents the top-of-atmosphere apparent reflectance in the near-infrared band, 0<ρ NIR <1; the threshold T of the water body index takes a value of 0.7.

[0022] Further, step 5 is specifically as follows: The glacier lake feature maps corresponding to the original map at different scales in the two branches obtained in step 2 are respectively used as upsampling blocks. Each upsampling block contains an upsampling layer and two convolutional layers. The upsampling layer uses a transposed convolution operation, and each convolutional layer is activated using the ReLU function. The finally restored result is the glacier lake extraction result of the original map.

[0023] Further, in step 6, the position loss is calculated using the L2 norm. The calculation formula of the L2 norm is:

[0024]

[0025] where I represents the input image; B represents the glacier lake mask obtained after thresholding based on the water body index NDWI; f(·) represents the glacier lake extraction model; u represents any position in the image; f u (I) represents the extraction result at any position in the input image; B u represents the mask result at any position in the image.

[0026] Compared with the prior art, the beneficial effects of the present invention are:

[0027] 1. The unsupervised contrastive learning glacier lake extraction method of the present invention is based on a convolutional neural network and adopts an unsupervised training method. By performing transformation processing on the original map of the glacier lake remote sensing image, a transformed map is obtained, and the original map and the transformed map are combined into a sample pair containing two branches. Then, downsampling processing and mapping processing are respectively performed on the sample pair to obtain a contrastive learning module for glacier lakes. At the same time, the water body index NDWI spectral feature map is used as a pseudo-label for contrastive learning, and a position learning module for glacier lakes is obtained by calculating the position loss. Finally, a glacier lake extraction model is obtained, and the glacier lake information can be automatically extracted by inputting any glacier lake remote sensing image into this model. This method does not require making labels for training data, greatly simplifies the preparatory work of the model training process, makes the glacier lake extraction process more convenient, time-saving and labor-saving, and significantly improves the extraction efficiency of glacier lakes.

[0028] 2. The unsupervised contrastive learning glacier lake extraction method of the present invention sets the water body index threshold to 0.7 to ensure that all noise interference is excluded. NDWI pixels greater than this threshold are marked as water bodies, thereby improving the accuracy of the model in extracting glacier lake information.

[0029] 3. The unsupervised contrastive learning glacier lake extraction method of the present invention uses a pseudo-label as the label for contrastive learning, and the obtained glacier lake model has strong applicability. By inputting any image into the trained model, a relatively accurate glacier lake extraction result can be obtained, thereby realizing the automatic extraction of glacier lakes. Description of the Drawings

[0030] Figure 1Schematic diagram of the extraction process of an unsupervised contrast learning ice lake extraction method according to the present invention. Detailed implementation manners

[0031] Contrast learning is a deep learning model that has emerged in the past two years. It transforms the input image and simultaneously inputs the transformed images into two networks with shared weights to discover targets with similar features in the image. The method based on data feature learning can learn ice lake features as much as possible. Based on this principle, the present invention proposes an unsupervised contrast learning ice lake extraction method that does not require a large number of training labels, avoids using auxiliary data or a large amount of preprocessing and postprocessing work, and obtains a good ice lake extraction effect, realizing the automatic extraction of ice lakes.

[0032] The following elaborates in detail on an unsupervised contrast learning ice lake extraction method of the present invention with reference to the accompanying drawings.

[0033] As Figure 1 shown, an unsupervised contrast learning ice lake extraction method provided by the present invention is based on a convolutional neural network to realize the automatic extraction of ice lakes in remote sensing images under unlabeled data, and specifically includes the following steps:

[0034] Step 1, perform transformation processing on the original image of the ice lake remote sensing image to obtain a transformed image, and form a sample pair containing two branches with the original image and the transformed image.

[0035] Perform transformation processing on the original image of the obtained ice lake remote sensing image to obtain a transformed image, and form a sample pair with the original image and the transformed image. In order to obtain mutually contrasting sample pairs, this embodiment provides 4 ways of image transformation, specifically including:

[0036] ① Color mapping transformation: The specific parameters include the transformation of brightness, contrast, saturation, and hue. Among them, the maximum adjustment range of brightness is 0.4; the maximum adjustment range of contrast is 0.4; the maximum adjustment range of saturation is 0.4; the maximum adjustment range of hue is 0.2;

[0037] ② Flip transformation: That is, perform horizontal transformation and vertical transformation;

[0038] ③ Grayscale transformation: Perform grayscale transformation with a probability of 0.2;

[0039] ④ Blur transformation: This embodiment can adopt a blur transformation with a Gaussian kernel, where the blur parameter sigma is set to 1.

[0040] Generally, the algorithm for image transformation is written in the Python language to implement the transformation processing of images. When performing image transformation, one of the above four methods is randomly selected, and noise is randomly added. At the same time, the image needs to be processed into a size of 448*448. Then, the transformed image obtained after processing and the original image are jointly formed into a sample pair as the input of the network.

[0041] Step 2: Input the sample pair into the network with shared weights for downsampling processing, and extract the corresponding feature maps of the two branches at different scales respectively.

[0042] Input the sample pair composed of the original image and the transformed image into the network. In this embodiment, a network with shared weights is selected. The parameters for downsampling the two branches in this network are kept consistent, and downsampling processing is performed on them respectively. The different scales include 4 scales, which are 224×224, 112×112, 56×56, and 28×28 respectively. One scale represents a downsampling block. Each downsampling block includes two convolutional layers and one downsampling layer. Each convolutional layer is activated using ReLU, and each downsampling layer uses a convolutional operation with a stride of 2. The activation result of the last convolutional layer before each downsampling layer is the ice lake feature map corresponding to the original image and the transformed image at different scales. The ice lake feature maps at different scales contain different ice lake details.

[0043] Step 3: Input the corresponding feature maps of the two branches obtained in Step 2 at different scales into the projection layer for mapping respectively to obtain the mapped feature vectors corresponding to them at the corresponding scales. Based on these feature vectors, the similarity loss is used to measure the similarity between the two. Among them, the similarity loss is calculated using cosine similarity, and thus the contrast learning module of the ice lake is obtained.

[0044] Specifically, input the corresponding feature maps of the two branches obtained in Step 2 at different scales into the projection layer for mapping respectively, that is, input the ice lake feature maps corresponding to the original image at different scales and the ice lake feature maps corresponding to the transformed image at different scales into the projection layer respectively. The projection layer includes three fully connected layers. Mapping means that the outputs of the feature maps entering the two fully connected layers are respectively activated using ReLU, so as to obtain the mapping results of the two branches at the corresponding scales in the third fully connected layer. The mapping results are the feature vectors corresponding to different scales. This feature vector is used to calculate the similarity loss of the mapping results of the two branches at the corresponding scales. The similarity loss refers to the similarity degree of the features of two similar images after passing through the contrast branch with shared weights. Thus, the contrast learning module of the ice lake is obtained. This module can learn the similar features in the feature maps corresponding to the two branches at different scales.

[0045] In this embodiment, the similarity loss is measured using cosine similarity, and the calculation formula is as follows:

[0046]

[0047] Among them, q represents the ice lake feature map of the original image, q' represents the ice lake feature map of the transformed image, i represents different scales, i is a positive integer, and 1 ≤ i ≤ 4; p(q i ) represents the feature vector after mapping the ice lake feature map of the original image; p(q' i ) represents the feature vector after mapping the ice lake feature map of the transformed image; ||p(q i )||2 represents the L2 norm of the feature vector after mapping the ice lake feature map of the original image; ||p(q' i )||2 represents the L2 norm of the feature vector after mapping the ice lake feature map of the transformed image. Since there are 4 features of different scales, the similarities at the 4 scales are accumulated, and then the contrast learning module of the ice lake is obtained.

[0048] Step 4: Calculate the water body index NDWI spectral feature map of the original image, and set the water body index threshold to T, where T takes the value of 0.7, and then obtain the rough ice lake distribution area map of the original image. Use this rough ice lake distribution area map as the pseudo-label for the contrast learning in Step 3 to guide the recognition of ice lake features.

[0049] Since contrast learning can only learn the similar features in the image, and cannot learn whether the pixels in the image are ice lakes. In this embodiment, the water body index NDWI (normalized difference water index, NDWI) spectral feature map is combined, and at the same time, a relatively high water body index threshold T is used to obtain the rough ice lake distribution area map. Use this result as the pseudo-label for the contrast learning module in Step 3, so as to guide the ice lake training model to further recognize the ice lake features and improve the accuracy of the ice lake training model. The calculation formula of the water body index NDWI spectral feature map is:

[0050]

[0051] Among them, ρ Green represents the top-of-atmosphere apparent reflectance in the green band, 0 < ρ Green < 1, ρ NIR represents the top-of-atmosphere apparent reflectance in the near-infrared band, 0 < ρ NIR < 1.

[0052] In this embodiment, in order to ensure that all noise interference can be excluded, a relatively high water body index threshold T of 0.7 is selected. If the NDWI pixel greater than this threshold is marked as water body.

[0053] Step 5: Upsample the ice lake feature maps corresponding to the original image at different scales in the two branches obtained in Step 2 in a network with shared weights, and output the ice lake extraction result corresponding to the original image.

[0054] Take the original images of the four scales obtained in step 2 corresponding to the ice lake feature maps as upsampling blocks respectively. Each upsampling block contains an upsampling layer and two convolutional layers. Among them, the upsampling layer uses deconvolution operation, and each convolutional layer is activated by the ReLU function. The finally restored result is the ice lake extraction result of the original image.

[0055] Step 6: Combine the pseudo-labels obtained in step 4 and the ice lake extraction result obtained in step 5 to calculate the position loss between the two. The position loss is calculated using the L2 norm, and then a position learning module for the ice lake is obtained. Combine this position learning module with the ice lake contrast learning module obtained in step 3 to obtain the ice lake extraction model.

[0056] In this embodiment, the position loss is measured using the L2 norm, and the formula is:

[0057]

[0058] where I represents the input image; B represents the ice lake mask obtained after thresholding according to the water body index NDWI; f(·) represents the ice lake extraction model; u represents any position in the image; f u (I) represents the extraction result at any position in the input image; B u represents the mask result at any position in the image.

[0059] Step 7: Input any ice lake remote sensing image into the ice lake extraction model in step 6, and the ice lake extraction result can be output.

[0060] Generally speaking, the present invention is an unsupervised contrast learning ice lake extraction method, which can effectively and accurately extract the ice lake boundary without any labeled training data.

[0061] The effect of the present invention can be further illustrated by the following experiments.

[0062] 1. Experimental conditions

[0063] The present invention is implemented by programming in the Python language on a central processing unit of i5-9400F 2.9GHz CPU, GTX 1660T 6G GPU, 16G of memory, and the WINDOWS 10 operating system. The model also involves a deep learning framework, and the deep learning framework used in this experiment is tensorflow 1.14. The data used in the experiment are all collected from Landsat-8 images, and the size of all images is 256×256×7.

[0064] 2. Experimental content

[0065] The calculation method for experimental accuracy evaluation is to use the F1 score. The specific calculation method of the F1 score is as follows:

[0066] Precision = Number of correctly extracted glacier lake pixels / All extracted pixels

[0067] Recall = Number of correctly extracted glacier lake pixels / All glacier lake pixels

[0068] F1 score = 2 × Precision × Recall / (Precision + Recall)

[0069] The size of the input image is set to 224×224×7. Regarding the training parameters, the batch size is set to 8, the epoch is set to 100, the dropout is set to 0.5 to prevent overfitting, the optimizer is selected as AdamOptimizer, and the learning rate is set to 0.0005. To verify the effectiveness of the present invention, this embodiment is compared with other extraction methods, including supervised and unsupervised methods. Among them, the supervised glacial lake extraction methods include the random forest segmentation algorithm, refer to (Wangchuk S.; Bolch T. Mapping of glacial lakes using Sentinel-1 and Sentinel-2 data and a random forest classifier: Strengths and challenges. Science of Remote Sensing, 2020, 2.), the U-net model, refer to (Qayyum N.; Ghuffar S.; Ahmad H.M.; Yousaf A.; Shahid I. Glacial Lakes Mapping Using Multi Satellite PlanetScope Imagery and Deep Learning. ISPRS International Journal of Geo-Information, 2020, 9.) and (Wu R.; Liu G.; Zhang R.; Wang X.; Li Y.; Zhang B.; Cai J.; Xiang W. A Deep Learning Method for Mapping Glacial Lakes from the Combined Use of Synthetic-Aperture Radar and Optical Satellite Images. Remote Sensing, 2020, 12.) and the generative adversarial network GAN-GL, refer to (Zhao H.; Zhang M.; Chen F. GAN-GL: Generative Adversarial Networks for Glacial Lake Mapping. Remote Sensing, 2021, 13.).The unsupervised ice lake extraction algorithm includes an improved C-V model, referring to (Zhao H., Chen F., Zhang M. A Systematic Extraction Approach for Mapping Glacial lakes in High Mountain Regions of Asia. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2018, 11, 2788-2799.), and a global-local iterative segmentation algorithm, referring to (Li, J., Sheng, Y. An automated scheme for glacial lake dynamics mapping using Landsat imagery and Digital Elevation Models: A Case Study in the Himalayas. Int. J. Remote Sens. 2012, 33, 5194–5213.). In the random forest method, approximately 1.98 million sample points were randomly sampled, with the number of ice lake pixels and background pixels remaining the same. 70% of these sample points were used for training, and 30% were used for testing. The improved C-V model algorithm is based on the principle of region segmentation. The model has good noise resistance, and at the same time, the quadratic term in the C-V model is simplified to a linear term to accelerate the model iteration process. The global-local iterative segmentation algorithm first uses the water body index NDWI for the rough extraction of ice lakes, then buffers the rough extraction results, and further extracts ice lake information using the bimodal threshold segmentation algorithm in the buffer. The dataset used in this embodiment is Landsat-8 remote sensing imagery, and after cropping, an ice lake image with a size of 256×256×7 is obtained. Similarly, 70% of the images in the dataset are used for the training process in the supervised model, and 30% are used for testing. The unsupervised model directly uses 30% of the images for testing. The accuracy comparison results of different models are shown in Table 1 below:

[0070] Table 1 Comparison of accuracy of ice lake extraction by typical models

[0071]

[0072] As can be seen from Table 1, compared with the traditional ice lake extraction model and the deep learning model, the model proposed by the present invention can learn similar targets in images by means of comparison due to the introduction of unsupervised contrastive learning. At the same time, by introducing pseudo-labels, the ice lake information in the model can be directly output, thus avoiding the heavy process of making a large number of training labels. At the same time, the model can also achieve an effect close to that of the supervised method and far exceed the results of the unsupervised method.

[0073] In the above, although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art in the technical field, as long as it is within the scope of the substantial spirit of the present invention, the changes and modifications to the above embodiments should be regarded as within the scope of the claims of the present invention.

Claims

1. An unsupervised contrastive learning method for extracting ice lakes, characterized in that, It includes the following steps: Step 1: Perform transformation processing on the original image of the ice lake remote sensing image to obtain a transformed image, and form a sample pair with two branches by combining the original image and the transformed image; Step 2: Input the sample pair into a network with shared weights for downsampling processing, and extract the corresponding feature maps of the two branches at different scales respectively; Step 3: Input the corresponding feature maps of the two branches obtained in Step 2 at different scales into the projection layer for mapping respectively to obtain the feature vectors mapped at the corresponding scales for the two, and use a similarity loss to measure the similarity degree between the two. Among them, the similarity loss is calculated using cosine similarity, and then a contrast learning module for the ice lake is obtained; Step 4: Calculate the water body index NDWI spectral feature map of the original image, set the water body index threshold as T, and then obtain a rough ice lake distribution area map of the original image. Use this rough result map as the pseudo-label for learning by the contrast learning module obtained in Step 3 to guide the identification of ice lake features; Step 5: Upsample the ice lake feature maps corresponding to the original image at different scales in the two branches obtained in Step 2 in a network with shared weights, and output the ice lake extraction result corresponding to the original image; Step 6: Calculate the position loss by combining the pseudo-label obtained in Step 4 and the ice lake extraction result obtained in Step 5, and then obtain a position learning module for the ice lake. Combine this position learning module and the contrast learning module obtained in Step 3 to obtain an ice lake extraction model; Step 7: Input any ice lake remote sensing image into the ice lake extraction model obtained in Step 6, and the ice lake extraction result can be output.

2. An unsupervised contrast learning ice lake extraction method according to claim 1, characterized in that: In Step 1, the transformation processing is performed by any one of color mapping transformation, flipping transformation, grayscale transformation, and blur transformation, and noise is randomly added, and at the same time, the image size is processed to 448×448.

3. An unsupervised contrast learning ice lake extraction method according to claim 2, characterized in that: In Step 2, the scales include 4 scales, with sizes of 224×224, 112×112, 56×56, and 28×28 respectively. One scale is one downsampling block, and each downsampling block includes two convolutional layers and one downsampling layer respectively. Each convolutional layer is activated using ReLU, and each downsampling layer uses a convolutional operation with a stride of 2; the ice lake feature map of the original image and the ice lake feature map of the transformed image are respectively the results activated by the last convolutional layer before each downsampling layer at the corresponding scale.

4. An unsupervised contrast learning ice lake extraction method according to claim 3, characterized in that: In Step 3, the projection layer includes three fully connected layers. The mapping means that the feature maps entering the two fully connected layers are respectively activated using ReLU, so as to obtain a mapping result in the third fully connected layer. The mapping result is the feature map vector corresponding to different scales, and this vector is used to calculate the similarity loss of the mapping result. The similarity loss is calculated using cosine similarity, and the calculation formula is as follows: Among them, q represents the ice lake feature map of the original image, q' represents the ice lake feature map of the transformed image, i represents different scales, i is a positive integer, and 1 ≤ i ≤ 4; p(q i ) represents the feature vector after mapping of the ice lake feature map of the original image; p(q' i ) represents the feature vector after mapping of the ice lake feature map of the transformed image; ||p(q i )||2 represents the L2 norm of the feature vector after mapping of the ice lake feature map of the original image; ||p(q' i )||2 represents the L2 norm of the feature vector after mapping of the ice lake feature map of the transformed image.

5. An unsupervised contrastive learning method for ice lake extraction according to claim 4, characterized in that: In step 4, the calculation formula of the water body index NDWI spectral feature map is: Among them, ρ Green represents the apparent reflectance at the top of the atmosphere in the green light band, 0 < ρ Green < 1, ρ NIR represents the apparent reflectance at the top of the atmosphere in the near-infrared band, 0 < ρ NIR < 1; the threshold T of the water body index takes a value of 0.

7.

6. An unsupervised contrastive learning method for ice lake extraction according to claim 5, characterized in that: Step 5 is specifically: taking the ice lake feature maps corresponding to the original images in different scales obtained in step 2 in the two branches as upsampling blocks respectively, each upsampling block includes an upsampling layer and two convolutional layers, the upsampling layer uses a transposed convolution operation, and each convolutional layer is activated using the ReLU function. The finally restored result is the ice lake extraction result of the original image.

7. An unsupervised contrastive learning method for ice lake extraction according to claim 6, characterized in that: In step 6, the position loss is calculated using the L2 norm, and the calculation formula of the L2 norm is: Among them, I represents the input image; B represents the ice lake mask obtained after thresholding by the water index NDWI; f(·) represents the ice lake extraction model; u represents any position in the image; f u (I) represents the extraction result at any position in the input image; B u represents the masking result at any position in the image.

Images