Optical-SAR (Synthetic Aperture Radar) collaboration-based time sequence ice lake area intelligent extraction method

Abstract

The invention discloses a time sequence ice lake area intelligent extraction method based on optics-SAR cooperation, and the method comprises the steps: 1, collecting a multi-time sequence optical remote sensing image, and constructing a target detection data set of an ice lake sample Pascal VOC format; 2, pre-training a YOLOv9-E weight by adopting an MSCOCO data set and carrying out transfer learning, and realizing global positioning of the glacial lake by combining a leak detection error optimization method to obtain a boundary frame coordinate of the glacial lake; 3, performing localization processing on the SAR image based on the boundary frame coordinates of the glacial lake, constructing a self-adaptive glacial lake extraction model fused with a multivariate unsupervised method, and outputting an accurate contour of the glacial lake; 4, constructing a multi-source error detection system, and performing secondary abnormal value identification and correction by adopting an isolated forest algorithm to obtain time sequence ice lake area data after quality control; and 5, based on the time sequence ice lake area data and SAR local feature analysis, identifying newly added and disappeared ice lakes, and calculating an ice lake area change rate to obtain an ice lake evolution analysis result. According to the invention, high-precision intelligent extraction of ice lake time sequence evolution under the multi-cloud condition is realized.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing technology, specifically a time-series intelligent extraction method for ice lake area based on optical-SAR synergy. Background Technology

[0002] Glacial lakes, as unique water bodies in high-altitude regions, have significant ecological and hydrological importance. Monitoring the dynamic evolution of glacial lakes is crucial for preventing glacial lake disasters and alleviating water shortages. In recent years, remote sensing technology has been widely used in the field of glacial lake monitoring. However, traditional methods for extracting glacial lake features, including water index methods based on optical images, threshold segmentation methods, and backscatter feature analysis methods based on SAR images, mainly rely on a single remote sensing data source. Their application in glacial lake monitoring has obvious limitations. This is because, although optical remote sensing images have advantages such as high resolution and intuitive recognition, and can accurately capture the details and features of the glacial lake surface, their application is severely restricted by cloud and fog obstruction. Typically, mountainous areas where glacial lakes are distributed generally have a large amount of cloud and fog, and there are significant seasonal variations in snowfall and freezing. Commonly used time-series cloud and fog removal algorithms are extremely ineffective in this region. As a result, optical remote sensing studies in glacial lake distribution areas (especially in southeastern Tibet) are mostly concentrated from November to February of the following year. During this period, the glacial lake surface is frozen and covered with snow, and the features of the glacial lake are almost invisible in optical remote sensing images. Manual visual interpretation can only rely on the texture and topographic differences between the lake shore slope and the lake surface, which significantly reduces the reliability of the results.

[0003] Synthetic Aperture Radar (SAR) technology, as an active remote sensing method, utilizes microwave wavelengths that can effectively penetrate clouds and fog, unaffected by the angle of solar incidence, providing an effective alternative for monitoring glacial lakes in cloudy areas. Glacial lakes in SAR images typically exhibit low backscattering characteristics due to specular reflection of incident electromagnetic waves on the continuous and static surface of the glacial lake, resulting in extremely weak radar echo signal strength. However, SAR images also face many challenges in glacial lake feature extraction: 1) There are various confounding factors, including ground features with strong radar wave absorption characteristics (such as wet snow), specular reflection caused by smooth ground features, deviation of the reflection direction of slopes facing away from the radar, and geometric distortion. These factors can all lead to low echo signal characteristics similar to those of glacial lakes. In particular, the southeastern Tibetan region has a high altitude and experiences multiple snowfalls throughout the year. Wet snow is a key factor causing radar wave signal absorption, making it difficult to distinguish the signals of glacial lakes from those of wet snow in snow-covered areas; 2) Geometric distortion caused by topographic relief significantly interferes with glacial lake feature extraction. Although image quality can be improved through ascent-descent fusion, complete complementarity cannot be achieved.

[0004] The development of deep learning technology has brought new opportunities for glacial lake feature extraction. Semantic segmentation methods based on convolutional neural networks have made significant progress in the accuracy of glacial lake mapping, effectively making up for the shortcomings of traditional band ratio and machine learning techniques in feature dimensionality reduction. They can extract more complex and multi-scale high-dimensional features. However, existing deep learning methods still have the following shortcomings: 1) False positives and false negatives are still common, which is not only related to the complexity of model structure and parameters, but also affected by the quality of data sources; 2) In different remote sensing images, the differences in physicochemical characteristics between glacial lakes and surrounding water bodies may lead to an imbalance in imaging features. Many land features have similar characteristics to glacial lakes, increasing the risk of false positives; 3) Most existing models are statistical models, and the significant seasonal changes exhibited by glacial lakes pose a major challenge to these models in the recognition of complex image features.

[0005] Currently, target detection technology is increasingly widely used in remote sensing image processing. The YOLO series of algorithms has attracted much attention due to its balanced performance in speed and accuracy. However, when applying target detection technology to ice lake identification, the following challenges still exist: 1) Ice lakes are small to medium-sized targets, and the probability of missed detection in target detection is extremely high; 2) Existing pre-trained models are mostly based on natural image datasets, which have significant domain differences from remote sensing images; 3) There is a lack of loss function optimization strategies tailored to the characteristics of remote sensing images.

[0006] Multi-source remote sensing image fusion is considered an effective way to overcome the limitations of a single data source. Some researchers have adopted multi-source image fusion methods to try to achieve accurate extraction of glacial lake features through the complementarity of different remote sensing information. However, existing multi-source fusion techniques have obvious defects: 1) They lack the necessary prior weights to guide the model to avoid potential error terms, which usually generate erroneous samples, causing the model to learn incorrect features; 2) Simple feature-level fusion often ignores the differences in imaging mechanisms of different data sources, which may amplify errors in certain error types; 3) Existing fusion methods mostly adopt global processing strategies and fail to make full use of the relative advantages of different data sources in different application scenarios.

[0007] Currently, existing research on the temporal evolution analysis of glacial lakes suffers from the following limitations: 1) Insufficient temporal resolution: most studies only provide information on changes over annual or longer timescales, making it difficult to capture the seasonal evolution characteristics of glacial lakes; 2) Limited spatial coverage: due to limitations in data quality and processing capabilities, most studies focus on individual glacial lakes or small areas; 3) Lack of effective outlier detection and data quality control methods, affecting the reliability of long-term time-series analysis results; 4) Insufficient automatic identification capability for newly added and disappearing glacial lakes, relying mainly on manual interpretation, which is inefficient and highly subjective; 5) Existing glacial lake feature extraction methods face challenges in practical applications, including low computational efficiency and ease of deployment. While traditional manual visual interpretation methods are highly accurate, they are labor-intensive, and cognitive biases among different operators lead to significant differences between datasets. Although existing automated methods have improved processing efficiency, they often require substantial computational resources and complex parameter tuning processes, limiting their widespread adoption in operational applications.

[0008] In summary, current glacial lake feature extraction technologies suffer from significant shortcomings in terms of applicability in cloudy and foggy environments, effectiveness of multi-source data fusion, robustness of deep learning models, and automation levels in temporal evolution analysis. Particularly in high-altitude, cloudy areas like southeastern Tibet, existing technologies struggle to achieve high-precision and high-frequency temporal monitoring of wide-area glacial lakes, severely hindering research on glacial lake evolution patterns and the improvement of disaster early warning capabilities. Therefore, there is an urgent need to develop an intelligent method that integrates the advantages of optical and SAR imagery to achieve high-precision extraction of the temporal evolution of wide-area glacial lakes under cloudy conditions, providing reliable technical support for glacial lake disaster early warning, water resource management, and climate change response research. Summary of the Invention

[0009] The purpose of this invention is to provide a time-series intelligent extraction method for ice lake area based on optical-SAR collaboration, which combines optical global positioning and SAR local contour extraction to achieve high-precision intelligent extraction of ice lake time-series evolution under cloudy conditions.

[0010] This invention is achieved through the following technical solution: A time-series intelligent extraction method for ice lake area based on optical-SAR synergy includes the following steps: Step 1: Collect multi-temporal optical remote sensing images and construct a deep learning target detection dataset containing glacial lake samples in Pascal VOC format based on existing glacial lake catalog data, thereby generating a glacial lake sample dataset suitable for model training and validation. Step 2: Pre-train the weights of the YOLOv9-E framework using the MSCOCO dataset and perform transfer learning. Combine this with a deep learning object detection dataset in Pascal VOC format based on glacial lake samples. Freeze the low-level features of layers 0-28 in the pre-trained model and fine-tune only the high-level features and classification layers to adapt the model to the remote sensing image features. Then, construct a missed detection function. Sum of error functions This optimizes the loss, and then records the confidence of the ice lake detection box by spatially associating multi-period detection coordinate boxes. In the overlapping area, non-maximum suppression is used to obtain high-confidence localization information, and the coordinates of the ice lake bounding box are output. Step 3: Based on the coordinates of the ice lake boundary box, perform localization processing on the SAR image and construct an adaptive ice lake extraction model that integrates multiple unsupervised methods to output the accurate outline of the ice lake. Step 4: Construct a multi-source error detection system that integrates sliding window detection, seasonal change detection, and standard deviation outlier detection to screen out outliers in the precise contour of the glacial lake. Then, use the isolated forest algorithm for secondary outlier identification and correction to obtain the quality-controlled time-series glacial lake area data. Step 5: Based on time-series glacial lake area data, identify newly added and disappeared glacial lakes based on SAR local feature analysis and calculate the rate of change of glacial lake area. Comprehensively analyze the evolution of glacial lakes to obtain the glacial lake evolution analysis results.

[0011] Furthermore, the specific process of step 1 is as follows: Step 1.1: Obtain existing ice lake catalog data, merge ice lake catalog data from different data sources, and obtain the initial dataset of ice lake samples; Step 1.2: Based on the distribution characteristics of the Sentinel-1 GRD data, construct a quality assessment system for glacial lake samples. Specifically, use a Gaussian mixture model to determine whether the glacial lake samples in the initial dataset exhibit a bimodal distribution. If so, proceed to step 1.3; otherwise, classify them as erroneous samples and delete them. The bimodal distribution is expressed by the formula: In the formula: Represents glacial lake sample data The probability density function; Indicates the number of Gaussian components; Indicates the first The mixing weights of Gaussian components; Indicates the first The mean vector of Gaussian components; Indicates the first The covariance matrix of Gaussian components; Step 1.3: Calculate the difference between the bimodal distribution of the glacial lake sample and the correct glacial lake characteristic distribution using Kullback-Leibler divergence. , is represented as: In the formula: Represents distribution Relative to distribution of Divergence; The distribution of characteristics of the corresponding glacial lakes is correct; Corresponding to the bimodal distribution of the glacial lake characteristics to be detected, when When the value is greater than 0.1, Mark as an error sample and delete; otherwise, proceed to step 1.4. Step 1.4: Construct a cloud and fog detection method based on the RGB dark channel prior, expressed as: In the formula: Representative position Color channel Pixel values; Indicated by The local neighborhood centered on; when If the value is greater than 0.5, the sample is determined to be a cloud-polluted image and deleted; otherwise, proceed to step 1.5. Step 1.5: After excluding cloud-polluted images and erroneous samples in steps 1.2 to 1.4, and combining manually vectorized supplementary samples, a deep learning object detection dataset containing ice lake samples in Pascal VOC format is obtained, and it is divided into ice lake sample datasets suitable for model training and validation.

[0012] Furthermore, the specific process of step 2 is as follows: Step 2.1: Use the MSCOCO dataset to pre-train the weights of the YOLOv9-E framework and perform transfer learning. Combine the deep learning object detection dataset in Pascal VOC format for glacial lake samples. Freeze the low-level features of layers 0-28 in the pre-trained model and fine-tune only the high-level features and classification layer to adapt the model to the features of remote sensing images. The formulas for calculating the cross-union ratio, precision, recall, and average precision of YOLOv9-E are expressed as follows: In the formula: This represents the correct detection target; This represents a misjudgment of the target; Represents the target that was missed; The first representing the sequence item; The first representing the sequence - 1 item; Representing the The weight of the item; Represents the number of terms in the sequence; express Average accuracy at the threshold; Represents average accuracy; represent Threshold range; Step 2.2: Construct the missed detection function Sum of error functions , used to optimize the loss, is expressed as: In the formula: This represents the false negative rate. This is due to the error of missed detection; The detection frame as ;Will The detection frame as ; The weight for missed detection loss is usually set to 1.0; Step 2.3: Synthesize the third quarter image data in each image period. The YOLOv9-E framework outputs detection bounding boxes for each synthesized quarterly image data period. By spatially associating the detection bounding boxes from multiple periods, the confidence of the ice lake detection boxes is recorded. In the overlapping areas, non-maximum suppression is used to obtain high-confidence localization information, and the coordinates of the ice lake bounding box are output.

[0013] Furthermore, the specific process of step 3 is as follows: Step 3.1: Construct a buffer using the minimum envelope rectangle to perform SAR image localization. The formula for calculating the buffer radius is as follows: In the formula: The adjustment factor has a default value of 2. This represents the ratio of the shorter side to the longer side of the coordinate frame. This represents the length of the main diagonal of the coordinate frame; The radius of the buffer zone; Step 3.2: Perform histogram adjustment on the localized SAR image, truncating 1% of the pixel values ​​before and after the cutoff to optimize the data distribution characteristics; Step 3.3: Integrate multiple unsupervised algorithms, including adaptive thresholding and clustering methods, to obtain a multivariate unsupervised adaptive ice lake extraction model. Based on maximizing the intersection-over-union ratio (IoU), determine the weights of the multivariate unsupervised adaptive ice lake extraction model. The loss function is expressed as: In the formula: The number of algorithms remaining after filtering: For the weights of each algorithm and ; For the first The segmentation result of each ice lake after weighted combination of various algorithms; For the first A truth mask for an ice lake; Represents any unsupervised algorithm; Indicates weight; Step 3.4: Based on the start dates of the four seasons in southeastern Tibet (February 25 for spring, May 31 for summer, September 15 for autumn, and November 28 for winter), the S1-GRD images for each season are mean-fused using a multivariate unsupervised adaptive glacial lake extraction model. Then, morphological post-processing is used to reduce speckle noise and the influence of short-term local changes to obtain the accurate contour of the glacial lake.

[0014] Furthermore, the adaptive threshold method is Threshold isodata, minimum, otsu, or yen, and the clustering method is Cascade Kmeans, Kmeans, LVQ, SNIC Kmeans, or Xmeans.

[0015] Furthermore, the specific process of step 4 is as follows: Step 4.1: Construct a multi-source error detection system including sliding window detection, seasonal variation detection, and standard deviation outlier detection to screen out outliers in the precise contour of the glacial lake, wherein: The sliding window detection is represented as follows: In the formula: Indicates the first The area corresponding to a quarterly glacial lake; This is the threshold for the sliding window, typically set to 0.5. The seasonal change detection is represented as follows: In the formula: This represents the average quarterly area of ​​glacial lakes between 2015 and 2024. The threshold for abnormal seasonal variations is typically set to 0.2. The standard deviation outlier detection is expressed as: In the formula: and These represent the mean and standard deviation of the annual glacial lake area, respectively. Step 4.2: Construct an isolated forest model and perform secondary outlier identification, represented as: In the formula: For initial outlier estimation; These are the finally detected outliers; Represents sample points based on The score for an isolated forest of trees; Represents a threshold; This represents an outlier value in the sliding window; This represents seasonal outliers; Outlier values ​​representing the standard deviation; Step 4.3: Replace outliers with the average value of the same quarter in adjacent years. If outliers exist in the same quarter of adjacent years, the valid value from the same quarter of the most recent year will be used to replace them. ， Finally, the time-series ice lake area data was obtained after quality control.

[0016] Furthermore, the specific process of step 5 is as follows: Step 5.1: Based on the identification vector, extract salient targets from the time-series ice lake area data, analyze the pixel features of the salient targets and the distribution features of local SAR images, and make a comparative judgment. Step 5.2: Calculate the difference between the mean pixel value of significant targets and the mean value of local SAR images. The difference follows an approximately normal distribution with a mean of -4.81 and a standard deviation of 3.01. When the difference does not exceed 1.5 times the standard deviation, it is determined to be a glacial lake area, and proceed to step 5.3; when the difference exceeds 1.5 times the standard deviation, it is determined to be a non-glacial lake area, and proceed to step 5.4. Step 5.3 Identify newly added glacial lakes: Select summer and autumn image data. If no glacial lake was detected in the initial year, but the pixel features of the area meet the glacial lake criteria in subsequent years, it is determined to be a newly added glacial lake. Step 5.4 Identify vanished glacial lakes: If the pixel features of a certain glacial lake change to non-glacial lake features over time, and the area remains non-glacial lake in subsequent years, it is determined to be a vanished glacial lake. Step 5.5: Perform evolutionary analysis based on the rate of change of glacial lake area, where: the annual rate of change of area is expressed as: The quarterly area change rate is expressed as: In the formula: Annual rate of change; This represents the quarterly rate of change. For the first Annual ice lake area ； For the first The area of ​​the frozen lake in a certain season of the year ； Step 5.6: Based on the output results of steps 5.3 to 5.5, comprehensively analyze the evolution of the glacial lake to obtain the analysis results of the glacial lake evolution.

[0017] The present invention has the following beneficial technical effects: First, the proposed "optical global positioning combined with SAR local contour" method fully leverages the advantages of optical images' clear-sky features and SAR images' ability to penetrate clouds and fog, achieving synergistic and complementary advantages between optics and SAR. Accurate location of the ice lake is achieved through stacking multiple temporal optical images, providing prior information for SAR local extraction and effectively avoiding interference from confounding factors during SAR global extraction. Second, based on the YOLOv9-E target detection framework, transfer learning is performed using MSCOCO pre-trained weights, combining deep learning and transfer learning. By freezing low-level features and fine-tuning high-level features, rapid convergence is achieved under small sample conditions. Third, addressing the issue of ice lakes being easily missed as small-to-medium-sized targets, a missed detection error function is innovatively introduced. This function adjusts the loss function... The paper proposes several key technologies: First, it significantly reduces the false negative rate by using a weighted approach, effectively improving the detection rate of glacial lakes and providing more complete candidate targets for subsequent manual review. Second, it proposes a weight determination method based on maximizing the intersection-union ratio (IU / UWR), integrating multiple unsupervised algorithms from adaptive thresholding and clustering methods. It evaluates model stability by normalizing the coefficient of variation and assigns reasonable weights to different algorithms, fully leveraging the advantages of each algorithm in handling complex boundaries and image noise. Third, it constructs a buffer radius calculation formula and adaptively determines the buffer radius, making the localized SAR image exhibit a clear bimodal distribution characteristic, significantly improving the separability of glacial lakes from the background. Fourth, it automatically identifies newly added and disappeared glacial lakes by comparing the salient target pixel features contained in the vector with local SAR images, and achieves a comprehensive analysis of glacial lake evolution by combining the glacial lake area change rate. Attached Figure Description

[0018] Figure 1(a) is a flowchart of steps 1 to 3 of the present invention; Figure 1(b) is a flowchart of steps 4 to 5 of the present invention; Figure 2 This is an analysis diagram of the characteristics and obfuscation features of glacial lakes in optical remote sensing images; Figure 3 This is an analysis diagram of the characteristics and obfuscation features of glacial lakes in SAR remote sensing images.

[0019] Figure 4 This is a schematic diagram of cloud mask filtering based on dark channels; Figure 5 This is a comparison chart of the mean changes in average accuracy based on different migration methods in YOLOv9-E; Figure 6 It is the overall error distribution of the YOLOv9-E framework under different missed detection loss weights; Figure 7 This involves extracting the crossover ratio (CRO) distribution and filtering of ice lakes under different conditions. Figure 8 This is a time-series evolution diagram of glacial lakes in southeastern Tibet from 2015 to 2024; Figure 9 This is a statistical analysis chart of newly added and disappeared glacial lakes. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to specific embodiments. These descriptions are for explanation purposes only and are not intended to limit the scope of the invention.

[0021] Referring to Figures 1(a) and 1(b), a time-series intelligent extraction method for ice lake area based on optical-SAR collaboration includes the following steps: Step 1: Collect multi-temporal optical remote sensing images. Based on existing ice lake cataloging data, construct a deep learning object detection dataset containing ice lake samples in Pascal VOC format. Generate an ice lake sample dataset suitable for model training and validation. The specific process is as follows: Step 1.1: Obtain existing ice lake catalog data, merge ice lake catalog data from different data sources, and obtain the initial dataset of ice lake samples; Step 1.2: Based on the distribution characteristics of the Sentinel-1 GRD data, construct a quality assessment system for glacial lake samples. Specifically, use a Gaussian mixture model to determine whether the glacial lake samples in the initial dataset exhibit a bimodal distribution. If so, proceed to step 1.3; otherwise, classify them as erroneous samples and delete them. The bimodal distribution is expressed by the formula: In the formula: Represents glacial lake sample data The probability density function; Indicates the number of Gaussian components; Indicates the first The mixing weights of Gaussian components; Indicates the first The mean vector of Gaussian components; Indicates the first The covariance matrix of Gaussian components; Step 1.3: Calculate the difference between the bimodal distribution of the glacial lake sample and the correct glacial lake characteristic distribution using Kullback-Leibler divergence. , is represented as: In the formula: Represents distribution Relative to distribution of Divergence; The distribution of characteristics of the corresponding glacial lakes is correct; Corresponding to the bimodal distribution of the glacial lake characteristics to be detected, when When the value is greater than 0.1, Mark as an error sample and delete; otherwise, proceed to step 1.4. Step 1.4: Construct a cloud and fog detection method based on the RGB dark channel prior, expressed as: In the formula: Representative position Color channel Pixel values; Indicated by The local neighborhood centered on; when If the value is greater than 0.5, the sample is determined to be a cloud-polluted image and deleted; otherwise, proceed to step 1.5. Step 1.5: After excluding cloud-polluted images and erroneous samples in steps 1.2 to 1.4, and combining manually vectorized supplementary samples, a deep learning object detection dataset containing ice lake samples in Pascal VOC format is obtained, and it is divided into ice lake sample datasets suitable for model training and validation. Step 2: Based on the improved YOLOv9-E object detection framework, and using transfer learning and missed detection error optimization methods, global localization of ice lakes under cloudy conditions is achieved. The specific process is as follows: Step 2.1: Using the YOLOv9-E framework as the base model, the weights of the YOLOv9-E framework are pre-trained using the MSCOCO dataset and transfer learning is performed. The deep learning object detection dataset in Pascal VOC format for glacial lake samples is combined. By freezing the low-level features of layers 0-28 in the pre-trained model, only the high-level features and classification layers are fine-tuned to adapt the model to the features of remote sensing images. The YOLOv9-E, under a 640-pixel input condition, exhibits strong sensitivity for small target detection. Its Intersection over Union (IoU), Precision (P), Recall (R), and Average Precision (AP) are calculated using the following formulas: In the formula: This represents the correct detection target; This represents a misjudgment of the target; Represents the target that was missed; The first representing the sequence item; The first representing the sequence - 1 item; Representing the The weight of the item; Represents the number of terms in the sequence; express Average accuracy under threshold ； Represents average accuracy ； represent Threshold range; Step 2.2: Construct the missed detection function Sum of error functions , used to optimize the loss, is expressed as: In the formula: This represents the false negative rate. This is due to the error of missed detection; The detection frame as ;Will The detection frame as ; The weight for missed detection loss is usually set to 1.0; Step 2.3: Fusion of all remote sensing data: Synthesize the image data from July to September in each image period. The YOLOv9-E framework outputs detection bounding boxes for each synthesized quarterly image data period. By spatially associating the detection bounding boxes from multiple periods, the confidence of the ice lake detection boxes is recorded. Non-maximum suppression is used in overlapping areas to obtain high-confidence positioning information and output the coordinates of the ice lake bounding box. Step 3: Based on the coordinates of the ice lake boundary box, perform localization processing on the SAR image and construct an adaptive ice lake extraction model that integrates multiple unsupervised methods to output the accurate outline of the ice lake. The specific process is as follows: Step 3.1: Perform SAR image localization based on the coordinates of the ice lake boundary box. This involves constructing a buffer using the minimum envelope rectangle and performing SAR image localization. The formula for calculating the buffer radius is as follows: In the formula: The adjustment factor has a default value of 2. This represents the ratio of the shorter side to the longer side of the coordinate frame. This represents the length of the main diagonal of the coordinate frame; The radius of the buffer zone; Step 3.2: Perform histogram adjustment on the localized SAR image, truncating 1% of the pixel values ​​before and after the cutoff to optimize the data distribution characteristics; Step 3.3: Integrate multiple unsupervised algorithms, including adaptive thresholding methods (e.g., Threshold isodata, minimum, Otsu, or Yen) and clustering methods (e.g., Cascade Kmeans, Kmeans, LVQ, SNIC Kmeans, or Xmeans), to obtain a fused multivariate unsupervised adaptive ice lake extraction model. Based on maximizing the intersection-over-union ratio, determine the weights of the multivariate unsupervised adaptive ice lake extraction model. The loss function is expressed as: In the formula: The number of algorithms remaining after filtering: For the weights of each algorithm and ; For the first The segmentation result of each ice lake after weighted combination of various algorithms; For the first A truth mask for an ice lake; Represents any unsupervised algorithm; Indicates weight; Step 3.4: Based on the start dates of the four seasons in southeastern Tibet (February 25 for spring, May 31 for summer, September 15 for autumn, and November 28 for winter), the S1-GRD images for each season are fused using a multivariate unsupervised adaptive glacial lake extraction model. Then, morphological post-processing is used to reduce speckle noise and the influence of short-term local changes to obtain the accurate outline of the glacial lake. Step 4: Construct a multi-source error detection system integrating sliding window detection, seasonal variation detection, and standard deviation outlier detection. Use the isolated forest algorithm for outlier identification and correction to obtain quality-controlled time-series glacial lake area data. The specific process is as follows: Step 4.1: Construct a multi-source error detection system to filter out outliers in the precise contour of the glacial lake. The multi-source error detection system includes sliding window detection, seasonal variation detection, and standard deviation outlier detection, among which: The sliding window detection is represented as follows: In the formula: Indicates the first The area corresponding to a quarterly glacial lake; This is the threshold for the sliding window, typically set to 0.5. The seasonal change detection is represented as follows: In the formula: This represents the average quarterly area of ​​glacial lakes between 2015 and 2024. The threshold for abnormal seasonal variations is typically set to 0.2. The standard deviation outlier detection is expressed as: In the formula: and These represent the mean and standard deviation of the annual glacial lake area, respectively. Step 4.2: Construct an isolated forest model and perform secondary outlier identification, represented as: In the formula: For initial outlier estimation; These are the finally detected outliers; Represents sample points based on The score for an isolated forest of trees; Represents a threshold; This represents an outlier value in the sliding window; This represents seasonal outliers; Outlier values ​​representing the standard deviation; Step 4.3: Replace outliers with the average value of the same quarter in adjacent years. If outliers exist in the same quarter of adjacent years, the valid value from the same quarter of the most recent year will be used to replace them. Finally, the time-series ice lake area data after quality control was obtained; Step 5: Based on time-series glacial lake area data, identify newly added and disappearing glacial lakes using SAR local feature analysis and calculate the rate of change of glacial lake area. Conduct a comprehensive analysis of glacial lake evolution to obtain the glacial lake evolution analysis results. The specific process is as follows: Step 5.1: Based on the identification vector, extract salient targets from the time-series ice lake area data, analyze the pixel features of the salient targets and the distribution features of local SAR images, and make a comparative judgment. Step 5.2: Calculate the difference between the mean pixel value of significant targets and the mean value of local SAR images. The difference follows an approximately normal distribution with a mean of -4.81 and a standard deviation of 3.01. When the difference does not exceed 1.5 times the standard deviation, it is determined to be a glacial lake area, and proceed to step 5.3; when the difference exceeds 1.5 times the standard deviation, it is determined to be a non-glacial lake area, and proceed to step 5.4. Step 5.3 Identify newly added glacial lakes: Select summer and autumn image data. If no glacial lake was detected in the initial year, but the pixel features of the area meet the glacial lake criteria in subsequent years, it is determined to be a newly added glacial lake. Step 5.4 Identify vanished glacial lakes: If the pixel features of a certain glacial lake change to non-glacial lake features over time, and the area remains non-glacial lake in subsequent years, it is determined to be a vanished glacial lake. Step 5.5: Perform evolutionary analysis based on the rate of change of glacial lake area, where: the annual rate of change of area is expressed as: The quarterly area change rate is expressed as: In the formula: Annual rate of change; This represents the quarterly rate of change. For the first Annual glacial lake area; For the first The area of ​​a frozen lake in a certain season of a year; Step 5.6: Through the output results of steps 5.3 to 5.5, a comprehensive analysis of the evolution of glaciers and lakes is achieved, thereby obtaining the analysis results of the evolution of glaciers and lakes.

[0022] Figure 2 The study demonstrates the impact of optical imagery on glacial lake extraction. It shows that: 1) Clouds and fog obscure optical images, significantly reducing the visibility of glacial lakes; 2) Shadows formed on the ground due to cloud cover blocking sunlight are typically dark in tone, creating cloud shadows. These cloud shadows reduce the visibility of glacial lakes and cause the color and texture features of the glacial lakes to resemble those of the cloud shadows, increasing the risk of confusion; 3) Because the imagery uses passive remote sensing, the angle of solar incidence affects the formation of mountain shadows on adjacent slopes or the ground surface, leading to a decrease in the intensity of surface reflected signals.

[0023] Figure 3 The study demonstrates the impact of radar imagery on glacial lake extraction. It shows that: 1) Ground features with strong radar wave absorption characteristics (such as wet snow) result in weaker echo signals, leading to misjudgments; 2) Smooth ground surfaces cause specular reflection of radar waves, creating features similar to those of glacial lakes; 3) When radar waves encounter a slope facing away from the radar, the reflection direction deviates from the radar direction, resulting in reduced signal strength and generally darker pixel brightness on slopes facing away from the SAR; 4) Geometric distortion in SAR images causes shadowing effects on glacial lakes and surrounding ground features, affecting signal strength and clarity. Some shadows overlap with the glacial lake boundary, significantly reducing the accuracy of glacial lake extraction.

[0024] Figure 4The image was shown to be removed by dark channel prior, which can be seen that the dark channel prior weakens the pixel reflectance brightness in cloudless areas and the cloud features are significantly highlighted. Based on the normalized cloud transmittance constructed by the dark channel, pixels with transmittance less than 0.3 are marked as cloud pixels. Through spatial cross-correlation, samples and image blocks that have intersection with the cloud mask are retrieved and excluded.

[0025] Figure 5 The results show a comparison of the mean accuracy changes of different transfer learning methods based on YOLOv9-E. It can be seen that the initial mAP index of transfer learning is higher, and the model with fewer learnable parameters converges faster, reducing the possibility of model divergence and making the gradient optimization direction clearer. At the mAP index of 0.5:0.95, the differences among the four transfer learning methods (no freezing, freezing 0-28 layers, freezing 0-41 layers, and freezing 0-48 layers) are small. However, freezing 0-28 layers achieves the best mAP index.

[0026] Figure 6 The overall error distribution of the model under different false negative loss weights is shown. It can be seen that the accuracy of the model changes little under different weights, and the false negative rate of the model decreases significantly with the increase of the false negative loss weight, indicating that the localization performance of the bounding box is relatively stable and the false negative loss can effectively reduce the false negative rate.

[0027] Figure 7 The cross-union ratio distribution and filtering results of ice lake extraction under different conditions are shown. It can be seen that the present invention retains the optimal polarization combination and image fusion method of ascending and descending orbit SAR images by comparison and selection, thereby improving image quality and information extraction capability.

[0028] Figure 8 This paper demonstrates the temporal evolution results of glacial lakes in southeastern Tibet from 2015 to 2024 obtained by applying the method proposed in this invention to intelligent extraction of lake surface data. Figure 9 This paper demonstrates the application of the method proposed in this invention to intelligent lake surface extraction in southeastern Tibet, identifying newly added and disappeared glacial lakes from 2015 to 2024; through... Figure 8 and Figure 9 It can be seen that the time-series intelligent extraction method for glacial lake area proposed in this invention can achieve high-precision and high-frequency time-series monitoring of glacial lakes over a wide area in the cloudy and foggy southeastern Tibet region.

Claims

1. A method for intelligent extraction of temporal ice lake area based on optical-SAR synergy, characterized in that, Includes the following steps: Step 1: Collect multi-temporal optical remote sensing images and construct a deep learning target detection dataset containing glacial lake samples in PascalVOC format based on existing glacial lake catalog data, thereby generating a glacial lake sample dataset suitable for model training and validation. Step 2: Pre-train the weights of the YOLOv9-E framework using the MSCOCO dataset and perform transfer learning. Combine this with a deep learning object detection dataset in Pascal VOC format based on glacial lake samples. Freeze the low-level features of layers 0-28 in the pre-trained model and fine-tune only the high-level features and classification layers to adapt the model to the remote sensing image features. Then, construct a missed detection function. Sum of error functions This optimizes the loss, and then records the confidence of the ice lake detection box by spatially associating multi-period detection coordinate boxes. In the overlapping area, non-maximum suppression is used to obtain high-confidence localization information, and the coordinates of the ice lake bounding box are output. Step 3: Based on the coordinates of the ice lake boundary box, perform localization processing on the SAR image and construct an adaptive ice lake extraction model that integrates multiple unsupervised methods to output the accurate outline of the ice lake. Step 4: Construct a multi-source error detection system that integrates sliding window detection, seasonal change detection, and standard deviation outlier detection to screen out outliers in the precise contour of the glacial lake. Then, use the isolated forest algorithm for secondary outlier identification and correction to obtain the quality-controlled time-series glacial lake area data. Step 5: Based on time-series glacial lake area data, identify newly added and disappeared glacial lakes based on SAR local feature analysis and calculate the rate of change of glacial lake area. Comprehensively analyze the evolution of glacial lakes to obtain the glacial lake evolution analysis results.

2. The intelligent extraction method for time-series ice lake area based on optical-SAR synergy according to claim 1, characterized in that, The specific process of step 1 is as follows: Step 1.1: Obtain existing ice lake catalog data, merge ice lake catalog data from different data sources, and obtain the initial dataset of ice lake samples; Step 1.2: Based on the distribution characteristics of the Sentinel-1 GRD data, construct a quality assessment system for glacial lake samples. Specifically, use a Gaussian mixture model to determine whether the glacial lake samples in the initial dataset exhibit a bimodal distribution. If so, proceed to step 1.3; otherwise, classify them as erroneous samples and delete them. The bimodal distribution is expressed by the formula: In the formula: Represents glacial lake sample data The probability density function; Indicates the number of Gaussian components; Indicates the first The mixing weights of Gaussian components; Indicates the first The mean vector of Gaussian components; Indicates the first The covariance matrix of Gaussian components; Step 1.3: Calculate the difference between the bimodal distribution of the glacial lake sample and the correct glacial lake characteristic distribution using Kullback-Leibler divergence. , is represented as: In the formula: Represents distribution Relative to distribution of Divergence; The distribution of characteristics of the corresponding glacial lakes is correct; Corresponding to the bimodal distribution of the glacial lake characteristics to be detected, when When the value is greater than 0.1, Standard Record it as an error sample and delete it; otherwise, proceed to step 1.

4. Step 1.4: Construct a cloud and fog detection method based on the RGB dark channel prior, expressed as: In the formula: Representative position Color channel Pixel values; Indicated by The local neighborhood centered on; when If the value is greater than 0.5, the sample is determined to be a cloud-polluted image and deleted; otherwise, proceed to step 1.

5. Step 1.5: After excluding cloud-polluted images and erroneous samples in steps 1.2 to 1.4, and combining manually vectorized supplementary samples, a deep learning object detection dataset containing ice lake samples in Pascal VOC format is obtained, and it is divided into ice lake sample datasets suitable for model training and validation.

3. The intelligent extraction method for time-series ice lake area based on optical-SAR synergy according to claim 1, characterized in that, The specific process of step 2 is as follows: Step 2.1: Use the MSCOCO dataset to pre-train the weights of the YOLOv9-E framework and perform transfer learning. Combine the deep learning object detection dataset in Pascal VOC format for glacial lake samples. Freeze the low-level features of layers 0-28 in the pre-trained model and fine-tune only the high-level features and classification layer to adapt the model to the features of remote sensing images. The formulas for calculating the cross-union ratio, precision, recall, and average precision of YOLOv9-E are expressed as follows: In the formula: This represents the correct detection target; This represents a misjudgment of the target; Represents the target that was missed; The first representing the sequence item; The first representing the sequence -1 item; Representing the The weight of the item; Represents the number of terms in the sequence; express Average accuracy at the threshold; Represents average accuracy; represent Threshold range; Step 2.2: Construct the missed detection function Sum of error functions , used to optimize the loss, is expressed as: In the formula: This represents the false negative rate. This is due to the error of missed detection; The detection frame as ;Will The detection frame as ; The weight for missed detection loss is usually set to 1.0; Step 2.3: Synthesize the third quarter image data in each image period. The YOLOv9-E framework outputs detection bounding boxes for each synthesized quarterly image data period. By spatially associating the detection bounding boxes from multiple periods, the confidence of the ice lake detection boxes is recorded. In the overlapping areas, non-maximum suppression is used to obtain high-confidence localization information, and the coordinates of the ice lake bounding box are output.

4. The intelligent extraction method for temporal ice lake area based on optical-SAR synergy according to claim 1, characterized in that, The specific process of step 3 is as follows: Step 3.1: Construct a buffer using the minimum envelope rectangle to perform SAR image localization. The formula for calculating the buffer radius is as follows: In the formula: The adjustment factor has a default value of 2. This represents the ratio of the shorter side to the longer side of the coordinate frame. This represents the length of the main diagonal of the coordinate frame; The radius of the buffer zone; Step 3.2: Perform histogram adjustment on the localized SAR image, truncating 1% of the pixel values ​​before and after the cutoff to optimize the data distribution characteristics; Step 3.3: Integrate multiple unsupervised algorithms, including adaptive thresholding and clustering methods, to obtain a multivariate unsupervised adaptive ice lake extraction model. Based on maximizing the intersection-over-union ratio (IoU), determine the weights of the multivariate unsupervised adaptive ice lake extraction model. The loss function is expressed as: In the formula: The number of algorithms remaining after filtering: For the weights of each algorithm and ; For the first The segmentation result of each ice lake after weighted combination of various algorithms; For the first A truth mask for an ice lake; Represents any unsupervised algorithm; Indicates weight; Step 3.4: Based on the start dates of the four seasons in southeastern Tibet (February 25 for spring, May 31 for summer, September 15 for autumn, and November 28 for winter), the S1-GRD images for each season are mean-fused using a multivariate unsupervised adaptive glacial lake extraction model. Then, morphological post-processing is used to reduce speckle noise and the influence of short-term local changes to obtain the accurate contour of the glacial lake.

5. The intelligent extraction method for temporal ice lake area based on optical-SAR synergy according to claim 4, characterized in that, The adaptive threshold method is Threshold isodata, minimum, otsu, or yen, and the clustering method is Cascade Kmeans, Kmeans, LVQ, SNIC Kmeans, or Xmeans.

6. The intelligent extraction method for temporal ice lake area based on optical-SAR synergy according to claim 1, characterized in that, The specific process of step 4 is as follows: Step 4.1: Construct a multi-source error detection system including sliding window detection, seasonal variation detection, and standard deviation outlier detection to screen out outliers in the precise contour of the glacial lake, wherein: The sliding window detection is represented as follows: In the formula: Indicates the first The area corresponding to a quarterly glacial lake; This is the threshold for the sliding window, typically set to 0.

5. The seasonal change detection is represented as follows: In the formula: This represents the average quarterly area of ​​glacial lakes between 2015 and 2024. The threshold for abnormal seasonal variations is typically set to 0.

2. The standard deviation outlier detection is expressed as: In the formula: and These represent the mean and standard deviation of the annual glacial lake area, respectively. Step 4.2: Construct an isolated forest model and perform secondary outlier identification, represented as: In the formula: For initial outlier estimation; These are the finally detected outliers; Represents sample points based on The score for an isolated forest of trees; Represents a threshold; This represents an outlier value in the sliding window; This represents seasonal outliers; Outlier values ​​representing the standard deviation; Step 4.3: Replace outliers with the average value of the same quarter in adjacent years. If outliers exist in the same quarter of adjacent years, the valid value from the same quarter of the most recent year will be used to replace them. Finally, the time-series ice lake area data after quality control was obtained.

7. The intelligent extraction method for temporal ice lake area based on optical-SAR synergy according to claim 1, characterized in that, The specific process of step 5 is as follows: Step 5.1: Based on the identification vector, extract salient targets from the time-series ice lake area data, analyze the pixel features of the salient targets and the distribution features of local SAR images, and make a comparative judgment. Step 5.2: Calculate the difference between the mean pixel value of significant targets and the mean value of local SAR images. The difference follows an approximately normal distribution with a mean of -4.81 and a standard deviation of 3.

01. When the difference does not exceed 1.5 times the standard deviation, it is determined to be a glacial lake area, and proceed to step 5.3; when the difference exceeds 1.5 times the standard deviation, it is determined to be a non-glacial lake area, and proceed to step 5.

4. Step 5.3 Identify newly added glacial lakes: Select summer and autumn image data. If no glacial lake was detected in the initial year, but the pixel features of the area meet the glacial lake criteria in subsequent years, it is determined to be a newly added glacial lake. Step 5.4 Identify vanished glacial lakes: If the pixel features of a certain glacial lake change to non-glacial lake features over time, and the area remains non-glacial lake in subsequent years, it is determined to be a vanished glacial lake. Step 5.5: Perform evolutionary analysis based on the rate of change of glacial lake area, where: the annual rate of change of area is expressed as: The quarterly area change rate is expressed as: In the formula: Annual rate of change; This represents the quarterly rate of change. For the first Annual glacial lake area; For the first The area of ​​a frozen lake in a certain season of a year; Step 5.6: Based on the output results of steps 5.3 to 5.5, comprehensively analyze the evolution of the glacial lake to obtain the analysis results of the glacial lake evolution.

Images