A method for automatically extracting ice lake contours

Abstract

The present invention discloses a method for automatically extracting the outline of an ice surface lake, which belongs to the technical field of remote sensing image analysis and processing, and includes obtaining a multispectral remote sensing image of an ice surface lake covering a glacier area, automatically generating ice surface lake prompt points in combination with spectral features and shape features, obtaining the outline of the ice surface lake and related geometric information, and eliminating and masking the non-ice surface lake area. The present invention performs continuous reconstruction and refined interpolation operations on ice surface lake image blocks, and automatically generates ice surface lake prompt points using the spectral features and shape features of the ice surface lake in the remote sensing image, so that the model can generate semantically discernible high-precision ice surface lake edge segmentation results without relying on artificial prompt points. The present invention provides an efficient and reliable technical tool for the study of ice surface lakes, and is of great significance for revealing the evolution of ice surface lakes in their life cycle and their response to climate change, as well as the impact of ice surface lake drainage on ice dynamics and mass balance.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing image analysis and processing, and in particular relates to an automatic extraction method for an ice lake contour based on automatic generation of prompt points by fusion of multiple features. Background Art

[0002] Glacial lakes are formed in low-lying areas by meltwater from the ice surface due to thermal effects on the glacier surface. The low albedo of the water in the glacial lake can form a positive feedback mechanism, which in turn exacerbates the melting of surrounding ice and snow. The meltwater from the ice surface is transported into the land area in front of the ice through the glacial water system or directly flows into the ocean, leaving the ice sheet, directly causing the loss of ice sheet material; some glacial lakes will enter the interior of the ice sheet through rapid drainage, driving the subglacial hydrological system, thereby affecting the movement of the ice sheet and indirectly affecting the material balance of the ice sheet. Material loss caused by surface meltwater such as glacial lakes is one of the important reasons for the mass loss of the Greenland Ice Sheet. It is crucial to understand the location of the formation of glacial lakes and how their distribution evolves over time. The emergence and continuous development of satellite remote sensing technology has provided a new means for large-scale glacial hydrological investigation and research.

[0003] At present, there are three main methods for extracting ice lakes from remote sensing images: manual delineation, threshold extraction, and machine learning. Among them, deep learning methods, as a more advanced machine learning technology, can capture more complex and subtle features of ice lakes, and have shown excellent performance in ice lake extraction. At present, the use of deep learning methods to extract ice lakes has higher accuracy than other methods, but there are still the following three difficulties: first, a large amount of labeled data is required; second, due to the diversity of glacial geographical environments, the model will face the problem of insufficient generalization ability and is difficult to adapt to large-scale ice lake monitoring; third, the initial convolution layer of the convolutional neural network usually focuses on larger and obvious features, and has weaker ability to extract smaller and less obvious features. In recent years, large artificial intelligence models are large neural networks pre-trained based on large-scale data, which have significant advantages over traditional deep learning models in terms of zero-sample migration capabilities, image encoding and decoding efficiency, and generalization capabilities. Last year, the Segment Anything Model (SAM) proposed by Meta AI in 2023 was the first basic model in the field of image processing that can segment any object from an image without additional training, and showed excellent generalization ability in natural images. However, its heavy reliance on artificial cues has greatly limited its application in the field of remote sensing glacier monitoring. How to combine the spectral and shape characteristics of ice surface lakes with artificial intelligence large models to automatically segment ice surface lakes on remote sensing images has remained largely unexplored and verified. Summary of the invention

[0004] In view of the problems existing in the prior art, the present invention provides a method for automatically extracting the contour of an ice lake. The present invention determines whether there is an ice lake area in a window based on the spectral characteristics and shape characteristics of the ice lake, thereby automatically generating ice lake prompt points and further completing the contour extraction of the ice lake.

[0005] The present invention adopts the following technical solutions:

[0006] A method for automatically extracting the outline of an ice lake comprises the following steps:

[0007] Step 1: Obtain multispectral remote sensing images of glacial lakes covering the glacier area. The multispectral remote sensing images of glacial lakes include red band, green band, blue band and near-infrared band.

[0008] Step 2: Generate a false color remote sensing image using the red band, blue band, and near-infrared band in step 1, and divide the false color remote sensing image into small blocks with overlapping areas called ice lake image blocks. Perform continuous reconstruction on each ice lake image block and combine it with a high-precision interpolation algorithm to transform the size of each ice lake image block to twice its original size while maintaining the original aspect ratio, thereby improving the accuracy of the boundary features of small ice lakes.

[0009] Step 3: Use the improved normalized water index to transform the size of the ice lake image block The water body enhancement index image block is calculated and the specific calculation formula is as follows:

[0010] ;

[0011] Among them, B represents the blue band, R represents the red band, and NIR represents the near infrared band.

[0012] Step 4: Set the threshold T and perform binarization processing on the water body enhancement index image block obtained in step 3 to obtain a binary image.

[0013] Step 5: Set a rectangular structure window and slide it over the binary image calculated in step 4 to generate ice lake prompt points.

[0014] Step 6: Generate true color remote sensing images using the red band, green band, and blue band in step 1, and use a sliding window to divide the true color remote sensing images into small blocks with overlapping areas called real image blocks. Perform continuous reconstruction on each ice lake image block and combine it with a high-precision interpolation algorithm to transform the size of each ice lake image block to twice its original size while maintaining the original aspect ratio.

[0015] Step 7: Construct an automatic extraction model for ice surface lakes based on the Segment Anything Model; input the real image patches after size transformation in Step 6 and the ice surface lake prompt points obtained in Step 5 into the automatic extraction model for ice surface lakes to obtain the ice surface lake contour mask result of the real image patches.

[0016] Step 8: Restore the ice surface lake contour mask result of the real image patches to the size of the original divided patches, then map the ice surface lake mask result after size restoration back to the coordinate system of the corresponding original multi-spectral remote sensing image of the ice surface lake, and merge them to generate a complete binary ice surface lake mask result.

[0017] Preferably, Step 2 is specifically as follows: Use a sliding window to divide the false-color remote sensing image into ice surface lake image patches of size L×W pixels with a certain overlapping area. On the premise of maintaining the original aspect ratio, perform continuous reconstruction and interpolation operations on each ice surface lake image patch so that the size of each ice surface lake image patch is transformed to 2L×2W pixels to enhance the edge details and features of small ice surface lakes in the image patches.

[0018] Preferably, the process of binarizing the water body enhancement index image patches obtained in Step 3 in Step 4 is as follows: Judge the size of each pixel with T. If >T, then mark this pixel as an ice surface lake water body pixel; if <T, then mark this pixel as an ice surface lake non-water body pixel, where the ice surface lake water body pixel is assigned a value of 1 and the ice surface lake non-water body pixel is assigned a value of 0 to generate a binary image.

[0019] Preferably, Step 5 specifically includes:

[0020] Step 5.1: Set a rectangular structural window with a size of 5×5. The rectangular structural window starts from the upper left corner of the binary image obtained in Step 4 and slides on the image at a step size of 5 pixels to cover all areas of the image. For each sliding window, calculate the proportion of the area of the ice surface lake water body pixel region in the window to the entire window. The calculation formula is as follows:

[0021] ;

[0022] where P represents the proportion of ice surface lake water body pixels in the window, and N represents the number of ice surface lake water body pixels in the window.

[0023] Step 5.2: If P>80% in the current window, it is considered that there is an ice surface lake area in this window, and record the center coordinates of this window as the ice surface lake prompt points. The calculation formula for the center coordinates is:

[0024] ;

[0025] Among them, x and y represent the coordinates of the upper left corner of the current window, and Center represents the center coordinates.

[0026] Preferably, step 6 is specifically as follows: using a sliding window to divide the real image blocks into L×W pixels in size with a certain overlapping area, and performing continuous reconstruction and interpolation operations on each real image block while maintaining the original aspect ratio, so that the size of each real image block is transformed to 2L×2W pixels, so as to enhance the edge details and features of small ice lakes in the image blocks.

[0027] Preferably, step 7 specifically includes:

[0028] Step 7.1: The resized real image block obtained in step 6 is independently input into the ice lake automatic extraction model for detection. The Vision Transformer network of the image encoder module in the model is used to capture the structural information of the pixel points and neighborhood points of the ice lake according to the ice lake features, thereby extracting the local edge features and overall contour shape information of the ice lake, and generating the image features of the ice lake.

[0029] Step 7.2: Input the corresponding ice surface lake prompt points obtained in step 5 into the prompter module of the ice surface lake automatic extraction model for encoding, and the information of these ice surface lake coordinate prompt points is passed to the mask decoder module of the model.

[0030] Step 7.3: The ice lake image features and the corresponding ice lake cue point information are interconnected in the mask decoder module through the cross-attention mechanism, so that the two can work together to automatically generate an ice lake contour mask result for the real image block that conforms to the current cue point information.

[0031] Preferably, step 8 is specifically:

[0032] The mask result of the ice lake contour of the real image block is restored to the size of the original block to maintain spatial consistency, and the mask result is mapped back to the coordinate system of the corresponding original ice lake multispectral remote sensing image using the position information of the real image block, and the complete ice lake detection result is generated after merging; in the overlapping area of ​​the merging process, the ice lake mask value with the highest confidence is first retained, and then the degree of overlap between other ice lake masks and the currently selected ice lake mask is checked one by one; if the overlap rate exceeds the preset threshold of 70%, they are considered to belong to the same ice lake target, and then the ice lake detection frame with higher confidence is retained and the ice lake detection frame with lower confidence is removed to avoid repeated detection of the same ice lake target; the final mask value is the binarization result, where 0 represents the non-ice lake area and 1 represents the ice lake area.

[0033] The present invention has the following beneficial effects:

[0034] The present invention provides a method for automatically extracting the contour of an ice lake by automatically generating ice lake prompt points based on multi-feature fusion. The method utilizes continuous reconstruction and interpolation operations to effectively enhance the model's ability to capture smaller and less obvious ice lake contour features. The method utilizes the spectral characteristics and shape characteristics of the ice lake on remote sensing images to automatically generate ice lake prompt points, so that the model can generate semantically discernible high-precision ice lake edge segmentation results without relying on artificial prompt points. The method can complete the extraction of ice lake contours in an automated and high-precision manner on remote sensing images, providing an efficient and reliable technical tool for the study of ice lakes, which will be of great significance for revealing the evolution of ice lakes in their life cycle and their response to climate change, as well as the impact of ice lake drainage on ice dynamics and mass balance. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is the overall flow chart of the present invention.

[0036] Figure 2 This is a flow chart of the method for automatically generating ice surface lake prompt points by combining spectral features and shape features according to the present invention.

[0037] Figure 3 This is a flow chart for automatically extracting an ice lake in Example 1. DETAILED DESCRIPTION

[0038] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and specific examples.

[0039] Example 1: Combination Figures 1 to 3 , an automatic extraction method for ice surface lake contour, comprising the following steps: Step 1: obtaining a multispectral remote sensing image of an ice surface lake covering a glacier area, the multispectral remote sensing image of the ice surface lake including a red band, a green band, a blue band and a near infrared band.

[0040] Step 2: Based on the spectral characteristics of the ice surface lake water body, which has a higher reflectivity in the blue band, a lower reflectivity in the red band, and almost completely absorbs radiation in the near-infrared band, and its reflectivity is significantly lower than that of the ice surface, the red band, blue band, and near-infrared band in step 1 are used to generate a false color remote sensing image, and a sliding window is used to divide the false color remote sensing image into ice surface lake image blocks of size L×W pixels and a certain overlapping area, such as Figure 3 Under the premise of maintaining the original aspect ratio, each ice lake image block is continuously reconstructed and interpolated so that the size of each ice lake image block is transformed to 2L×2W pixels to enhance the edge details and features of small ice lakes in the image block.

[0041] Step 3: Use the improved normalized water index to transform the size of the ice lake image block Perform calculations to obtain water body enhancement index image blocks. The specific calculation formula is as follows:

[0042] ;

[0043] Where B represents the blue band, R represents the red band, and NIR represents the near-infrared band.

[0044] This index highlights the characteristics of water bodies absorbing near-infrared radiation while effectively suppressing the spectral interference characteristics of non-target water bodies such as blue ice and terrain shadows.

[0045] Step 4: Set a threshold T, which is determined by the maximum inter-class variance method.

[0046] Perform binarization on the water body enhancement index image blocks obtained in Step 3 to obtain a binary image. Specifically: Judge the size of each pixel's compared with T. If >T, then mark this pixel as an ice-covered lake water body pixel; if <T, then mark this pixel as an ice-covered lake non-water body pixel, where the ice-covered lake water body pixel is assigned a value of 1 and the ice-covered lake non-water body pixel is assigned a value of 0 to generate a binary image, as shown in the binary image in Figure 3

[0047] Step 5: Set a rectangular structural window and slide it through the binary image calculated in Step 4 to generate ice-covered lake hint points.

[0048] Specifically, it includes: Step 5.1: Since the shape of an ice-covered lake is usually oval and has a large area, set a rectangular structural window with a size of 5×5. The rectangular structural window starts from the upper left corner of the binary image obtained in Step 4 and slides on the image with a step size of 5 pixels, covering all areas of the image. For each sliding window, calculate the ratio of the area of the ice-covered lake water body pixels (pixel value is 1) in the window to the entire window. The calculation formula is as follows:

[0049] ;

[0050] Where P represents the ratio of the ice-covered lake water body pixels in the window, and N represents the number of ice-covered lake water body pixels in the window.

[0051] Step 5.2: If P>80% in the current window, it is considered that there is an ice-covered lake area in this window, and record the center coordinates of this window as the ice-covered lake hint point. The calculation formula for the center coordinates is:

[0052] ;

[0053] Where x and y respectively represent the coordinates of the upper left corner of the current window, and Center represents the center coordinates.​

[0054] Step 6: Generate true color remote sensing images using the red band, green band, and blue band in step 1, and use a sliding window to divide the real image blocks into L×W pixels with a certain overlapping area, such as Figure 3 Based on the real image blocks in the image, each ice lake image block is continuously reconstructed and combined with a high-precision interpolation algorithm to stretch the real image blocks to 2L×2W pixel size while maintaining the original aspect ratio, so as to enhance the edge details and features of small ice lakes in the real image blocks.

[0055] Step 7: Build an automatic ice lake extraction model based on the Segment Anything Model; send the real image block after the size transformation in step 6 and the ice lake prompt points obtained in step 5 to the automatic ice lake extraction model to obtain the ice lake contour mask result of the real image block.

[0056] Specifically including: Step 7.1: independently input the real image block after the size transformation in step 6 into the ice lake automatic extraction model for detection, and use the Vision Transformer network of the image encoder module in the model to capture the structural information of the pixel points of the ice lake and its neighborhood points according to the ice lake characteristics (such as the irregular edges of the ice lake, the curved shape of the ice lake, the texture and color of the ice lake boundary), thereby extracting the local edge features and overall contour shape information of the ice lake, and generating the image features of the ice lake.

[0057] Step 7.2: Input the corresponding ice lake cue points obtained in step 5 into the cue module of the ice lake automatic extraction model for encoding. The size of each cue point after passing through the cue encoder is 1×256. The information of these ice lake coordinate cue points is passed to the mask decoder module of the model.

[0058] Step 7.3: The ice lake image features and the corresponding ice lake prompt point information are interconnected in the mask decoder module through the cross-attention mechanism, so that the ice lake image features and the ice lake prompt information can interact with each other to guide the mask generation; after multiple layers of iteration of the cross-attention mechanism, the mask decoder further uses the multi-layer perceptron (MLP) to perform nonlinear processing on the extracted ice lake features; the multi-layer perceptron further refines the complex ice lake detail features and ice lake edge information by gradually adjusting and optimizing the feature representation, so that the generated ice lake mask boundary is clearer and more accurate, and finally automatically generates the ice lake contour mask result of the real image block that conforms to the current prompt point information, such as Figure 3 shown.

[0059] Step 8: Restore the ice lake contour mask result of the real image block to the size of the original block, then map the size-restored ice lake mask result back to the coordinate system of the corresponding original ice lake multispectral remote sensing image, and merge them to generate a complete ice lake mask binary result.

[0060] Specifically, the size of the ice lake contour mask result of the real image block is restored to the size of the original block to maintain spatial consistency, and the mask result is mapped back to the coordinate system of the corresponding original ice lake multispectral remote sensing image using the location information of the real image block, and the complete ice lake detection result is generated after merging; in the overlapping area of ​​the merging process, the ice lake mask value with the highest confidence is first retained, and then the degree of overlap between other ice lake masks and the currently selected ice lake mask is checked one by one; if the overlap rate exceeds the preset threshold of 70%, they are considered to belong to the same ice lake target, and then the ice lake detection frame with higher confidence is retained and the ice lake detection frame with lower confidence is removed to avoid repeated detection of the same ice lake target; the final mask value is the binarization result, where 0 represents the non-ice lake area and 1 represents the ice lake area.

[0061] Elimination of non-ice lake areas and mask optimization: For non-ice lake pixels (ice rivers, water-filled cracks, and slush, etc.) that still exist after the ice lake contour is extracted, mark all connected areas in the ice lake mask binarization result obtained in step 8, and generate the corresponding minimum circumscribed rectangle for each area. Then, based on the geometric properties of the rectangle (length and width), calculate its aspect ratio; for areas whose aspect ratio exceeds a specific threshold A, they can be determined as non-ice lake areas, so that the mask results of these areas are eliminated from the overall mask to optimize the extracted ice lake classification results, further improve the accuracy of the ice lake classification results, and finally generate high-quality ice lake binary mask results.

[0062] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for automatically extracting the outline of an ice lake, characterized in that: The following steps are involved: Step 1: Obtain multispectral remote sensing images of ice surface lakes covering the glacier area. The multispectral remote sensing images of ice surface lakes include red band, green band, blue band and near infrared band; Step 2: Generate a false color remote sensing image using the red band, blue band, and near infrared band in step 1, and divide the false color remote sensing image into small blocks with overlapping areas, called ice surface lake image blocks. Perform continuous reconstruction on each ice surface lake image block and combine it with a high-precision interpolation algorithm to transform the size of each ice surface lake image block to twice its original size while maintaining the original aspect ratio, thereby improving the accuracy of the boundary features of small ice surface lakes. Step 3: Use the improved normalized water index to transform the size of the ice lake image block The water body enhancement index image block is calculated and the specific calculation formula is as follows: ； Among them, B represents the blue band, R represents the red band, and NIR represents the near infrared band; Step 4: Set the threshold T, and perform binarization processing on the water body enhancement index image block obtained in step 3 to obtain a binary image; Step 5: Set a rectangular structure window and slide it on the binary image calculated in step 4 to generate ice lake prompt points; Specifically include: Step 5.1: Set a 5×5 rectangular window. The rectangular window starts from the upper left corner of the binary image obtained in step 4 and slides on the image with a step length of 5 pixels to cover all areas of the image. For each sliding window, calculate the proportion of the area of ​​the ice surface lake water pixel area in the window to the entire window. The calculation formula is as follows: ； Among them, P represents the proportion of ice-surface lake water pixels in the window, and N represents the number of ice-surface lake water pixels in the window; Step 5.2: If P>80% in the current window, it is considered that there is an ice lake area in the window, and the center coordinates of the window are recorded as the ice lake prompt point. The calculation formula of the center coordinates is: ； Among them, x and y represent the coordinates of the upper left corner of the current window, and Center represents the center coordinates; Step 6: Generate true color remote sensing images using the red band, green band, and blue band in step 1, and use a sliding window to divide the true color remote sensing images into small blocks with overlapping areas called real image blocks. Perform continuous reconstruction on each ice lake image block and combine it with a high-precision interpolation algorithm to transform the size of each ice lake image block to twice its original size while maintaining the original aspect ratio; Step 7: Build an automatic ice lake extraction model based on the Segment Anything Model; send the real image block after the size transformation in step 6 and the ice lake prompt points obtained in step 5 to the automatic ice lake extraction model to obtain the ice lake contour mask result of the real image block; Step 8: Restore the ice lake contour mask result of the real image block to the size of the original block, then map the size-restored ice lake mask result back to the coordinate system of the corresponding original ice lake multispectral remote sensing image, and merge them to generate a complete ice lake mask binarization result.

2. The method for automatically extracting the outline of an ice lake according to claim 1, characterized in that: Step 2 is as follows: Use a sliding window to divide the false color remote sensing image into ice lake image blocks of size L×W pixels with a certain overlapping area. Under the premise of maintaining the original aspect ratio, continuously reconstruct and interpolate each ice lake image block so that the size of each ice lake image block is transformed to 2L×2W pixels to enhance the edge details and features of small ice lakes in the image block.

3. The method for automatically extracting the outline of an ice lake according to claim 1, characterized in that: The process of binarizing the water body enhancement index image block obtained in step 3 in step 4 is as follows: Determine the with the size of T. If > T, then mark the pixel as an ice-covered lake water body pixel; if < T, then mark the pixel as an ice-covered lake non-water body pixel, where the ice-covered lake water body pixel is assigned a value of 1 and the ice-covered lake non-water body pixel is assigned a value of 0 to generate a binary image.

4. The method for automatically extracting the outline of an ice lake according to claim 1, characterized in that: Step 6 is as follows: Use a sliding window to divide the real image blocks into L×W pixels with a certain overlapping area. Under the premise of maintaining the original aspect ratio, continuously reconstruct and interpolate each real image block so that the size of each real image block is transformed to 2L×2W pixels to enhance the edge details and features of small ice lakes in the image blocks.

5. The method for automatically extracting the outline of an ice lake according to claim 1, characterized in that: Step 7 specifically includes: Step 7.1: The real image block after the size transformation obtained in step 6 is independently input into the ice lake automatic extraction model for detection, and the Vision Transformer network of the image encoder module in the model is used to capture the structural information of the pixel points of the ice lake and its neighborhood points according to the ice lake features, thereby extracting the local edge features and overall contour shape information of the ice lake, and generating the image features of the ice lake; Step 7.2: Input the corresponding ice surface lake prompt points obtained in step 5 into the prompter module of the ice surface lake automatic extraction model for encoding, and the information of these ice surface lake coordinate prompt points is transmitted to the mask decoder module of the model; Step 7.3: The ice lake image features and the corresponding ice lake cue point information are interconnected in the mask decoder module through the cross-attention mechanism, so that the two can work together to automatically generate an ice lake contour mask result for the real image block that conforms to the current cue point information.

6. The method for automatically extracting the outline of an ice lake according to claim 1, characterized in that: Step 8 is as follows: The mask result of the ice lake contour of the real image block is restored to the size of the original block to maintain spatial consistency, and the mask result is mapped back to the coordinate system of the corresponding original ice lake multispectral remote sensing image using the position information of the real image block. After merging, a complete ice lake detection result is generated; in the overlapping area of ​​the merging process, the ice lake mask value with the highest confidence is first retained, and then the overlap between other ice lake masks and the currently selected ice lake mask is checked one by one; If the overlap rate exceeds the preset threshold of 70%, they are considered to belong to the same ice lake target. Then the ice lake detection frame with high confidence is retained and the ice lake detection frame with low confidence is removed to avoid repeated detection of the same ice lake target. The final mask value is the binarization result, where 0 represents the non-ice lake area and 1 represents the ice lake area.

Images