Automatic detection and early warning method of tailings dry beach based on multi-sensor image joint processing

Abstract

The present invention discloses a tailings dry beach automatic detection and early warning method with multi-sensor image joint processing, which belongs to the field of image data processing. This method uses an unmanned aerial vehicle equipped with an RGB camera, a multi-spectral camera and a laser radar to collect RGB images, multi-spectral images and three-dimensional point cloud data, and then obtains a variety of feature data, and then obtains boundary information through fusion and random forest classifier, and then calculates the dry beach length and issues an early warning. The present invention not only improves the accuracy and efficiency of prediction, but also makes full use of the advantages of various types of data, so that tailings pond monitoring has stronger reliability and flexibility, and also has the advantages of good random forest classification effect, strong discrimination ability, and low false detection rate.

Description

Technical Field

[0001] The invention belongs to the field of image data processing, and in particular relates to an automatic detection and early warning method for dry tailings beaches by joint processing of multi-sensor images. Background Art

[0002] Tailings ponds are places used to store tailings produced by metal and non-metal mines after ore selection, waste produced during hydrometallurgy, or other industrial waste residues. Tailings ponds are built by damming the valley mouth or enclosing land. Their structure includes tailings dam system, tailings discharge system, flood discharge system, and backwater system. As mines continue to improve their management requirements for tailings ponds, monitoring technologies are becoming increasingly diversified, and monitoring points cover deformation displacement, seepage, dry beach, reservoir water level, precipitation, and other aspects.

[0003] Wet tailings ponds usually use hydraulic conveying to discharge tailings. Monitoring items for wet tailings ponds include dam displacement, wetting line, dry beach length and slope, precipitation and reservoir water level. Among them, monitoring of dry beach length helps to schedule the tail discharge port and control the discharge volume at different tail discharge points in the reservoir area, thereby effectively controlling the wetting line and dry beach length, avoiding the dry beach length being too low or the wetting line exceeding the dam control line due to uneven discharge, which affects the anti-sliding stability safety factor of the dam.

[0004] At present, the monitoring of the dry beach length of tailings ponds mainly relies on daily manual inspections and fixed-point monitoring by fixed cameras. When it is found that the dry beach length does not meet the requirements, an early warning is issued to remind the staff to take timely measures. However, manual inspections are highly subjective and greatly affected by various factors such as the environment, geography and weather, so they are inaccurate and inefficient. The limited monitoring range of fixed cameras often leads to many inspection contents being ignored or unrecognizable, making it difficult to detect safety hazards in a timely manner. Summary of the invention

[0005] The invention proposes an automatic detection and early warning method for dry tailings beach by multi-sensor image joint processing, the purpose of which is to solve the problems of poor accuracy, low efficiency and limited monitoring range of the existing method.

[0006] The technical solution of the present invention is as follows:

[0007] A tailings dry beach automatic detection and early warning method using multi-sensor image joint processing, comprising the following steps:

[0008] Step 1: Use a drone equipped with an RGB camera, a multispectral camera, and a lidar to take real-time photos of the tailings pond, and transmit the collected RGB images, multispectral images, and three-dimensional point cloud data to the server; the RGB images and multispectral images correspond to the same area and have the same size;

[0009] Step 2: After receiving the collected data, the server preprocesses the RGB image and the multispectral image respectively to obtain an RGB orthophoto image and a multispectral orthophoto image; the RGB orthophoto image and the multispectral orthophoto image have the same size;

[0010] Step 3: Obtain a digital elevation model based on the three-dimensional point cloud data;

[0011] Step 4: Obtain the starting boundary for calculating the dry beach length based on the digital elevation model;

[0012] Step 5: Based on the RGB orthoimage, multispectral orthoimage and digital elevation model, multiple features are extracted and fused, and then the random forest method is used to classify and predict the fusion results to obtain the termination boundary of the dry beach length calculation;

[0013] Step 6, obtaining the dry beach length according to the starting boundary and the ending boundary calculated by the dry beach length;

[0014] Step 7: Determine whether the length of the dry beach exceeds a preset range, and issue an early warning if so.

[0015] As a further improvement of the automatic detection and early warning method for tailings dry beach by joint processing of multi-sensor images: in step 4, based on the digital elevation model, the elevation of the sedimentary beach surface is used as the reference elevation, and a three-dimensional analysis algorithm is used to extract the contour line contour of the outer slope of the dam body in the digital elevation model corresponding to the reference elevation, and the contour line contour is used as the starting boundary for calculating the dry beach length.

[0016] As a further improvement of the tailings dry beach automatic detection and early warning method by multi-sensor image joint processing, the specific process of step 5 is as follows:

[0017] Step 5-1, extracting depth features and texture features from the RGB orthoimage, extracting index features from the multispectral orthoimage, and extracting elevation features about the termination boundary from the digital elevation model;

[0018] Step 5-2, the depth feature, texture feature, index feature and elevation feature are fused, and each pixel in the fused feature map corresponds to a one-dimensional feature vector;

[0019] Step 5-3: Input the one-dimensional feature vector corresponding to each pixel into the constructed random forest to obtain the prediction result corresponding to each pixel;

[0020] Step 5-4: Based on the prediction results of each pixel obtained by random forest, the termination boundary of the dry beach length calculation is determined.

[0021] As a further improvement of the tailings dry beach automatic detection and early warning method by multi-sensor image joint processing, the depth feature extraction method is as follows:

[0022] Step A-1, inputting the RGB orthophoto image into the image segmentation model to obtain the segmentation result of water body and non-water body, wherein the segmentation result is a probability map, wherein the probability value indicates the credibility that the corresponding pixel belongs to the water body; in the processing process of the image segmentation model, extracting the intermediate feature map generated by the intermediate layer of the model;

[0023] Step A-2: Fuse the probability map and the intermediate feature map by channel to obtain The depth features of the channels.

[0024] As a further improvement of the tailings dry beach automatic detection and early warning method by multi-sensor image joint processing: the image segmentation model in step A-1 is a U-Net Transformer model;

[0025] The RGB orthophoto image is input into the U-Net Transformer model to obtain the probability map, and the decoder of the U-Net Transformer model is used to obtain the probability map. The intermediate feature map of each channel.

[0026] As a further improvement of the tailings dry beach automatic detection and early warning method by multi-sensor image joint processing, the channel fusion in step A-2 refers to: splicing the probability map and the intermediate feature map by channel to obtain the depth feature;

[0027] Alternatively, the eigenvalues ​​of the intermediate feature map are weighted according to the probability map and then concatenated with the probability map by channel.

[0028] As a further improvement of the tailings dry beach automatic detection and early warning method of the multi-sensor image joint processing, the process of weighting the eigenvalues ​​of the intermediate feature map according to the probability map is as follows: for the first pixels, let the corresponding probability value in the probability map be , the eigenvalues ​​corresponding to the pixel in the intermediate feature map are , , then respectively according to calculate Weighted eigenvalues To replace the corresponding Obtaining a weighted intermediate feature map, and then splicing the weighted intermediate feature map with the probability map by channel to obtain the depth feature;

[0029] Weighted Eigenvalue The calculation method is: ;

[0030] in, According to the probability value The calculated weighted values ​​corresponding to the eigenvalues ​​of each channel are:

[0031] ;

[0032] In the formula, the peak point .

[0033] As a further improvement of the automatic detection and early warning method of tailings dry beach by multi-sensor image joint processing:

[0034] In step 5-2, the depth features, texture features, index features and elevation features are spliced ​​by channel, and each pixel in the spliced ​​feature map corresponds to a one-dimensional feature vector, which contains the features from the depth features, texture features, index features and elevation features. feature values.

[0035] As a further improvement of the tailings dry beach automatic detection and early warning method by multi-sensor image joint processing, the random forest used in step 5-3 is constructed as follows:

[0036] First, we construct each decision tree in the random forest by randomly selecting features, and then calculate the importance of each feature: ,in Refers to Features, Features The importance of is the number of decision trees constructed, Features In the The gain value in each decision tree is calculated; then, the features are sorted by importance, and the decision trees in the random forest are reconstructed again. When the nodes of the decision tree are split again, the features with the highest importance ranking in the currently selected feature subset are preferentially selected as the split features; the reconstructed random forest is used as the random forest for prediction.

[0037] As a further improvement of the tailings dry beach automatic detection and early warning method of the multi-sensor image joint processing: when using random forest to predict a certain pixel, first input the corresponding one-dimensional feature vector into each decision tree to obtain the prediction distribution of each decision tree, and set The prediction distribution obtained by the decision tree is ,in Indicates The decision tree predicts that the current input one-dimensional feature vector belongs to The probability of the categories, is the number of categories, including several categories under the water body category and several categories under the non-water body category; then calculate the weight of each decision tree for different categories:

[0038] ;

[0039] in, is the number of decision trees in the random forest, Indicates The decision tree The weight of each category;

[0040] Then, calculate the score of each category for the current input one-dimensional feature vector:

[0041] ;

[0042] in, For the The category with the highest score is the prediction result of the current pixel.

[0043] Compared with the prior art, the present invention has the following positive effects:

[0044] 1. The present invention combines RGB images, multispectral images and three-dimensional point clouds to obtain multiple features, fully integrates multi-source data, and then quantizes the integrated features and inputs them into the random forest to obtain the required boundary information. This not only improves the accuracy and efficiency of the prediction, but also makes full use of the advantages of various types of data, making the tailings pond monitoring more reliable and flexible.

[0045] 2. Based on the traditional random forest, the present invention introduces a feature selection optimization mechanism, which reconstructs the decision tree according to the importance of the features, thereby providing more splitting opportunities for more important features and improving the classification effect.

[0046] 3. Based on the traditional random forest, the present invention also introduces a multi-level weighted voting mechanism, which takes into account the prediction probability of each tree for all categories and distinguishes them, allowing the algorithm to retain more probability information, especially when the categories are close and the samples are difficult to distinguish, which can achieve a smoother and more robust effect.

[0047] 4. The present invention adds the elevation features obtained by the digital elevation model to the discriminant features of the random forest, which effectively constrains the prediction of the termination boundary and reduces the false detection rate of dry beach monitoring.

[0048] 5. In the feature fusion process, the present invention also performs peak weighted calculation on the feature values ​​of each channel according to the probability map obtained by the preliminary segmentation. For pixels that are preliminarily determined to belong to water bodies, the channels close to the decoder input end enjoy higher weight values, so that the model can pay more attention to abstract information when random forest fine classification is used subsequently. For pixels that are preliminarily determined to belong to non-water bodies, the channels close to the decoder output end enjoy higher weight values, so that the model can pay more attention to detail information when random forest fine classification is used subsequently, thereby effectively enhancing the discrimination ability of the random forest classifier when processing fine-grained features. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0050] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0051] like Figure 1 , a tailings dry beach automatic detection and early warning method using multi-sensor image joint processing, the steps comprising:

[0052] Step 1: Use a drone equipped with an RGB camera, a multispectral camera, and a lidar to take real-time photos of the tailings pond, and transmit the collected RGB images, multispectral images, and 3D point cloud data to the server. The RGB image and the multispectral image correspond to the same area and have the same size.

[0053] Step 2: After receiving the collected data, the server preprocesses the RGB image and the multispectral image to obtain the RGB orthophoto image and the multispectral orthophoto image. The RGB orthophoto image and the multispectral orthophoto image have the same size.

[0054] Specifically, the collected RGB images and multispectral images are uploaded to a fixed path on the server. When the server finds new images, it will automatically trigger the preprocessing program to stitch and correct the RGB images and multispectral images respectively to obtain the corresponding orthophoto images.

[0055] Step 3: Obtain a digital elevation model based on the three-dimensional point cloud data.

[0056] The specific process is:

[0057] Step 3-1, preprocessing: First, denoise the 3D point cloud data to remove outliers in the data; then perform filtering to eliminate unnecessary details and noise; finally, perform point cloud registration to ensure the spatial consistency of the data, thereby improving the overall accuracy of the data.

[0058] Step 3-2, ground point extraction: Use filtering methods based on parameters such as distance and slope, as well as multi-scale progressive filtering (PMF) technology to extract points representing the ground from the pre-processed point cloud. At the same time, perform sparse processing on the extracted ground points to reduce the amount of data and improve the efficiency of subsequent processing.

[0059] Step 3-3, interpolation processing: Use techniques such as Kriging interpolation or triangulated network interpolation (TIN) to convert the extracted ground points into a regular two-dimensional elevation grid.

[0060] Step 3-4, smoothing: Smooth the generated elevation grid to eliminate elevation changes and sharp edges and corners, and ensure the smoothness and continuity of the elevation grid.

[0061] Step 3-5, export data: export the processed elevation grid to common geographic data formats, such as ESRIGRID, GeoTIFF, etc., to complete the generation of the digital elevation model.

[0062] Step 4: Obtain the starting boundary for calculating the dry beach length based on the digital elevation model.

[0063] The specific process is as follows: based on the digital elevation model, the elevation of the sedimentary beach surface is used as the benchmark elevation, and the three-dimensional analysis algorithm is used to extract the contour line contour of the outer slope of the dam body in the digital elevation model corresponding to the benchmark elevation, and the contour line contour is used as the starting boundary for the calculation of the dry beach length.

[0064] Step 5: Based on the RGB orthoimage, multispectral orthoimage and digital elevation model, multiple features are extracted and fused, and then the random forest method is used to classify and predict the fusion results to obtain the termination boundary of the dry beach length calculation.

[0065] The specific process is:

[0066] Step 5-1: extract depth features and texture features from the RGB orthoimage, extract index features from the multispectral orthoimage, and extract elevation features about the termination boundary from the digital elevation model.

[0067] A. The extraction method of deep features is:

[0068] Step A-1, input the RGB orthophoto image into the image segmentation model to obtain the segmentation result of water body and non-water body, wherein the segmentation result is a probability map, in which the probability value represents the credibility that the corresponding pixel belongs to the water body; during the processing of the image segmentation model, the intermediate feature map generated by the intermediate layer of the model is extracted.

[0069] Specifically, the image segmentation model is a U-Net Transformer model, which combines the characteristics of the U-Net model and the advantages of the Transformer module. It is usually used in various image segmentation scenarios, including but not limited to medical image analysis, remote sensing image processing, and scene understanding in autonomous driving.

[0070] In this embodiment, the RGB orthophoto image is input into the U-Net Transformer model to obtain a probability map, and a decoder containing Since the feature map in the decoder is constructed by upsampling and concatenating with the corresponding layer of the encoder through skip connections, the channels close to the decoder input in the intermediate feature map retain higher abstract information, while the channels close to the decoder output retain more detailed information.

[0071] Step A-2: Fuse the probability map and the intermediate feature map by channel to obtain The depth features of the channels.

[0072] Fusion by channel means:

[0073] Method 1: Concatenate the probability map and the intermediate feature map by channel to obtain the depth feature.

[0074] Or use:

[0075] Method 2: Weight the eigenvalues ​​of the intermediate feature map according to the probability map and then concatenate them with the probability map by channel.

[0076] Specifically, for the first pixels, let the corresponding probability value in the probability map be , the eigenvalues ​​corresponding to the pixel in the intermediate feature map are , , then respectively according to calculate Weighted eigenvalues To replace the corresponding A weighted intermediate feature map is obtained, and then the weighted intermediate feature map and the probability map are spliced ​​by channel to obtain the depth feature.

[0077] Weighted Eigenvalue The calculation method is: ;

[0078] in, According to the probability value The calculated weighted values ​​corresponding to the eigenvalues ​​of each channel are:

[0079] ;

[0080] In the formula, the peak point That is, if the probability that a pixel belongs to a water body is greater, the channel close to the decoder input end enjoys a higher weight value, and the subsequent further classification will pay more attention to the abstract information in the intermediate feature map; conversely, the channel close to the decoder output end enjoys a higher weight value, and the subsequent further classification will pay more attention to the detailed information in the intermediate feature map.

[0081] B. Extraction of texture features:

[0082] In this embodiment, texture features are extracted from the RGB orthoimage using a gray level co-occurrence matrix.

[0083] Gray-level co-occurrence matrix (GLCM) is a feature extraction technique that describes texture information in detail by counting the grayscale values ​​of pixels in an image and their spatial distribution relationships. The core of GLCM is to analyze the second-order statistical characteristics of an image, including the roughness, directionality, and uniformity of the texture. The feature set extracted by GLCM, such as statistics such as contrast, uniformity, energy, and entropy, provides us with a comprehensive description of the distribution and structural characteristics of image textures.

[0084] C. Extraction of index features:

[0085] Based on the multispectral orthophoto, the NDVI index value and NDWI index value of each pixel are obtained to constitute the index characteristics of the two channels.

[0086] The calculation method of NDVI index value is: NDVI = (NIR-R) / (NIR + R); the calculation method of NDWI index value is: NDWI = (Green-NIR) / (NIR + R). Among them, NIR is the value of the pixel in the near infrared band of the multispectral orthophoto, R is the value of the pixel in the red band of the multispectral orthophoto, and Green is the value of the pixel in the green band of the multispectral orthophoto.

[0087] The Normalized Difference Vegetation Index (NDVI) helps to enhance the characteristics of vegetation attributes, distinguish between vegetation and non-vegetation, and assist in the confirmation of water bodies. The Normalized Difference Water Index (NDWI) helps to enhance the characteristics of water attributes, distinguish between water bodies and non-water bodies. The higher the value, the greater the possibility of the existence of water bodies.

[0088] D. Extraction of elevation features:

[0089] An initial feature map with the same size as the RGB orthoimage and the multispectral orthoimage is constructed. The slope value of each pixel in the initial feature map is calculated according to the digital elevation model. The slope value refers to the degree of inclination of the corresponding position of the pixel in the digital elevation model relative to the horizontal plane. The initial feature map is filled with the slope value of each pixel to obtain the elevation feature.

[0090] Step 5-2: Fuse the depth features, texture features, index features, and elevation features. Each pixel in the fused feature map corresponds to a one-dimensional feature vector.

[0091] Specifically, the depth features, texture features, index features and elevation features are spliced ​​by channel, and each pixel in the spliced ​​feature map corresponds to a one-dimensional feature vector, which contains the features from the depth features, texture features, index features and elevation features. feature values.

[0092] Step 5-3: Input the one-dimensional feature vector corresponding to each pixel into the constructed random forest to obtain the prediction result corresponding to each pixel.

[0093] Random forest is a classification method based on ensemble learning. It builds multiple decision trees and votes on their prediction results to determine the final classification result.

[0094] In a random forest, each decision tree is usually constructed by the following steps: First, the training set is randomly sampled with replacement to form a new sample set, which will be used to train a decision tree. Then, at each split node, a subset of all features is randomly selected, and then the optimal features and split points are found in the selected feature subset to split the node.

[0095] In this embodiment, the one-dimensional feature vector corresponding to each pixel contains eigenvalues ​​(the exponential feature contains two channels), so according to features to build a decision tree.

[0096] The present invention introduces a feature selection optimization mechanism when constructing a random forest: first, each decision tree in the random forest is constructed by randomly selecting features, and then the importance of each feature is calculated: ,in Refers to Features, Features The importance of is the number of decision trees constructed, Features In the The gain value in the decision tree. After that, the features are sorted by importance, and the decision trees in the random forest are rebuilt again. When splitting the nodes of the decision tree, the features with the highest importance in the currently selected feature subset are preferentially selected as the splitting features. The rebuilt random forest is used as the random forest for prediction.

[0097] When using random forest to predict a pixel, first input the corresponding one-dimensional feature vector into each decision tree to obtain the prediction distribution of each decision tree. The prediction distribution obtained by the decision tree is ,in Indicates The decision tree predicts that the current input one-dimensional feature vector belongs to The probability of the categories, is the number of categories, including several categories under the water body category and several categories under the non-water body category. Then calculate the weight of each decision tree for different categories:

[0098] ;

[0099] in, is the number of decision trees in the random forest, Indicates The decision tree The weight of each category.

[0100] Then, calculate the score of each category for the current input one-dimensional feature vector:

[0101] .

[0102] in, For the The category with the highest score is the prediction result of the current pixel.

[0103] Step 5-4: Based on the prediction results of each pixel obtained by random forest, the termination boundary of the dry beach length calculation is determined.

[0104] Specifically, the pixels predicted to belong to the water body in step 5-3 are filtered, and the pixels of the water body are converted into vector polygons. The boundary of the vector polygon close to the dam body is the termination boundary of the dry beach length calculation.

[0105] Step 6: Obtain the dry beach length based on the starting boundary and the ending boundary calculated from the dry beach length.

[0106] For each point in the starting boundary, the distance between it and each point in the ending boundary is calculated respectively, and then the shortest distance among all the obtained distances is taken as the length of the dry beach.

[0107] Step 7: Determine whether the length of the dry beach exceeds the preset range. If so, issue an early warning to indicate the risk that the expansion of the dry beach may pose to the stability of the dam, thereby providing real-time support for the safety management of the tailings pond.

[0108] It should be noted that it is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. The scope of the present invention is defined by the claims rather than the above description.

Claims

1. A tailings dry beach automatic detection and early warning method using multi-sensor image joint processing, characterized in that the steps include: Step 1: Use a drone equipped with an RGB camera, a multispectral camera, and a lidar to take real-time photos of the tailings pond, and transmit the collected RGB images, multispectral images, and three-dimensional point cloud data to the server; the RGB images and multispectral images correspond to the same area and have the same size; Step 2: After receiving the collected data, the server preprocesses the RGB image and the multispectral image to obtain an RGB orthoimage and a multispectral orthoimage; The size of RGB orthophotos and multispectral orthophotos are consistent; Step 3: Obtain a digital elevation model based on the three-dimensional point cloud data; Step 4: Obtain the starting boundary for calculating the dry beach length based on the digital elevation model; Step 5: Based on the RGB orthoimage, multispectral orthoimage and digital elevation model, multiple features are extracted and fused, and then the random forest method is used to classify and predict the fusion results to obtain the termination boundary of the dry beach length calculation; Step 6, obtaining the dry beach length according to the starting boundary and the ending boundary calculated by the dry beach length; Step 7: Determine whether the length of the dry beach exceeds a preset range, and issue an early warning if so.

2. The tailings dry beach automatic detection and early warning method of multi-sensor image joint processing as claimed in claim 1 is characterized by: In step 4, based on the digital elevation model, the elevation of the sedimentary beach surface is used as the reference elevation, and the three-dimensional analysis algorithm is used to extract the contour line contour of the outer slope of the dam body in the digital elevation model corresponding to the reference elevation, and the contour line contour is used as the starting boundary for calculating the dry beach length.

3. The tailings dry beach automatic detection and early warning method of multi-sensor image joint processing as claimed in claim 1 is characterized in that: The specific process of step 5 is: Step 5-1, extracting depth features and texture features from the RGB orthoimage, extracting index features from the multispectral orthoimage, and extracting elevation features about the termination boundary from the digital elevation model; Step 5-2, the depth feature, texture feature, index feature and elevation feature are fused, and each pixel in the fused feature map corresponds to a one-dimensional feature vector; Step 5-3: Input the one-dimensional feature vector corresponding to each pixel into the constructed random forest to obtain the prediction result corresponding to each pixel; Step 5-4: Based on the prediction results of each pixel obtained by random forest, the termination boundary of the dry beach length calculation is determined.

4. The tailings dry beach automatic detection and early warning method of multi-sensor image joint processing as claimed in claim 3 is characterized in that: The deep features are extracted as follows: Step A-1, inputting the RGB orthophoto image into the image segmentation model to obtain the segmentation result of water body and non-water body, wherein the segmentation result is a probability map, wherein the probability value indicates the credibility that the corresponding pixel belongs to the water body; in the processing process of the image segmentation model, extracting the intermediate feature map generated by the intermediate layer of the model; Step A-2: Fuse the probability map and the intermediate feature map by channel to obtain The depth features of the channels.

5. The tailings dry beach automatic detection and early warning method of multi-sensor image joint processing as claimed in claim 4 is characterized by: The image segmentation model in step A-1 is a U-Net Transformer model; The RGB orthophoto image is input into the U-Net Transformer model to obtain the probability map, and the decoder of the U-Net Transformer model is used to obtain the probability map. The intermediate feature map of each channel.

6. The tailings dry beach automatic detection and early warning method of multi-sensor image joint processing as claimed in claim 5 is characterized in that: The channel-by-channel fusion in step A-2 refers to: splicing the probability map and the intermediate feature map by channel to obtain the depth feature; Alternatively, the eigenvalues ​​of the intermediate feature map are weighted according to the probability map and then concatenated with the probability map by channel.

7. The tailings dry beach automatic detection and early warning method of multi-sensor image joint processing as claimed in claim 6 is characterized in that: The process of weighting the eigenvalues ​​of the intermediate feature map according to the probability map is as follows: pixels, let the corresponding probability value in the probability map be , the eigenvalues ​​corresponding to the pixel in the intermediate feature map are , , then respectively according to calculate Weighted eigenvalues To replace the corresponding Obtaining a weighted intermediate feature map, and then splicing the weighted intermediate feature map with the probability map by channel to obtain the depth feature; Weighted Eigenvalue The calculation method is: ; in, According to the probability value The calculated weighted values ​​corresponding to the eigenvalues ​​of each channel are: ； In the formula, the peak point .

8. The tailings dry beach automatic detection and early warning method according to any one of claims 5 to 7, characterized in that: In step 5-2, the depth features, texture features, index features and elevation features are spliced ​​by channel. The index features include NDVI index value and NDWI index value. Each pixel in the spliced ​​feature map corresponds to a one-dimensional feature vector, which contains the values ​​from the depth features, texture features, index features and elevation features. feature values.

9. The tailings dry beach automatic detection and early warning method according to any one of claims 3 to 7 using multi-sensor image joint processing is characterized in that: The random forest used in step 5-3 is constructed as follows: First, we construct each decision tree in the random forest by randomly selecting features, and then calculate the importance of each feature: ,in Refers to Features, Features The importance of is the number of decision trees constructed, Features In the The gain value in each decision tree is calculated; then, the features are sorted by importance, and each decision tree in the random forest is rebuilt. When the nodes of the decision tree are split again, the features with the highest importance in the currently selected feature subset are preferentially selected as the splitting features; The reconstructed random forest is used as the random forest for prediction.

10. The tailings dry beach automatic detection and early warning method according to any one of claims 3 to 7, characterized in that: When using random forest to predict a pixel, first input the corresponding one-dimensional feature vector into each decision tree to obtain the prediction distribution of each decision tree. The prediction distribution obtained by the decision tree is ,in Indicates The decision tree predicts that the current input one-dimensional feature vector belongs to The probability of the categories, is the number of categories, including several categories under the water body category and several categories under the non-water body category; then calculate the weight of each decision tree for different categories: ； in, is the number of decision trees in the random forest, Indicates The decision tree The weight of each category; Then, calculate the score of each category for the current input one-dimensional feature vector: ； in, For the The category with the highest score is the prediction result of the current pixel.

Images