Landslide risk point classification method based on multimodal decision fusion

Abstract

The present invention discloses a landslide risk point classification method based on multimodal decision fusion, which mainly solves the problem that the existing method cannot obtain the information contained in the digital elevation data from a small number of samples, thereby causing the unstable improvement of the landslide risk point classification accuracy. The implementation scheme is: extracting the digital elevation model corresponding to the high-resolution remote sensing image to be detected; dividing the extracted digital elevation model according to "flat land" and "mountainous land"; training the high-resolution remote sensing image according to whether it contains landslide risk points; training the digital elevation model according to whether it is "mountainous land" or "flat land"; using the high-resolution remote sensing image classification model to distinguish whether the sample to be detected is a landslide risk point; using the terrain classification result of the digital elevation model classification model to screen the high-resolution remote sensing image classification result to obtain the final classification result. The present invention improves the classification accuracy of landslide risk point detection and can be used for geological disaster detection and early warning.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing image interpretation, and in particular relates to a landslide risk point classification method, which can be used for detecting landslide risk points and making geological disaster warnings. Background Art

[0002] Geological disasters refer to natural disasters that cause great changes in geological morphology due to human and natural factors, thus causing great impacts on nearby areas. Being able to predict and notify the occurrence of natural disasters in advance plays a vital role in responding to natural disasters. Common geological disasters include crustal activity disasters, slope rock and soil movement disasters, ground deformation disasters, etc., and the most frequent of these disasters is slope rock and soil movement disasters. The most common landslides, collapses, and mudslides belong to this category of disasters. The characteristics of landslides in satellite images after landslides occur are more obvious, which has a clear guiding significance for the hazard level assessment of disasters, personnel evacuation, and post-disaster rescue. Although after a landslide disaster occurs, the detection of landslides by technical means can have a certain early warning effect on secondary disasters caused by landslides, but it cannot effectively predict the occurrence of landslides before they occur. If the occurrence of landslide disasters can be predicted and the threatening landslide risk points can be detected in advance, the loss of personnel and economy will be greatly reduced. Therefore, the detection of landslide risk points is more valuable in practical terms than the detection of landslides.

[0003] The problem of landslide risk point detection belongs to a complex scenario problem. First, the landslide risk point has lower significance than the landslide feature, that is, the signal-to-noise ratio is lower, and the background has a greater impact on the target. Most algorithms for classification or detection are based on whether the samples have the same features, which undoubtedly greatly increases the difficulty of the problem for geological disaster risk points that are slightly different from the background. Second, different landslides have different shapes, resulting in a large amount of data required for learning, which is undoubtedly a more serious problem for landslide risk points with limited high-quality labeled data.

[0004] In response to the problem of limited data information due to low signal-to-noise ratio, some geological disaster researchers have added digital elevation models (DEMs) to landslide detection, hoping to reflect the topographic changes of images through DEMs and provide more data features for solving related problems. For example, Ji Shunping et al. improved the classification performance of the landslide classification model by adding DEM in their study on landslide detection, "Landslide detection from an open satellite imagery and digital elevation model dataset using attentionboosted convolutional neural networks"; Liu Jia et al. improved the accuracy of rapid and automatic landslide identification by adding DEM, slope and other information in "Co-seismic landslide identification method based on GEE and U-net model".

[0005] The DEM processing methods in the above literature all directly add extra dimensions to store, process and extract DEM data features. This form of data processing can classify landslide risk points by extracting features from DEM models of different landslide risks when the amount of data is large, so as to improve the detection accuracy of landslide risk points. However, due to the diversity of terrain, the forms of landslide risk points are also diverse, and there are few high-quality annotated data of landslide risk points. Therefore, after the DEM features obtained by training are migrated to the test set, they cannot cover all types of landslide risk points, making it difficult to obtain the relationship information between slope, slope direction and landslide hidden in DEM in a small number of sample learning, resulting in unstable performance improvement of classification. Summary of the invention

[0006] The purpose of the present invention is to overcome the shortcomings of the existing technology and propose a landslide risk point classification method based on multimodal decision fusion to make full use of part of the information in DEM, improve the accuracy of landslide risk point classification while ensuring the stability of accuracy improvement.

[0007] To achieve the above object, the technical solution adopted by the present invention includes the following steps:

[0008] (1) Extract the height data corresponding to all RGB high-resolution remote sensing images from the digital elevation model to construct a regional digital elevation model (DEM) that corresponds to the RGB high-resolution remote sensing images one by one;

[0009] (2) The regional digital elevation models (DEMs) are merged to form a dataset HS, and the HS is divided into “flat land” and “mountainous land”;

[0010] (3) All RGB high-resolution remote sensing image samples are randomly divided into RGB training set and RGB test set in a ratio of 8:2;

[0011] (4) according to the one-to-one correspondence between the RGB training set and the RGB test set and the regional digital elevation model, the regional digital elevation model DEM extracted and obtained in step (1) is divided into a DEM training set and a DEM test set;

[0012] (5) Use the RGB high-resolution remote sensing image training set data and the regional digital elevation model DEM training set data to train the corresponding classifiers respectively:

[0013] (5a) Load the RGB high-resolution remote sensing image data in the RGB high-resolution remote sensing image training set into the existing classification model and train it using the stochastic gradient descent optimization method to obtain the RGB high-resolution remote sensing image classifier f RGBS ;

[0014] (5b) Load the regional digital elevation model DEM in the regional digital elevation model DEM training set into another existing classification model and train it using the stochastic gradient descent optimization method to obtain the regional digital elevation model DEM classifier f GS ;

[0015] (6) Use the classifier obtained in step (5) to perform detection on the RGB high-resolution remote sensing image test set and the regional digital elevation model DEM test set:

[0016] (6a) Using the RGB high-resolution remote sensing image classifier f obtained in (5a) RGBS Check each sample in the RGB high-resolution remote sensing image test set one by one to see if it contains landslide risk points;

[0017] (6b) Using the regional digital elevation model DEM classifier f obtained in (5b) HS Classify each sample in the regional digital elevation model DEM test set as "flat land" or "mountainous land" one by one;

[0018] (7) For the samples of “landslide risk points” detected in (6a), the classification results of their corresponding regional digital elevation model (DEM) are determined one by one, namely, “mountainous area” or “flat land”, and the samples whose corresponding regional digital elevation model (DEM) classification results are “flat land” are adjusted to “no landslide risk points”.

[0019] Compared with the existing method, the present invention has the following advantages:

[0020] First, the present invention improves the use of digital elevation model (DEM), extracts the terrain information contained in the regional digital elevation model (DEM), uses the regional digital elevation model (DEM) classifier to accurately classify the simple question of "mountain" or "flat", adjusts the corresponding RGB high-resolution remote sensing image classification results, and improves the classification performance of RGB high-resolution remote sensing images with or without landslide risk points;

[0021] Second, the present invention adds an extra branch for processing the digital elevation model DEM. Since it has a very high accuracy in classifying "flat land" and "mountainous land" and is not coupled with the classification branch of the original RGB high-resolution remote sensing image, compared with the application form of the existing digital elevation model DEM processing method in classification, the present invention can improve the accuracy and stability of classification after adding the digital elevation model DEM information. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a flow chart for realizing the present invention;

[0023] Figure 2 Schematic diagram of pixel distribution of the slope map in the present invention. DETAILED DESCRIPTION

[0024] The specific embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings.

[0025] Reference Figure 1 ,This example of landslide risk point classification method based on multi-modal decision fusion includes the following steps:

[0026] Step 1: extract the height data corresponding to the RGB high-resolution remote sensing image from the digital elevation model to construct a regional digital elevation model DEM corresponding to the RGB high-resolution remote sensing image.

[0027] RGB high-resolution remote sensing images are raw data, and their corresponding digital elevation model data play an auxiliary role in the detection of landslide risk points. In order to make better use of digital elevation model data, this example clips the digital elevation model according to the correspondence between geographic coordinate systems. At the same time, since the altitude data represented by DEM varies greatly between different regions even if the terrain is the same, this has no effect on whether a landslide will occur. Therefore, in order to eliminate the altitude differences between regions, for each DEM point data on the DEM corresponding to each image, the minimum value of each DEM data needs to be subtracted. The specific implementation is as follows:

[0028] 1.1) Obtain a specific RGB high-resolution remote sensing image from the original RGB high-resolution remote sensing image dataset, read the geographic coordinate system information contained therein and the size width w and height h of the RGB high-resolution remote sensing image;

[0029] 1.2) Map the geographic coordinate information obtained in 1.1) to the digital elevation model to obtain the coordinate information (x, y) of the upper left corner of the position to be cropped;

[0030] 1.3) According to the size information obtained in 1.1) and the coordinate information obtained in 1.2), the area with (x, y) as the upper left corner and (x+w, y+h) as the lower right corner in the digital elevation model is cropped as the regional digital elevation model DEM corresponding to the RGB high-resolution remote sensing image in 1.1);

[0031] 1.4) Calculate the mapping height h′(i,j) of each pixel in the regional digital elevation model DEM image:

[0032] h′(i,j)=h(i,j)-min(h(i,j))

[0033] Where h(i,j) is the value corresponding to the coordinate position (i,j) in the digital elevation model, and min is the function that takes the global minimum value.

[0034] Step 2: construct a dataset HS based on the regional digital elevation model DEM, and divide the HS into "flat land" and "mountainous land".

[0035] In order to improve the accuracy of the classification of the original data by the overall classification model after adding the regional digital elevation model DEM, the present invention adds some prior knowledge to the existing classification model to improve the performance of the classification model, and provides additional classification basis for the classification of landslide risk points based on the prior information that landslides cannot occur in areas with flat terrain and landslides may occur in areas with rugged terrain, that is, "flat land" does not include "landslide risk points". Therefore, the regional digital elevation model DEM in the data set HS is divided into a "flat land" part and a "mountainous land" part, so as to realize the subsequent classification of the aforementioned terrain information.

[0036] Reference Figure 2 , the specific implementation of this step is as follows:

[0037] 2.1) Calculate the east-west change rate of each pixel in each image in the dataset HS and the north-south rate of change

[0038]

[0039] Among them, {a,b,c,d,e,f,g,h,i} are all the numbers of the nine-grid pixels centered on a certain pixel in the height map image in the dataset HS. The nine pixels are distributed as a whole into three layers, with three pixels in each layer. The first layer is {a,b,c}, the second layer is {d,e,f}, and the third layer is {g,h,i}. we_cellsize represents the pixel size in the east-west direction, and sn_cellsize represents the pixel size in the north-south direction.

[0040] 2.2) According to the results of 2.1), calculate the slope value S corresponding to each pixel of the digital elevation model DEM image of each area in the data set HS (i,j) :

[0041]

[0042] Among them, the constant C is the conversion ratio between angle and radian, and its value is 57.29578;

[0043] 2.3) merging the slope values ​​belonging to the same regional digital elevation model DEM image into a slope map;

[0044] 2.4) Calculate the maximum slope value S in the slope map corresponding to the digital elevation model DEM image of each area in the dataset HS max ;

[0045] 2.5) Set the threshold t to 5° and take the maximum slope value S in the slope map corresponding to the DEM image of each region as max Compare with this threshold:

[0046] If S max >t, the image is labeled as “mountainous”;

[0047] If S max ≤t, the image is labeled as “flat land”.

[0048] Step 3: Divide the RGB training set and the RGB test set.

[0049] The original RGB high-resolution remote sensing image samples to be classified for landslide risk points are randomly divided into RGB training set and RGB test set, respectively denoted as RGBS train and RGBS test .

[0050] Step 4: For the regional digital elevation model DEM with terrain annotation information obtained in step 2, it is divided into a DEM training set and a DEM test set according to the one-to-one correspondence between the RGB training set and the RGB test set and the regional digital elevation model.

[0051] Since the classification task of the branch corresponding to the RGB high-resolution remote sensing image is "whether there is a landslide risk point", but the classification task of the regional digital elevation model DEM branch is "flat land" or "mountainous land", if the regional digital elevation model DEM data set is divided in the form of random division, some RGB high-resolution remote sensing image samples will be divided into the RGB training set, and the regional digital elevation model DEM corresponding to these samples will be divided into the DEM test set, which will result in inaccurate calculation of the final indicator.

[0052] In order to ensure that the training and testing of the model and the calculation of indicators can meet the objective situation, this example uses the one-to-one correspondence when cutting the regional digital elevation model DEM in step 1 and the division of the RGB training set and the RGB test set in step 3 to divide the regional digital elevation model DEM dataset into DEM training set HS train and DEM test set HS test .

[0053] Step 5: Use the RGB high-resolution remote sensing image training set data and the regional digital elevation model DEM training set data to train the corresponding classifiers respectively.

[0054] 5.1) Use the stochastic gradient descent optimization method to train the RGB high-resolution remote sensing image classification model to obtain the classifier f RGBS :

[0055] For the RGB high-resolution remote sensing image training set RGBS obtained in step 3 train , select an existing suitable deep learning classification model for relevant classification experiments. Due to the limited amount of data, transfer learning is used here to train the model. Use the pre-training files of the public network model on the ImageNet dataset for training. In this way, relatively good training results can be achieved even with a small amount of data. For the classification task of the two categories of "points with landslide risk" and "points without landslide risk" in this example, change the category output layer to two categories, and then train the classifier. The implementation is as follows:

[0056] 5.1.1) Generate an RGB high-resolution remote sensing image classifier f with random parameters according to the network structure corresponding to the trained model RGBS ;

[0057] 5.1.2) Use the network parameters in the public pre-training file to replace the RGB high-resolution remote sensing image classifier f RGBs Random parameters of

[0058] 5.1.3) RGB high-resolution remote sensing image data x RGB Load into RGB high-resolution remote sensing image classifier f RGBSIn the above example, we get the high-resolution image data x RGB The preliminary classification result y′ RGB ;

[0059] 5.1.4) RGB high-resolution remote sensing image classifier f RGBS The classification result y′ RGB The corresponding classification label y RgB Input to the cross entropy loss function f(y′ RGB ,y RGB ), and obtain the RGB high-resolution remote sensing image classifier f RGBS Loss value l RGB ;

[0060] 5.1.5) Calculate the classifier f RGBS The loss value l RGB For the classifier f RGBS The gradient of the parameters and updates the classifier f in the negative direction of the gradient RGBS parameter;

[0061] 5.1.6) Repeat (5.1.3) to (5.1.5) until the RGB high-resolution remote sensing image classifier f RGBS The loss value l RGB Converge and obtain the RGB high-resolution remote sensing image classifier f RGBS ;

[0062] 5.2) Using the stochastic gradient descent optimization method to train the regional digital elevation model (DEM) data, the classifier f is obtained. HS :

[0063] This example uses the Vision In Transformer (VIT) model to classify DEM branch data. Compared with other convolutional neural network classification models, VIT has stronger interpretability. The unique location encoding feature of VIT makes the model pay more attention to areas with higher slope values, rather than just calculating the classification surfaces between different categories.

[0064] Due to the limited amount of data, this example uses transfer learning to train the DEM branch classification model. For the two-category classification task of "flat land" and "mountain" in this example, the category output layer is changed to two categories, and then the classification model is trained using the pre-trained file corresponding to the VIT model on the ImageNet 21K dataset. The specific steps are as follows:

[0065] 5.2.1) Generate a classifier f with random parameters according to the network structure corresponding to the regional digital elevation model DEM classification model HS ;

[0066] 5.2.2) Replace the classifier f with the network parameters obtained from the public pre-training file HS Random parameters of

[0067] 5.2.3) Regional digital elevation model DEM data x H Load into the regional digital elevation model DEM classifier f HS In the above example, we can get the digital elevation model DEM data x H The preliminary classification result y′ H ;

[0068] 5.2.4) Regional digital elevation model DEM classifier f HS The classification result y′ H The corresponding classification label y H Input to the loss cross entropy function f(y′ H ,y H ), and obtain the regional digital elevation model DEM classifier f HS Loss value l H ;

[0069] 5.2.5) Calculate the regional digital elevation model DEM classifier f HS The loss value l H For the regional digital elevation model DEM classifier f HS The gradient of the parameters in , and update the classifier f in the negative direction of the gradient HS parameter.

[0070] 5.2.6) Repeat (5.2.3) to (5.2.5) until the regional digital elevation model DEM classifier f HS The loss value l H Converge and get the regional digital elevation model DEM classifier f HS .

[0071] Step 6: Use RGB high-resolution remote sensing image classifier f RGBS and regional digital elevation model DEM classifier f HS The detection was carried out on the RGB high-resolution remote sensing image test set and the regional digital elevation model DEM test set respectively.

[0072] 6.1) Using the RGB high-resolution remote sensing image classifier f obtained in 5.1) RGBS Check each sample in the RGB high-resolution remote sensing image test set one by one to see if it contains landslide risk points;

[0073] For the RGB high-resolution remote sensing image test set RGBS obtained in step 3 test This example uses the RGB high-resolution remote sensing image classifier f obtained in step 5.1) RGBSEach sample to be classified is classified to see whether it contains a "land slide risk point". The sample with the classification result of "land slide risk point" is recorded as true and classified into the "land slide risk point" sample set TR. t , the samples classified as “no landslide risk point” are recorded as false and classified into the “no landslide risk point” sample set TR f .

[0074] 6.2) Using the regional digital elevation model DEM classifier f obtained in 5.2) HS Classify each sample in the regional digital elevation model DEM test set as "flat land" or "mountainous land" one by one;

[0075] For the regional digital elevation model DEM test set HS obtained in step 4 test In this example, the regional digital elevation model DEM classifier f obtained in step 6b) is used. HS Classify each sample to be classified as "flat land" or "mountainous land", and record the image classified as "flat land" as plate and divide it into TH p , the image classified as "mountain" is recorded as mountain and divided into TH m .

[0076] Step 7: Adjust the classification result set TR of the RGB branch in 6.1) according to the classification result of the DEM branch in 6.2). t and TR f . And output the final classification results.

[0077] For the two classification sets corresponding to the classification results in 6.1) the sample set TR of “points with landslide risk” t and the sample set TR of “no landslide risk point” f , determine one by one whether the classification result of the corresponding regional digital elevation model DEM is "mountain" or "flat", and adjust some of the classification results. The specific implementation steps are as follows:

[0078] 7.1) Go through the sample set TR of “points with landslide risk” one by one t The various samples:

[0079] If the regional digital elevation model DEM corresponding to each sample is passed through the classifier f HS If the classification result is “flat land”, the sample result is adjusted from “with landslide risk point” to “without landslide risk point” and the result set TR t Adjust to result set TR f ;

[0080] If the regional digital elevation model DEM corresponding to each sample is passed through the classifier f HSIf the classification result is "mountainous", no processing is done;

[0081] 7.2) For the sample set TR of “no landslide risk point” f The samples in are not processed;

[0082] 7.3) Output RGB high-resolution remote sensing image test set RGBS test The final classification result set TR t and TR f .

[0083] The following is a description of the technical effects of the present invention in combination with simulation experiments:

[0084] 1. Simulation conditions

[0085] Intel Core i7-7800X CPU with a main frequency of 3.50GHz, 64.0GB of memory, Ubuntu18.04 operating system, python3.7.1+pytorch1.7.0 development environment.

[0086] The experimental data is a high-resolution image with a resolution of 1m, and the digital elevation model DEM data is a 30m resolution digital elevation model data of the corresponding area.

[0087] 2. Simulation content

[0088] Simulation 1: The above data are classified and simulated on Resnet34, Resnet50, Resnet101, Res_att, Dense-Net, Efficient-Net, EfficientNet-v2 and VIT models using the present invention and the existing mainstream RGB high-resolution remote sensing image classification method respectively, and the simulation results are compared using four indicators: positive sample precision rate Pre_pos, positive sample recall rate Rec_pos, positive sample F1-score and overall sample accuracy rate ACC. The results are shown in Table 1.

[0089] Table 1 Classification results of “points with landslide risk” category

[0090]

[0091] As can be seen from Table 1, the experimental results of the nine mainstream classification models used in the experiment show that the present invention can screen out the "flat ground" samples that are misclassified as "no landslide wind point" while ensuring that the recall rate of the "landslide risk point" category is not reduced. And in this way, the detection accuracy of the "landslide risk point" samples is improved. Compared with the existing methods, in most cases, the classification indicators of the "landslide risk point" category can be improved by 1 to 3 percentage points.

[0092] Simulation 2: The processing methods of the present invention and the existing mainstream RGB high-resolution remote sensing images and regional digital elevation model DEM are used to perform classification simulation on the above data on Resnet34, Resnet50, Resnet101, Res_att, Dense-Net, Efficient-Net, EfficientNet-v2 and VIT models, and four indicators including positive sample precision rate Pre_pos, positive sample recall rate Rec_pos, positive sample F1-score and overall sample accuracy rate ACC are used for comparison. The results are shown in Table 2.

[0093] Table 2 Classification results of “points with landslide risk” category

[0094]

[0095] As can be seen from Table 2, due to the instability of DEM information and the additional noise caused by low-resolution DEM interpolation, the classification separability of the dataset with additional information will increase or decrease and become unstable. Although additional input is added, the actual classification performance is worse in most indicators, and even worse than the results before adding DEM information. However, the processing method of the present invention can still improve by 1 to 5 percentage points compared with other methods.

Claims

1. A landslide risk point classification method based on multimodal decision fusion, characterized in that: The steps include: (1) Extract the height data corresponding to all RGB high-resolution remote sensing images from the digital elevation model to construct a regional digital elevation model (DEM) that corresponds to the RGB high-resolution remote sensing images one by one; (2) The regional digital elevation models (DEMs) are merged to form a dataset HS, and the HS is divided into "flat land" and "mountainous land"; (3) All RGB high-resolution remote sensing image samples are randomly divided into RGB training set and RGB test set; (4) according to the one-to-one correspondence between the RGB training set and the RGB test set and the regional digital elevation model, the regional digital elevation model DEM extracted and obtained in step (1) is divided into a DEM training set and a DEM test set; (5) Use the RGB high-resolution remote sensing image training set data and the regional digital elevation model DEM training set data to train the corresponding classifiers respectively: (5a) Load the RGB high-resolution remote sensing image data in the RGB high-resolution remote sensing image training set into the existing classification model and train it using the stochastic gradient descent optimization method to obtain the RGB high-resolution remote sensing image classifier f RGBS ; (5b) Load the regional digital elevation model DEM in the regional digital elevation model DEM training set into another existing classification model and train it using the stochastic gradient descent optimization method to obtain the regional digital elevation model DEM classifier f HS ; (6) Use the classifier obtained in step (5) to perform detection on the RGB high-resolution remote sensing image test set and the regional digital elevation model DEM test set: (6a) Using the RGB high-resolution remote sensing image classifier f obtained in (5a) RGBS Check each sample in the RGB high-resolution remote sensing image test set one by one to see if it contains landslide risk points; (6b) Using the regional digital elevation model DEM classifier f obtained in (5b) HS Classify each sample in the regional digital elevation model DEM test set as "flat land" or "mountainous land" one by one; (7) For the samples of "with landslide risk points" detected in (6a), the classification results of the corresponding regional digital elevation model (DEM) are determined one by one, namely "mountainous area" or "flat land", and the samples whose corresponding regional digital elevation model (DEM) classification results are "flat land" are adjusted to "without landslide risk points".

2. The method according to claim 1, characterized in that: In (1), the height data corresponding to all RGB high-resolution remote sensing images are extracted from the digital elevation model, which is implemented as follows: (1a) The digital elevation model information corresponding to the landslide risk is cut according to the geographic information and stored in a single-channel image corresponding to the size of the corresponding landslide risk point; (1b) Calculate the image mapping height h′(i,j) for each pixel: h′(i,j)=h(i,j)-min(h(i,j)) Where h(i,j) is the value corresponding to the coordinate position (i,j) in the digital elevation model, and min is the function that takes the global minimum value.

3. The method according to claim 1, characterized in that: In (2), the data set HS is divided into "flat land" and "mountainous land", which is implemented as follows: (2a) Calculate the east-west change rate of each pixel in each image in the dataset HS and the north-south rate of change Among them, {a,b,c,d,e,f,g,h,i} are all the numbers of the nine-grid pixels centered on a certain pixel in the height map image in the dataset HS. The nine pixels are distributed as a whole into three layers, with three pixels in each layer. The first layer is {a,b,c}, the second layer is {d,e,f}, and the third layer is {g,h,i}. We_cellsize is the pixel size in the east-west direction, and sn_cellsize is the pixel size in the north-south direction. (2b) Calculate the slope value S corresponding to each pixel of each image in the data set HS (i,j) : The constant C is the conversion factor between angle and radian, and its value is 57.29578; (2c) merging the slope values ​​belonging to the same image into a slope map; (2d) Calculate the maximum slope value S in the slope map corresponding to each image in the dataset HS max ; (2e) Set the threshold t to 5° and calculate the maximum slope value S in the slope map corresponding to each image. max Compare with this threshold: If S max >t, the image is classified as "mountainous"; If S max ≤t, the image is classified as "flat land".

4. The method according to claim 1, characterized in that: In (5a), the RGB high-resolution remote sensing image classification model is trained using the stochastic gradient descent optimization method, which is implemented as follows: (5a1) Generate a classifier f with random parameters according to the network structure corresponding to the trained model RGBS ; (5a2) The RGB high-resolution image data x RGB Load into classifier f RGBS In the above example, we get the high-resolution image data x RGB The preliminary classification result y′ RGB ; (5a3) The classifier f RGBS The classification result y′ RGB The corresponding classification label y RGB Input to the cross entropy loss function f(y′ RGB ,y RGB ), and get the loss value l RGB ; (5a4) Calculate the loss value l RGB For the classifier f RGBS The gradient of the parameters and updates the classifier f in the negative direction of the gradient RGBS parameter; (5a5) Repeat (5a2) to (5a4) until the final loss value is l RGB Converge and obtain the RGB high-resolution remote sensing image classifier f RGBS .

5. The method according to claim 1, characterized in that: In (5b), the regional digital elevation model DEM classification model is trained using the stochastic gradient descent optimization method, which is implemented as follows: (5b1) Generate a classifier f with random parameters according to the network structure corresponding to the regional digital elevation model DEM classification model Hs ; (5b2) Regional digital elevation model DEM data x H Load into classifier f HS In the above example, we can get the digital elevation model DEM data x H The preliminary classification result y′ H ; (5b3) The classifier f HS The classification result y′ H The corresponding classification label y H Input to the loss cross entropy function f(y′ H ,y H ), and get the loss value l H ; (5b4) Calculate the loss value l H For the classifier f HS The gradient of the parameters in , and update the classifier f in the negative direction of the gradient HS parameter; (5b5) Repeat (5b2) to (5b4) until the final loss value is l H Convergence, get the regional digital elevation model DEM classifier f HS .

Images