A semi-supervised deep learning method for intelligent remote sensing identification of large landslides in mountainous areas

Abstract

The application discloses a kind of semi-supervised deep learning mountain large-scale landslide intelligent remote sensing identification method, belong to surveying and mapping technical field.The application includes the following steps:S1: selecting containing landslide remote sensing image and corresponding area's digital elevation data are carried out data synthesis, obtain data set, label annotation is carried out to data and obtains sample set;S2: constructing student training network model and teacher training network model, training and judging to sample set, prediction is carried out to sample set and obtains prediction result, the landslide feature type of judging prediction result and obtains training result;S3: the training result obtained in S2 is carried out cross entropy loss value calculation, calculate mean square error, obtain exponential moving average line, according to exponential moving average line updates weight in student training network model, and continuously predicted weighted average is used to update the weight of teacher training network model.The application is efficiently and accurately identified landslide in remote sensing image under the support of small amount of sample and limited computing power.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of surveying and mapping, in particular to a semi-supervised deep learning intelligent remote sensing identification method for large landslides in mountainous areas. BACKGROUND

[0002] China has vast territory, diverse landforms, and complex and changeable geographical and geological environment. Overall, the southwest region has a large terrain, with more mountainous terrain, and the mountainous area in the plateau and plain transition area is even more so. Due to the proximity to the subtropics, the annual precipitation is relatively high, and the precipitation time is also relatively long, so geological disasters occur frequently, and the disasters are extremely serious. One of the typical disasters that are more harmful and serious is landslides. In addition, landslides also damage road traffic and pose a serious threat to nearby critical facilities, such as important projects such as mountainous railways.

[0003] The frequent occurrence of geological disasters reminds us of the severity of disaster prevention and mitigation. In view of this situation, it is necessary to map the landslide area, distinguish the landslide boundary, and identify the distribution of the landslide when identifying the landslide, so as to use this information to provide a data basis for landslide investigation, secondary disaster warning, and risk assessment. The primary work of landslide identification is to select the identification method. Over the years, the method of landslide identification has undergone a long development. In the early days, landslide identification was mainly relied on manual work, and traditional methods were used for field geological mapping. The results obtained are accurate but time-consuming and labor-intensive. After the development of aerospace remote sensing technology, many methods have emerged for unmanned aerial vehicle or satellite remote sensing images. For image interpretation and analysis, from visual interpretation to computer interpretation, all landslide identification methods need to consider the accuracy of identification. At present, most methods have achieved a certain degree of accuracy, and how to improve the efficiency of rapid identification while ensuring accuracy for a large amount of data has become a research hotspot. The current landslide identification mainly uses machine learning algorithm classification, which is difficult to deal with complex landslides with high identification difficulty. SUMMARY

[0004] In view of the problems in the prior art, the present application provides a semi-supervised deep learning intelligent remote sensing identification method for large landslides in mountainous areas, which aims to efficiently and accurately identify landslides in remote sensing images under the support of a small amount of samples and limited computing power.

[0005] The technical scheme adopted by the present application is as follows:

[0006] A semi-supervised deep learning intelligent remote sensing identification method for large landslides in mountainous areas, comprising the following steps:

[0007] S1: select landslide-containing remote sensing images and corresponding regional digital elevation data for data synthesis to obtain a landslide remote sensing image data set containing digital elevation data, label part of the landslide remote sensing image data containing digital elevation data to obtain a sample set containing label annotation, and the rest of the landslide remote sensing image data containing digital elevation data form a sample set not containing label annotation;

[0008] S2: input the sample set containing label annotation and the sample set not containing label annotation obtained in S1 into a semantic segmentation network to respectively build a student training network model and a teacher training network model, train the sample set containing label annotation and the sample set not containing label annotation through the student training network model and the teacher training network model, and determine whether the training is completed, predict the sample set corresponding to the student training network model and the teacher training network model that have completed the training through the student training network model and the teacher training network model, and obtain prediction results, return the sample set that has not completed the training to the semantic segmentation network, and input the prediction results into a discriminator to respectively determine the landslide feature type of the prediction results corresponding to the student training network model and the teacher training network model and respectively obtain student training results and teacher training results;

[0009] S3: calculate the cross-entropy loss value of the student training results obtained in S2, and calculate the mean square error of the student training network model and the teacher training network model, calculate the exponential moving average line according to the cross-entropy loss value of the student training network model and the mean square error of the teacher training network model, update the weight in the student training network model according to the feedback adjustment of the exponential moving average line, and update the weight of the teacher training network model by continuously predicting and weighting the updated student training network model.

[0010] Preferably, in S1, landslide remote sensing images with a spatial resolution of 0.2m-0.9m and corresponding regional digital elevation data models are selected for data synthesis to obtain a landslide remote sensing image data set containing digital elevation data.

[0011] Preferably, the landslide remote sensing image data set containing digital elevation data in S1 is subjected to image enhancement, and the image enhancement methods include cropping transformation, rotation, flipping, scaling, and translation. The sample images in the landslide remote sensing image data set containing digital elevation data are cropped to make the cropped sample images consistent in size with the corresponding digital elevation data, and the sample images are uniformly scaled to squares. The sample images that are not squares are cropped to squares without damaging the integrity of the landslide.

[0012] Preferably, the sample normalization processing in S1 is:

[0013] The landslide remote sensing image data set containing digital elevation data obtained in S1 is subjected to image normalization processing, and the original image of the input layer is unified, and the characteristic value size range is reduced, and the formula is as follows:

[0014]

[0015] x * Corresponding to the pixel value of the pixel point after the image normalization operation, x is the pixel value of each point of the image to be processed, min(x) and max(x) are the minimum and maximum pixel values found in the image respectively.

[0016] Preferably, the process of labeling part of the landslide remote sensing image data containing digital elevation data in S1 is as follows: using the editor of Arcmap to draw a small amount of feature area vector map spot to label, generate a bounding box, and the element envelope rectangle in the Arcmap tool directly generates the smallest wrapping rectangle, and then output and save as a slice, wherein the ratio of the sample containing label annotation to the sample not containing label annotation is 1:8.

[0017] Preferably, the training process of constructing the student training network model and the teacher training network model in S2 is as follows:

[0018] The Mask RCNN algorithm based on the deep learning framework Tensorflow is used for training and testing, artificial annotation and part of the label annotation sample of the public data set are used for identification training, the Feature Map is obtained and output based on the backbone network ResNet, and then the large landslide features are sampled and merged through the feature pyramid FPN.

[0019] Further, the process of predicting the sample set corresponding to the student training network model and the teacher training network model by the student training network model and the teacher training network model in S2 is as follows:

[0020] The semantic segmentation network in S2 is a candidate region network, which generates an anchor box with each pixel as the center based on the large landslide features obtained in S2, wherein the position coordinates of the real box are x_a, y_a, the width and height of the real box are w_a, h_a, the position coordinates of the anchor box are x * ,y * , the width and height of the anchor box are w * ,h * , and the offset formula is as follows:

[0021]

[0022]

[0023] The anchor frame is moved on each image of the large landslide data set to help train the candidate region network, and then the anchor frame is subjected to a classification task and a regression task. The classification task calculates the score of each anchor frame to determine the probability of landslide. The regression task finds the position of the anchor frame, and the position of the anchor frame is represented by the upper left point and the lower right point. The anchor frames are sorted according to the scores, and the positive and negative examples are divided according to the probability.

[0024] The classification of the candidate region network is based on a Softmax function, and the Softmax function is as follows:

[0025]

[0026] where a j represents the probability of the j-th category forward calculation, S j is the probability calculated by the Softmax function.

[0027] The cross-entropy loss function used is as follows:

[0028]

[0029] where y j represents the true label, and S j represents the probability.

[0030] Preferably, the process of inputting the prediction result into the discriminator in S2 to judge the landslide feature type of the prediction result corresponding to the student training network model and the teacher training network model respectively is:

[0031] After processing the prediction result output by the semantic segmentation network in the student training network model, the corresponding large landslide features are extracted by ROI Align, and the ROI Align loss function is:

[0032]

[0033] The types of these large landslide features are judged in the pre-trained ResNet-50 discriminator, and the prediction frame before adjustment is also regressed.

[0034] Preferably, the process of calculating the mean square error of the student training network model and the teacher training network model in S3, calculating the exponential moving average line according to the cross-entropy loss value of the student training network model and the mean square error of the teacher training network model, and feeding back to adjust and update the weight in the student training network model is:

[0035] The loss value of the student training network model is calculated according to the loss function of Mask RCNN, the mask is generated by the full connection network in the mask branch, the information of each pixel point of the real mask and the predicted mask is compared, the binary classification cross entropy loss function is adopted, and the loss value of the student training network model is calculated as follows:

[0036]

[0037]

[0038] Where y i is the predicted probability, y1 i is the real probability, and the loss value is 0 when the values of the two are the same, reaching the ideal value. The greater the difference between the predicted and real probabilities, the greater the loss value, and the loss value is changed to realize back propagation according to the weight parameters;

[0039] The mean square error between the student training network model and the teacher training network model is calculated, that is, the expected distance between the prediction results of the student training network model and the teacher training network model, and the calculation formula is as follows:

[0040]

[0041] Where f(x, θ, η) represents the student training network model, f(x, θ', η') represents the teacher training network model, J(θ) represents the expected distance between the prediction results of the student training network model and the teacher training network model, represents the expected value between the student training network model and the teacher training network model, x represents the prediction result of the training, θ represents the weight of the student training network model, θ' represents the weight of the teacher training network model, η represents the noise of the student training network model, and η' represents the noise of the teacher training network model. The mean square error is calculated by inputting the weights (θ, θ') and the noise (η, η') and combining the prediction result of the training, the mean square error is taken as the consistency cost, and the accuracy and reliability of the training results of the student training network model and the teacher training network model are judged according to the mean square error;

[0042] The exponential moving average line is calculated according to the cross entropy loss value and the mean square error obtained by the student training network model, and the weight in the student training network model is adjusted according to the exponential moving average line. The formula for calculating the exponential moving average line is as follows:

[0043] avg = loss + εJ(θ) (10)

[0044] Where loss represents the loss value of the student training network model, and ε represents the weight number of the mean square error.

[0045] Preferably, the process of weighting the student training network model weight updated according to the adjustment in S3 to continuously predict the weighting average to update the teacher training network model weight is:

[0046] The following formula is used:

[0047] θ' t = alpha * theta t-1 + (1-alpha) * theta' t (11)

[0048] Wherein alpha represents a smoothing coefficient parameter, theta represents the weight of the student training network model, theta' represents the weight of the teacher training network model, t represents the number of training, the weight of the teacher training network model in the t-1 training is updated by the weight of the student training network model in the t training and the weight of the teacher training network model in the t training is updated by the weight of the student training network model in the t training t . t

[0049] In summary, due to the adoption of the above technical solutions, the beneficial effects of the present application are:

[0050] 1. The present application can improve the accuracy and reliability of remote sensing image recognition and extraction of landslide results when the similarity is high or the landslide feature is not obvious under the condition of containing less labeled sample; without additional auxiliary data, the landslide information can be effectively extracted through the input remote sensing image, the landslide image extraction process is based on landslide detection, and the similarity is high or the landslide feature is not obvious. The landslide image will not affect the recognition and extraction result. This method realizes pixel-level automatic landslide detection and similarity high landslide differentiation, and integrates a semi-supervised learning method, which trains the student training network model and the teacher training network model respectively for samples with and without label marking, only a small amount of labeled samples can obtain a model with good classification effect, and then the weight of the training network model can be continuously adjusted and iteratively optimized, and a large amount of manual label marking time is saved, and the training efficiency is improved;

[0051] 2. A landslide recognition scheme based on large-scale landslide region search is constructed, which combines the Mask RCNN algorithm framework with semi-supervised learning, obtains and outputs feature maps through the backbone network ResNet, and then samples and merges large-scale landslide features through the feature pyramid FPN to output, realizes the back propagation of weight parameters, and uses a candidate region network for auxiliary training, which greatly improves the training accuracy and reliability, and can better extract landslide features, effectively reduces the algorithm loss, greatly speeds up the extraction speed of the landslide, and improves the reliability of the landslide identification and extraction; ​

[0052] 3. A discriminator module is established based on ResNet-50, which takes the mean square error between the extracted landslide model output and the landslide sample model input as the training target, and through the discriminator module, the loss function of Mask RCNN and the mean square error between the two models, the quality of the output results of the student training network model and the teacher training network model can be effectively evaluated. According to the size of the loss value, the mean square error between the two training network models and the exponential moving average line, the model is automatically fine-tuned. BRIEF DESCRIPTION OF DRAWINGS

[0053] The present application will be described by way of example and with reference to the accompanying drawings, in which:

[0054] Figure 1 is a flowchart of the present application;

[0055] Figure 2 is a flowchart of the neural network of the present application;

[0056] Figure 3 is a structural diagram of the two training network models of the present application;

[0057] Figure 4 is a comparison diagram of the results of the model of the present application;

[0058] Figure 5 is a comparison diagram of the training convergence results of the model of the present application;

[0059] Figure 6 is a loss value curve diagram of the training results of the present application;

[0060] Figure 7 is a learning rate curve diagram of the training results of the present application. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0062] In the description of the embodiments of the present application, it should be noted that the terms "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship in which the product of the application is usually placed, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second", "third" and the like are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0063] The following will be described in conjunction with Figures 1-5 The present application will be described in detail.

[0064] Figure 1 The overall architecture flow design diagram of the present application is shown, which is divided into three parts S1, S2 and S3 according to steps, and is performed in sequence, each part has a separate output, and is the input of the next part;

[0065] Figure 2 The training network flowchart used by the model is shown, which can be divided into three parts of feature extraction, RPN network, classification and regression prediction,

[0066] Figure 3 The structure and relationship between the student training network model and the teacher training network model are shown, the student training network model trains the samples with label marks, and the teacher training network model trains the remaining samples without label marks, wherein the ratio of samples with label marks to samples without label marks is 1:8, and the weights of the student training network model can be updated to further update the weights of the teacher training network model;

[0067] Figure 4 The image comparison of landslide remote sensing image original image, digital elevation data, landslide area identified by artificial recognition, landslide segmented by training after semi-supervised learning and landslide segmented by training the image combined with digital elevation data is shown; A1, A2 and A3 are landslide remote sensing image original images, AA1, AA2 and AA3 are digital elevation data corresponding to the landslide image, B1, B2 and B3 are landslide areas identified by artificial recognition, C1, C2 and C3 are landslides segmented by training after semi-supervised learning, and D1, D2 and D3 are segmented landslide images combined with digital elevation data;

[0068] Figure 5The display is the prediction result obtained after the model individual discriminates Mask RCNN semantic segmentation and semi-supervised model fine-tuning, wherein the original Img is the original high-resolution remote sensing image, the Train Label is the label established by artificial, and the Pre is the prediction result of the model.

[0069] Figure 6 、 Figure 7 The display is the loss value and learning rate of the last landslide remote sensing image identification according to the loss function calculation of the student training network model.

[0070] Embodiments

[0071] A semi-supervised deep learning intelligent remote sensing identification method for large landslides in mountainous areas, comprising the following steps:

[0072] S1: Selecting remote sensing images containing landslides and corresponding digital elevation data for data synthesis to obtain landslide remote sensing image data sets containing digital elevation data, labeling part of the landslide remote sensing image data containing digital elevation data to obtain a sample set containing label annotations, and the remaining landslide remote sensing image data containing digital elevation data forming a sample set not containing label annotations;

[0073] S2: Inputting the sample set containing label annotations and the sample set not containing label annotations obtained in S1 into a semantic segmentation network to construct a student training network model and a teacher training network model, respectively, training the sample set containing label annotations and the sample set not containing label annotations by the student training network model and the teacher training network model, respectively, and determining whether the training is completed, predicting the sample set corresponding to the completed training by the student training network model and the teacher training network model, and obtaining prediction results, returning the sample set not completed training to the semantic segmentation network, and inputting the prediction results into a discriminator to determine the landslide feature type of the prediction results corresponding to the student training network model and the teacher training network model, respectively, and obtaining student training results and teacher training results, respectively;

[0074] S3: Calculating the cross-entropy loss value of the student training results obtained in S2, and calculating the mean square error of the student training network model and the teacher training network model, calculating the exponential moving average line according to the cross-entropy loss value of the student training network model and the mean square error of the teacher training network model, updating the weight in the student training network model according to the feedback adjustment of the exponential moving average line, and updating the weight of the teacher training network model by continuously predicting and weighting average according to the weight of the student training network model after the adjustment and update.

[0075] In S1, landslide remote sensing images with a spatial resolution of 0.2 m to 0.9 m are selected, and the corresponding regional digital elevation data model is used for data synthesis to obtain a landslide remote sensing image data set containing digital elevation data.

[0076] The landslide remote sensing image data set containing digital elevation data in S1 is subjected to image enhancement, and the image enhancement methods include cropping transformation, rotation, flipping, scaling, and translation. The sample images in the landslide remote sensing image data set containing digital elevation data are cropped to make the cropped sample images consistent in size with the corresponding digital elevation data, and the sample images are uniformly scaled to squares. The sample images that are not squares are cropped to squares without damaging the integrity of the landslide.

[0077] The sample normalization process in S1 is as follows:

[0078] The landslide remote sensing image data set containing digital elevation data obtained in S1 is subjected to image normalization processing to unify the original images of the input layer and reduce the range of feature values. The formula is as follows:

[0079]

[0080] x * Corresponding to the pixel value of the pixel point after the image normalization operation, x is the pixel value of each point of the image to be processed, and min(x) and max(x) are the minimum and maximum pixel values found in the image, respectively.

[0081] The process of labeling part of the landslide remote sensing image data containing digital elevation data in S1 is as follows: a small number of feature regions are vectorized and labeled using the editor of Arcmap, a bounding box is generated, and the smallest wrapping rectangle is directly generated using the feature envelope rectangle in the Arcmap tool. Then, it is output and saved as a slice, and the ratio of samples containing label annotations to samples not containing label annotations is 1:8. The training process of the student training network model and the teacher training network model in S2 is as follows:

[0082] The Mask RCNN algorithm based on the deep learning framework Tensorflow is used for training and testing, artificial labeling and part of the label annotation samples of the public data set are used for identification training, the Feature Map is obtained and output based on the backbone network ResNet, and then the large landslide features are sampled and merged by the feature pyramid FPN.

[0083] The process of predicting the sample set corresponding to the student training network model and the teacher training network model in S2 is as follows:

[0084] The semantic segmentation network in S2 is a candidate region network. The candidate region network generates an anchor box centered on each pixel of the large landslide features obtained in S2, wherein the position coordinates of the real box are x_a, y_a, the width and height of the real box are w_a, h_a, the position coordinates of the anchor box are x * ,y * , the width and height of the anchor box are w * ,h * , and the offset formula is:

[0085]

[0086]

[0087] The anchor box is moved on each image of the large landslide dataset to help the candidate region network training, and then the anchor box is subjected to a classification task and a regression task. The classification task calculates the score of each anchor box to determine the probability of landslide, and the regression task finds the position of the anchor box, which is represented by the upper left point and the lower right point. The anchor boxes are sorted according to the scores, and the positive and negative examples are divided according to the probability.

[0088] The classification of the candidate region network is based on the Softmax function, and the Softmax function is as follows:

[0089]

[0090] where a j represents the probability of j-class forward calculation, S j is the probability calculated by the Softmax function; the cross-entropy loss function used is as follows:

[0091]

[0092] where y j represents the real label, and S j represents the probability.

[0093] The process of inputting the prediction results in S2 into the discriminator to judge the landslide feature types corresponding to the prediction results of the student training network model and the teacher training network model is as follows:

[0094] After processing the prediction results output by the semantic segmentation network in the student training network model, the corresponding large landslide features are extracted by ROI Align, and the ROI Align loss function is as follows:

[0095]

[0096] The types of these large landslide features are judged in the pre-trained ResNet-50 discriminator, and the prediction box before adjustment is also regressed.

[0097] The mean square error of the student training network model and the teacher training network model is calculated in S3. The exponential moving average line is calculated according to the cross-entropy loss value of the student training network model and the mean square error of the teacher training network model. The process of adjusting and updating the weight in the student training network model according to the exponential moving average line is as follows:

[0098] The loss value of the student training network model is calculated according to the loss function of Mask RCNN. In the Mask branch, the mask is generated by the fully connected network. The information of each pixel point of the real mask and the predicted mask is compared, and the binary classification cross-entropy loss function is used. The loss value of the student training network model is calculated as follows:

[0099]

[0100]

[0101] where y i is the predicted probability, y1 i is the real probability, and the loss value is 0 when the values of the two are the same, reaching the ideal value. The greater the difference between the predicted and real probabilities, the greater the loss value. According to the loss value, the weight parameter is changed to realize backpropagation;

[0102] The mean square error between the student training network model and the teacher training network model is calculated, that is, the expected distance between the prediction results of the student training network model and the teacher training network model. The calculation formula is as follows:

[0103]

[0104] where f(x, θ, η) represents the student training network model, f(x, θ', η') represents the teacher training network model, J(θ) represents the expected distance between the prediction results of the student training network model and the teacher training network model, represents the expected value between the student training network model and the teacher training network model, x represents the prediction result of the training, θ represents the weight of the student training network model, θ' represents the weight of the teacher training network model, η represents the noise of the student training network model, and η' represents the noise of the teacher training network model. By inputting the weight (θ, θ') and the noise (η, η'), and combining the prediction result of the training, the mean square error is calculated as the consistency cost. According to the mean square error, the accuracy and reliability of the training results of the student training network model and the teacher training network model are judged;

[0105] According to the cross-entropy loss value and the mean square error of the student training network model, an exponential moving average line is calculated, and the weights in the student training network model are adjusted according to the exponential moving average line. The formula for calculating the exponential moving average line is as follows:

[0106] avg = loss + εJ(θ) (10)

[0107] Wherein, loss represents the loss value of the student training network model, and ε represents the weight number of the mean square error.

[0108] The process of continuously predicting and performing weighted average to update the weights of the teacher training network model according to the adjusted and updated weights of the student training network model in S3 is as follows:

[0109] The following formula is used:

[0110] θ' t = αθ' t-1 + (1-α)θ t (11)

[0111] Wherein, α represents a smoothing coefficient parameter, θ represents the weights of the student training network model, θ' represents the weights of the teacher training network model, and t represents the number of training. The weights θ t ' -1 of the teacher training network model in the t-1th training and the weights θ t of the student training network model in the tth training are used to update the weights θ t ' of the teacher training network model in the tth training.

[0112] The above embodiments only express the specific implementation of the present application, and the description is more specific and detailed, but it cannot be understood as the limitation of the protection scope of the present application. It should be pointed out that for ordinary skilled in the art, without departing from the technical scheme concept of the present application, a number of modifications and improvements can be made, which belong to the protection scope of the present application.

Claims

1. A semi-supervised deep learning-based intelligent remote sensing identification method for large landslides in mountainous areas, characterized in that, Includes the following steps: S1: Select remote sensing images containing landslides and digital elevation data of the corresponding areas to synthesize the data, and obtain a landslide remote sensing image dataset containing digital elevation data. Label some of the landslide remote sensing image data containing digital elevation data to obtain a sample set with labels. The remaining landslide remote sensing image data containing digital elevation data form a sample set without labels. S2: Input the labeled sample set and the unlabeled sample set obtained in S1 into the semantic segmentation network to construct the student training network model and the teacher training network model respectively. Train the labeled sample set and the unlabeled sample set with the student training network model and the teacher training network model respectively and determine whether the training is complete. Predict the corresponding completed sample set with the student training network model and the teacher training network model and obtain the prediction result. Return the untrained sample set to the semantic segmentation network. Input the prediction result into the discriminator to determine the landslide feature type of the prediction result corresponding to the student training network model and the teacher training network model respectively and obtain the student training result and the teacher training result respectively. S3: Calculate the cross-entropy loss value of the student training results obtained in S2, and simultaneously calculate the mean square error of the student training network model and the teacher training network model. Calculate the exponential moving average based on the cross-entropy loss value of the student training network model and the mean square error of the teacher training network model. Adjust and update the weights in the student training network model based on the feedback from the exponential moving average. Based on the adjusted and updated weights of the student training network model, continuously predict and perform a weighted average to update the weights of the teacher training network model.

2. The semi-supervised deep learning-based intelligent remote sensing identification method for large landslides in mountainous areas according to claim 1, characterized in that, In step S1, landslide remote sensing images with a spatial resolution of 0.2m to 0.9m are selected and combined with digital elevation data models of the corresponding areas to obtain a landslide remote sensing image dataset containing digital elevation data.

3. The semi-supervised deep learning-based intelligent remote sensing identification method for large landslides in mountainous areas according to claim 1, characterized in that, Image enhancement is performed on the landslide remote sensing image dataset containing digital elevation data in S1. The image enhancement methods include: cropping transformation, rotation, flipping, scaling, and translation. The sample images in the landslide remote sensing image dataset containing digital elevation data are cropped so that the size of the cropped sample images is consistent with the corresponding digital elevation data. The sample images are uniformly scaled to square. Non-square sample images are cropped to square without compromising the integrity of the landslide.

4. The semi-supervised deep learning-based intelligent remote sensing identification method for large landslides in mountainous areas according to claim 1, characterized in that, The sample normalization process in S1 is as follows: The landslide remote sensing image dataset containing digital elevation data obtained in S1 is subjected to image normalization processing. The original images of the input layer are unified, and the range of feature values ​​is reduced. The formula is as follows: （1） This corresponds to the pixel value after image normalization. These are the pixel values ​​of each point in the image to be processed. and These are the minimum and maximum pixel values ​​found in the image, respectively.

5. The semi-supervised deep learning-based intelligent remote sensing identification method for large landslides in mountainous areas according to claim 1, characterized in that, The process of labeling some landslide remote sensing image data containing digital elevation data in S1 is as follows: using the ArcMap editor, a small number of feature areas are drawn as vector patches for labeling, and bounding boxes are generated. The feature envelope rectangle in the ArcMap tool is converted into a polygon to directly generate the minimum wrapping rectangle, and then the output is saved as a slice. The ratio of samples with labels to samples without labels is 1:

8.

6. The semi-supervised deep learning-based intelligent remote sensing identification method for large landslides in mountainous areas according to claim 1, characterized in that, The training process for constructing the student training network model and the teacher training network model in S2 is as follows: The Mask R-CNN algorithm based on the Tensorflow deep learning framework was used for training and testing. Manually labeled samples and labeled samples from some public datasets were used for recognition training. The feature map was obtained and output based on the backbone network ResNet, and then sampled and merged into the large landslide feature output through the feature pyramid FPN.

7. A semi-supervised deep learning-based intelligent remote sensing identification method for large landslides in mountainous areas according to claim 1 or 6, characterized in that, The process in S2 of predicting the corresponding training sample sets through student-trained network models and teacher-trained network models is as follows: The semantic segmentation network in S2 is a candidate region network. This network generates anchor boxes centered on each pixel based on the large landslide features obtained in S2, where the coordinates of the ground truth boxes are... , The width and height of the actual frame are , The position coordinates of the anchor frame are , The offset formula is: （2） （3） By moving anchor boxes across various images in a large landslide dataset, the candidate region network is trained. Then, classification and regression tasks are performed on these anchor boxes. The classification task calculates the score of each anchor box to determine the probability of whether a landslide has occurred. The regression task finds the position of the anchor box, which is represented by the top left and bottom right points. The anchor boxes are sorted according to their scores and positive and negative examples are distinguished based on their probabilities. Candidate region network classification is based on the Softmax function, which is as follows: （4） in This represents the probability calculated forward for category j. The probability is calculated using the Softmax function. The cross-entropy loss function used is as follows: （5） in Indicates the true label, It represents probability.

8. The semi-supervised deep learning-based intelligent remote sensing identification method for large landslides in mountainous areas according to claim 1, characterized in that, The process in S2 of inputting the prediction results into the discriminator to determine the landslide feature type of the prediction results corresponding to the student-trained network model and the teacher-trained network model is as follows: After processing the prediction results output by the semantic segmentation network in the student-trained network model, the corresponding large landslide features are extracted using ROI Align, where the ROI Align loss function is: （6） The feature types of these large landslides are determined in a pre-trained ResNet-50 discriminator, while the previously adjusted predicted bounding boxes are regressed.

9. The semi-supervised deep learning-based intelligent remote sensing identification method for large landslides in mountainous areas according to claim 1, characterized in that, In step S3, the mean squared error of the student-trained network model and the teacher-trained network model is calculated. An exponential moving average is then calculated based on the cross-entropy loss of the student-trained network model and the mean squared error of the teacher-trained network model. The process of adjusting and updating the weights in the student-trained network model based on the feedback from the exponential moving average is as follows: The loss value of the student training network model is calculated based on the loss function of Mask R-CNN. In the Mask branch, a fully connected network generates a mask, and the information of each pixel of the real mask and the predicted mask is compared. A binary classification cross-entropy loss function is used. The loss value of the student training network model is calculated as follows: （7） （8） in To predict probabilities, The loss value is 0, which is the ideal value, when the predicted and actual probabilities are the same. The greater the difference between the predicted and actual probabilities, the greater the loss value. Backpropagation is achieved by changing the weight parameters according to the loss value. The mean squared error between the student-trained network model and the teacher-trained network model is calculated, which is the expected distance between the prediction results of the student-trained network model and the prediction results of the teacher-trained network model. The calculation formula is as follows: （9） in, This indicates that students are training the network model. This indicates that the teacher is training the network model. This represents the expected distance between the predictions of the student-trained network model and the teacher-trained network model. This represents the expected value between the student-trained network model and the teacher-trained network model. This represents the prediction result of the training. The weights representing the weights used by students to train the network model. The weights representing the weights used by the teacher to train the network model. This represents the noise used by students to train the network model. This represents the noise generated by the teacher training the network model, through the input weights ( , ) and noise ( , The mean squared error is calculated by combining the prediction results of the training and using the mean squared error as the consistency cost. The accuracy and reliability of the training results of the student-trained network model and the teacher-trained network model are judged based on the mean squared error. The exponential moving average is calculated based on the cross-entropy loss and mean squared error obtained from the student-trained network model. The weights in the student-trained network model are then adjusted and updated based on the exponential moving average. The formula for calculating the exponential moving average is as follows: （10） in This represents the loss value used by students to train the network model. The weights represent the mean squared error.

10. A semi-supervised deep learning-based intelligent remote sensing identification method for large landslides in mountainous areas according to claim 1, characterized in that, The process in S3 of updating the weights of the teacher's training network model by continuously predicting and weighting the updated weights of the student training network model is as follows: The following formula is used: （11） in Represents the smoothing coefficient parameter. The weights representing the weights used by students to train the network model. The weights representing the weights used by the teacher to train the network model. This represents the number of training iterations, determined by the weights of the network model trained by the teacher during the (t-1)th training iteration. The weights of the network model trained by the student in the t-th training session Update the weights of the network model trained by the teacher in the t-th training iteration. .

Images