Semantic segmentation method for landslide detection using medium-resolution multi-source remote sensing data

Abstract

The application provides a semantic segmentation method for landslide detection by using medium-resolution multi-source remote sensing data, mainly including five steps: data selection and download, data preprocessing, model construction, model training and result evaluation. The application faces medium-resolution multi-source remote sensing data, proposes a new double-encoder semantic segmentation network with a self-attention mechanism, realizes feature expression and hierarchical fusion of multi-modal data, and further improves the landslide detection precision, thereby providing a technical reference for landslide detection in a large spatial scale.

Description

TECHNICAL FIELD

[0001] The application provides a semantic segmentation method for landslide detection using medium-resolution multi-source remote sensing data, and belongs to the technical field of natural disaster information analysis. BACKGROUND

[0002] Landslide is a global natural disaster, which refers to the geological phenomenon of rock-soil body sliding along the layer under the action of gravity. There are many factors inducing landslide, such as natural factors such as earthquake, volcanic eruption, heavy rainfall, climate change, and human activity factors such as road construction, residential expansion, vegetation cutting, groundwater exploitation. The traditional landslide detection method based on field investigation usually takes a long time and costs a lot, and is affected by the inconvenience of transportation after the disaster and the safety of the investigators, so it is usually difficult to extract and map the landslides in the region continuously and completely. Therefore, landslide detection based on remote sensing technology has become the main research method in the related field.

[0003] Libo Cheng et al. modified the target detection network YOLOv4, added an attention mechanism in the network, and proposed a new YOLO-SA network for landslide detection. The method comes from the journal paper: Cheng L, Li J, Duan P, et al. A small attentional YOLO model for landslide detection from satellite remote sensing images [J]. Landslides, 2021, 18(8): 2751-2765. The shortcomings of this method are:

[0004] (1) This method finally marks the position and approximate range of landslide on the image block with a rectangular box, and does not extract the accurate landslide boundary, which has limited application in actual business;

[0005] (2) This method is for high spatial resolution remote sensing images, and is affected by the access, limited coverage and small image width of high spatial resolution remote sensing images, so it is difficult to be popularized in large spatial scale business.

[0006] Ji Shunping et al. introduced attention mechanism in convolutional neural network model to distinguish landslide scenes and non-landslide scenes in remote sensing images. This method comes from the journal paper: Ji, Shunping, et al. "Landslide detection from an open satellite imagery and digital elevation model dataset using attention boosted convolutional neural networks." Landslides 17.6 (2020): 1337-1352. The disadvantages of this method are:

[0007] (1) This method gives a semantic label to a remote sensing image, only visualizes potential landslides, and does not classify image elements or targets, so it cannot determine the detailed boundaries of landslides;

[0008] (2) Artificially defined threshold to determine whether the remote sensing image belongs to the landslide or non-landslide category, the selection of the threshold will affect the accuracy of the classification;

[0009] (3) Lack of reliable data fusion strategy for multi-modal. Optical bands and DEM data have different data meanings, the former represents reflectivity, and the latter contains terrain information. This method only concatenates them at the data level, without fusion at a higher level such as the feature level or decision level.

[0010] (4) This method is for high spatial resolution remote sensing images, and is affected by the high cost of accessing high spatial resolution remote sensing images, limited coverage, and small image width. It is difficult to generalize this method to large spatial scale business.

[0011] L. Bragagnolo et al. used the U-Net model to realize landslide detection on Landsat 8 images. This method comes from the journal paper Bragagnolo, L., et al. "Convolutional neural networks applied to semantic segmentation of landslide scars." Catena 201 (2021): 105189. The disadvantages of this method are:

[0012] (1) Manually selects three bands of multi-band medium-resolution remote sensing images as input data for the model, ignoring the possible contribution of other bands to landslide detection;

[0013] (2) The model has low extraction accuracy. SUMMARY

[0014] To solve the above technical problems, the present application provides a semantic segmentation method for landslide detection using medium-resolution multi-source remote sensing data, which aims to solve the technical problems:

[0015] (1) Accurate pixel-by-pixel landslide detection is achieved, and the extraction result is an accurate description of the landslide location, range and boundary;

[0016] (2) In the process of landslide detection, no high-resolution remote sensing image is used, only publicly available medium-resolution remote sensing data is used;

[0017] (3) In the process of landslide detection, no classification threshold is defined by human, and end-to-end and fully automatic landslide detection is achieved;

[0018] (4) For multi-source remote sensing data (optical bands and DEM data), a feature-level deep fusion strategy is adopted to improve the feature learning and expression ability of the model for different modal data.

[0019] (5) In view of the characteristics of complex scene and rich details of remote sensing image, more attention is paid to the target information in the image, and the attention to the background information is suppressed, the deep learning model is optimized, and the landslide detection precision is improved.

[0020] The specific technical scheme of the present application:

[0021] The semantic segmentation method for landslide detection using medium-resolution multi-source remote sensing data mainly includes five steps: data selection and download, data preprocessing, model construction, model training and result evaluation.

[0022] (1) Data download and label generation.

[0023] The present application collects Sentinel-2 images and NASA digital elevation model NASADEM data for landslide detection.

[0024] Firstly, all Sentinel-2 images of the study area are retrieved through Google Earth Engine, and the images with high cloud cover are filtered out, and then the composite image of the study area is generated by taking the minimum value; with the help of Google Earth, the landslide label of the study area is obtained by manual interpretation, so as to build the landslide label database. Finally, the generated landslide label is binary data, in which "0" represents background and "1" represents landslide.

[0025] (2) Data preprocessing

[0026] The data used has a total of 15 bands, of which the first to the twelfth bands are the first to the twelfth bands of the Sentinel-2 image, the thirteenth band is the slope, the fourteenth band is the elevation, and the fifteenth band is the aspect. Among them, the slope and aspect are calculated based on the NASADEM data.

[0027] The source data was reprojected, image normalized and image patch generated pre-processed. The obtained Sentinel-2 imagery, NASA DEM data and landslide labels were reprojected by WGS 84 / Pseudo-Mercator (EPSG:3857) and resampled at 10m resolution.

[0028] The images were normalized. Specifically, the average value of each channel data was calculated, thereby generating a set of average value vectors. The feature vector of each pixel was divided by the average value vector pixel by pixel to complete the normalization. For image patch generation, the present application used a sliding window with a size of 128x128 pixels and a span of 128 pixels to scan and crop the normalized images. In addition, the label patch was obtained by scanning and cropping the real landslide label of the ground surface.

[0029] (3) Model construction

[0030] An end-to-end landslide detection network was adopted, and the network architecture included two parts of encoding path and decoding path. The following is a detailed description of the two parts:

[0031] ① Encoding path and feature fusion

[0032] Two groups of encoders were included, wherein the main encoder included 4 encoding blocks, the first encoding block had two convolutional layers, and the other three encoding blocks each had one pooling layer and two convolutional layers; the accompanying encoder included 3 encoding blocks, the first and second encoding blocks had two convolutional layers and one pooling layer, and the third encoding block had two convolutional layers; the main encoder took the optical band as input, and the accompanying encoder took the DEM data as input;

[0033] The features extracted by the two encoders were fused layer by layer; the output features of the same layer convolutional block in the main encoder and the accompanying encoder were concatenated as the input of the next layer convolutional block of the main encoder; an intermediate structure was also included for connecting the encoding path and the decoding path. The intermediate structure included one maximum pooling layer, two convolutional layers and one up-sampling layer.

[0034] ② Decoding path and attention mechanism.

[0035] The up-sampling layer restored the feature map extracted by the encoder to the original size layer by layer, highlighting the details of the target in space; the convolutional layer further optimized the features. Four decoding blocks were included, each of which had an inverse convolutional layer as the up-sampling layer, and two convolutional layers for optimizing the features and adjusting the number of feature channels. In the last decoding block, a 1x1 convolutional layer was finally set to obtain the mapping classification result. Finally, the decoder generated a feature vector with a length of the number of categories at each pixel position.

[0036] The features generated by the same level of encoder and decoder are fused by using the skip connection, so that the decoder can retain more high-resolution detail information in the high-level feature layer when up-sampling the feature layer, and the self-attention mechanism is introduced in the skip connection, so that the model can suppress irrelevant background information and focus on useful information.

[0037] (4) Model training

[0038] The step-by-step optimization of the model parameters is realized by minimizing the loss function. The loss function L integrates the weighted cross-entropy loss function L WCE and the Dice loss function L Dice . The joint loss function L has the advantages of cross-entropy loss and Dice loss. The calculation method of L is as follows:

[0039] L=(1-α)L WCE +αL Dice (1)

[0040] Wherein α∈[0,1] is the balance parameter. The weighted cross-entropy loss function L WCE is a variant of cross-entropy loss, and the calculation method is as follows:

[0041]

[0042] Wherein N represents the number of training samples, K represents the number of classes, β k is the weight value set for the kth class, y n is the one-hot label of the nth sample, is the prediction value of the nth sample.

[0043] The Dice coefficient is between 0 and 1, and is used to evaluate the similarity between two samples. The calculation method of the Dice loss L Dice is as follows:

[0044]

[0045] (5) Model evaluation

[0046] Four indicators are used to evaluate the classification effect, namely accuracy precision, recall, overall accuracy OA and F1 score. These indicators are defined as follows:

[0047]

[0048]

[0049]

[0050]

[0051] Wherein, TP is the true positive number, that is, the pixel correctly classified as landslide; TN is the true negative number, that is, the pixel correctly classified as background; FP is the false positive number, that is, the ground real background pixel incorrectly classified as landslide; FN is the false negative number, that is, the ground real landslide pixel incorrectly classified as background. OA represents the percentage of pixels correctly classified among all pixels including background pixels and landslide pixels. F1 is considered to be the harmonic mean of precision and recall, which can balance between accuracy and recall.

[0052] Advantages brought by the technical scheme of the present application

[0053] (1) The present application constructs a set of landslide detection method based on medium resolution multi-source remote sensing data (including optical bands and DEM data), which realizes pixel-by-pixel landslide detection, so as to depict the position and boundary of landslide.

[0054] (2) The present application does not use any high-quality high-spatial-resolution remote sensing image, but achieves a landslide detection effect comparable to that of using high-resolution remote sensing image. For example, in the same study area, using Planet (spatial resolution 3m) image to carry out landslide detection, the final F1 precision is 79.33%, and the F1 precision obtained by the present application using medium resolution image (spatial resolution 10m) is 79.24%, the difference between the two is only 0.09%. The present application can provide technical reference for large-scale regional landslide detection.

[0055] (3) The landslide detection model designed by the present application is end-to-end and fully automatic, without any manual intervention.

[0056] (4) The present application uses a dual-encoder architecture to learn the deep features of optical bands and DEM data respectively, and layer-by-layer fuses the two types of features, so the present application fuses multi-modal data at the feature level, which has better feature expression ability compared with the data level fusion strategy.

[0057] (5) The present application is not limited to existing deep learning models, but optimizes the deep learning model. On the one hand, it fuses the features of multi-modal data, and on the other hand, it uses self-attention modules to improve the model's attention to the target and suppress the attention to the background. The present application has achieved a landslide detection effect better than the classic semantic segmentation network. Through experiments, the present application has exceeded the F1 precision of SegNet, U-Net and Att U-Net models by 5.67-19.83%, 2.66-3.58% and 2.62-3.33% respectively. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 The flowchart of the present application;

[0059] Figure 2 Fig. 1 is a schematic diagram of a landslide detection network structure according to the present application;

[0060] Figure 3 Fig. 2 is a flowchart showing the operation process of the self-attention mechanism according to the present application. DETAILED DESCRIPTION

[0061] The specific technical solutions of the present application will be described in conjunction with the drawings.

[0062] As shown in Fig. 1, the semantic segmentation method for landslide detection using medium-resolution multi-source remote sensing data mainly includes five steps: data selection and download, data preprocessing, model construction, model training, and result evaluation. Figure 1 (1) Data download and label generation

[0063] The present application collects Sentinel-2 images and NASA Digital Elevation Model (NASADEM) data for landslide detection. The Sentinel-2 image has 13 bands with different spatial resolutions, among which the visible and near-infrared bands have a resolution of 10 meters, the red edge and short-wave infrared bands have a resolution of 20 meters, and the atmospheric band has a resolution of 60 meters. The spatial resolution of NASADEM data is 30 meters, and the pixel value represents the elevation.

[0064] The Sentinel-2 and NASADEM data used in the present application are collected and downloaded for free using Python API and Google Earth Engine (https: / / developers.google.com / earth-engine / ). The present application first retrieves all Sentinel-2 images of the study area through Google Earth Engine, filters out images with high cloud cover, and then generates a composite image of the study area with the minimum value. Compared with a single image, the quality of the composite image is much better and the cloud cover is less.

[0065] With the help of Google Earth, the present application obtains the landslide labels of the study area through manual interpretation to build a landslide label database. The final landslide label is binary data, where "0" represents the background and "1" represents the landslide.

[0066] (2) Data preprocessing

[0067] The data used in the present application has a total of 15 bands, among which the first to the twelfth bands are the first to the twelfth bands of the Sentinel-2 image, the thirteenth band is the slope, the fourteenth band is the elevation, and the fifteenth band is the aspect. Among them, the slope and aspect are calculated based on the NASADEM data.

[0068]

[0069] ​The present application carries out re-projection, image normalization and image block generation and other preprocessing on the source data. The obtained Sentinel-2 image, NADASEM data and landslide label are re-projected by WGS 84 / Pseudo-Mercator (ESPG: 3857) and resampled at a resolution of 10 meters.

[0070] In addition, the present application carries out image normalization. Specifically, the average value of each channel data is calculated, thereby generating a set of average value vectors, and the feature vector of each pixel is divided by the average value vector pixel by pixel to complete the normalization. For image block generation, the present application uses a sliding window with a size of 128x128 pixels and a span of 128 pixels to scan and crop the normalized image. In addition, the generation of the label block is also obtained by scanning and cropping the real landslide label of the ground surface.

[0071] (3) Model construction

[0072] The present application proposes a new end-to-end landslide detection network, and the network architecture is as follows Figure 2

[0073] The network mainly includes two parts of encoding path and decoding path. The following is a detailed description of the two parts:

[0074] ① Encoding path and feature fusion. In the deep semantic segmentation network, the network encoder usually includes two components of convolutional layer and pooling layer. The convolutional layer extracts features from the input data, and the local range in the input data is linearly weighted by the convolution kernel. The calculation result is nonlinearly transformed by the activation function to obtain the output feature map. The convolution kernel parameters are trainable, and are adaptively adjusted by back propagation and parameter update. The pooling layer is used to change the size of the feature map, and the data is reduced in dimension to obtain higher level feature expression and reduce the computational complexity. The commonly used pooling methods include maximum pooling and average pooling.

[0075] The present application designs two groups of encoders, wherein the main encoder includes 4 encoding blocks, the first encoding block has two convolutional layers, and the next 3 encoding blocks each have one pooling layer and two convolutional layers; the accompanying encoder includes 3 encoding blocks, the first and second encoding blocks have two convolutional layers and one pooling layer, and the third encoding block has two convolutional layers without a pooling layer. The main encoder takes the optical band as the input, and the accompanying encoder takes the DEM data as the input. The present application uses the maximum pooling layer to reduce the computational complexity and aggregate the depth information in the feature map generated by different levels of convolution. By stacking the convolutional layer and the maximum pooling layer, the network realizes hierarchical feature expression and synthesis of the high-level semantic features of the input data.

[0076] ​Meanwhile, the features extracted by the two encoders are fused layer by layer. Specifically, the output features of the same convolution block in the main encoder and the accompanying encoder are concatenated as the input of the next convolution block in the main encoder. The fusion strategy of feature levels in the double-branch structure helps to improve the feature expression ability and consistency level of the network. Formula (1) describes the process of generating feature maps after the data passes through the encoding block of the main encoder (except the first encoding block):

[0077] wherein δ, respectively represent the main encoder and the accompanying encoder, and l represents the encoding block number; m represents the feature map size, d(l) represents the number of channels of the output feature map of the lth layer, then represents the output feature of the updated main encoding block, represents the output feature of the previous main encoding block, represents the output feature of the previous accompanying encoding block; || represents the feature concatenation operation; and represent the trainable parameter matrices of the two convolution layers, BN represents batch normalization, σ(·) represents an activation function, and MaxPool 2×2 (·) represents maximum pooling.

[0078] Formula (2) describes the process of generating feature maps after the data passes through an encoding block in the accompanying encoder θ:

[0079]

[0080] wherein represents the output feature map, represents the input feature map, and respectively represent the trainable parameter matrices of the two convolution layers in the encoding block.

[0081] In addition, an intermediate structure is used to connect the encoding path and the decoding path. The intermediate structure includes a maximum value pooling layer, two convolution layers, and an up-sampling layer.

[0082] ② Decoding path and attention mechanism. The decoder is usually composed of up-sampling layers and convolution layers. The up-sampling layers restore the feature maps extracted by the encoder to the original size level by level, highlighting the details of the target in space; the convolution layers further optimize the features. The network proposed in the present application includes 4 decoding blocks, each of which has a deconvolution layer as an up-sampling layer, and two convolution layers for optimizing features and adjusting the number of feature channels. In the last decoding block, a 1x1 convolution layer is finally set to obtain the mapping result. Finally, the decoder generates a feature vector with a length of the number of categories at each pixel position.

[0083] The U-Net series network adopts a skip connection (usually a concatenation layer) to fuse the features generated by the encoder and the decoder at the same level, so that the decoder can retain more high-resolution detail information in the high-level feature layer when up-sampling the feature layer, which helps to improve the pixel-by-pixel segmentation accuracy. A self-attention mechanism is introduced in the skip connection to better fuse the features, so that the model can suppress irrelevant background information and focus on useful information.

[0084] As shown in Figure 3 , the running process of the self-attention mechanism is shown. In the figure, X represents the up-sampled feature map from the decoder block; H represents the feature map from the corresponding encoder block; m is the size of the feature map; d_h and d_x represent the number of feature channels of H and X; d_int and d_out represent the number of intermediate feature channels and output feature channels. H and X are first updated by a 1x1 convolution layer. Then, the obtained feature maps are added element by element. The generated intermediate feature map is sent to another 1x1 convolution layer to generate attention coefficients P. The attention coefficients P are multiplied with the input X element by element to obtain the attentioned X. The updated attentioned X is connected with the input H to finally generate the output feature map.

[0085] (4) Model training

[0086] The network proposed in the present application is trained under the guidance of the joint loss function L, and the step-by-step optimization of the model parameters is realized by minimizing the loss function. The loss function L integrates the weighted cross-entropy loss function L WCE and the Dice loss function L Dice . The joint loss function L has the advantages of both cross-entropy loss and Dice loss. The calculation method of L is as follows:

[0087] L=(1-α)L WCE +αL Dice (3)

[0088] wherein α∈[0,1] is a balance parameter. The weighted cross-entropy loss function L WCE is a variant of the cross-entropy loss, and the calculation method is as follows:

[0089]

[0090] wherein N represents the number of training samples, K represents the number of classes, β k is the weight value set for the kth class, y n is the one-hot label of the nth sample, is the prediction value of the nth sample. The weighted cross-entropy is very useful for alleviating the imbalance problem of classes.

[0091] The Dice loss, originally from the Dice coefficient, was originally designed to cope with strong class imbalance. The Dice coefficient is between 0 and 1, used to evaluate the similarity between two samples. The Dice loss L Dice The calculation method of the Dice loss L is as follows:

[0092]

[0093] (5) Model evaluation

[0094] The present application adopts four indexes to evaluate the classification effect, i.e. precision, recall, overall accuracy (OA) and F1 score (F1). These indexes are defined as follows.

[0095]

[0096]

[0097]

[0098]

[0099] Wherein, TP is the true positive number, i.e. the pixel correctly classified as landslide; TN is the true negative number, i.e. the pixel correctly classified as background; FP is the false positive number, i.e. the ground true background pixel incorrectly classified as landslide; FN is the false negative number, i.e. the ground true landslide pixel incorrectly classified as background. OA represents the percentage of pixels correctly classified among all pixels including background pixels and landslide pixels. F1 is considered to be the harmonic mean of precision and recall, which can balance between precision and recall.

[0100] The present application faces the medium resolution multi-source remote sensing data, proposes a new dual-encoder semantic segmentation network with self-attention mechanism, realizes the feature expression and hierarchical fusion of multi-modal data, and further improves the landslide detection precision, providing technical reference for large spatial scale landslide detection.

Claims

1.A semantic segmentation method for landslide detection using medium-resolution multi-source remote sensing data, characterized in that, It includes five steps: data selection and download, data preprocessing, model construction, model training, and result evaluation. Specifically: (1) Data download and label generation All Sentinel-2 images of the study area are retrieved through Google Earth Engine, and images with high cloud cover are filtered out, and then the composite image of the study area is generated with the minimum value; With the help of Google Earth, the landslide label of the study area is obtained through manual interpretation, and the landslide label database is constructed; Finally, the generated landslide label is binary data, where "0" represents the background and "1" represents the landslide; (2) Data preprocessing The data used has a total of 15 bands, of which the first 12 bands are the first 12 bands of the Sentinel-2 image, the 13th band is the slope, the 14th band is the elevation, and the 15th band is the aspect; Among them, the slope and aspect are calculated based on NASADEM data; The source data is reprojected, image normalized and image block generated preprocessing; The obtained Sentinel-2 image, NASADEM data and landslide label are reprojected by WGS 84 / Pseudo-Mercator and resampled at 10m resolution; (3) Model construction An end-to-end landslide detection network is adopted, which has two parts: encoding path and decoding path. The encoding path part is specifically: It includes two groups of encoders, of which the main encoder includes 4 encoding blocks, the first encoding block has two convolutional layers, and the next three encoding blocks each have one pooling layer and two convolutional layers; The accompanying encoder includes three encoding blocks, the first and second encoding blocks have two convolutional layers and one pooling layer, and the third encoding block has two convolutional layers; The main encoder takes the optical band as input, and the accompanying encoder takes the DEM data as input; The features extracted by the two encoders are fused layer by layer; The output features of the convolutional blocks in the same layer of the main encoder and the accompanying encoder are concatenated as the input of the next convolutional block of the main encoder; An intermediate structure is used to connect the encoding path and the decoding path; The intermediate structure includes one maximum pooling layer, two convolutional layers and two up-sampling layers; The decoding path part is specifically: The up-sampling layer restores the feature map extracted by the encoder to the original size layer by layer, highlighting the details of the target in space; The convolutional layer further optimizes the features; It contains 4 decoding blocks, each of which has a deconvolution layer as an up-sampling layer, and 2 convolutional layers for optimizing features and adjusting the number of feature channels; In the last decoding block, a 1x1 convolutional layer is set at the end to obtain the mapping classification result; Finally, the decoder generates a feature vector with a length of the number of classes at each pixel position; The skip connection is used to fuse the features generated by the encoder and the decoder in the same layer, so that the decoder retains the high-resolution detail information in the high-level feature layer when up-sampling the feature layer, and the self-attention mechanism is introduced in the skip connection, so that the model suppresses irrelevant background information and focuses on useful information; (4) Model training The step-by-step optimization of the model parameters is achieved by minimizing a loss function; the loss function L integrates a weighted cross-entropy loss function L WCE and a Dice loss function L Dice ; (5) Model evaluation Four indicators are used to evaluate the classification results, i.e., precision, recall, overall accuracy (OA) and F1 score. 2.The semantic segmentation method for landslide detection using medium-resolution multi-source remote sensing data according to claim 1, characterized in that, In step (2) of data preprocessing, the image is normalized, and the average value of each channel data is calculated to generate a set of average value vectors. The feature vector of each pixel is divided by the average value vector pixel by pixel to complete the normalization. 3.The semantic segmentation method for landslide detection using medium-resolution multi-source remote sensing data according to claim 1, wherein, In step (2) of data preprocessing, for image block generation, a sliding window with a size of 128x128 pixels and a span of 128 pixels is used to scan and crop the normalized image. 4.The semantic segmentation method for landslide detection using medium-resolution multi-source remote sensing data according to claim 1, wherein, In step (2) of data preprocessing, the generation of label blocks is also obtained by scanning and cropping the ground truth landslide labels. 5.The semantic segmentation method for landslide detection using medium-resolution multi-source remote sensing data according to claim 1, wherein, In step (4), the calculation method of the loss function L is as follows: L = (1 - a)L WCE + aL Dice where a e [0, 1] is a balance parameter; the weighted cross-entropy loss function L WCE is a variant of the cross-entropy loss, computed as follows: where N represents the number of training samples, K represents the number of classes, β k is the weight value of the kth class, represents the model output value of the nth sample to the kth class; y nk represents the real one-hot label of the nth sample to the kth class. The Dice coefficient is between 0 and 1, and is used to evaluate the similarity between two samples. The Dice loss L Dice The calculation method is as follows: 6.The semantic segmentation method for landslide detection using medium-resolution multi-source remote sensing data according to claim 1, wherein, In step (5), the indicators are defined as follows: Where TP is the true positive number, i.e., the pixels correctly classified as landslide; TN is the true negative number, i.e., the pixels correctly classified as background; FP is the false positive number, i.e., the ground truth background pixels incorrectly classified as landslide; FN is the false negative number, i.e., the ground truth landslide pixels incorrectly classified as background; OA represents the percentage of pixels correctly classified among all pixels including background pixels and landslide pixels; F1 is considered to be the harmonic mean of precision and recall, which can balance between accuracy and recall.

Images