A landslide detection and early warning model and early warning method based on multi-model fusion

Abstract

The present invention discloses a landslide detection and early warning method based on multi-model fusion, comprising: (1) obtaining a remote sensing image of a region to be detected, and cutting it into a plurality of sub-remote sensing images; (2) classifying and identifying the remote sensing images as landslide images or non-landslide images in turn; (3) outputting a detection frame for each landslide image; (4) enlarging the detection frame size and filling the background with black; (5) obtaining a semantic segmentation result image; (6) marking the semantic segmentation result image obtained in step (5) as a reference image; (7) monitoring the sub-region of each landslide image, and giving an alarm by monitoring the area value of the deformed region. In the process of processing the landslide image into the semantic segmentation result image, the present invention improves target detection, detection frame enlargement, and semantic segmentation, and can improve the landslide detection accuracy under multi-scale and complex backgrounds, especially the landslide detection accuracy under small scale.

Description

Technical Field

[0001] The invention relates to a landslide detection and early warning method, and in particular to a landslide detection and early warning method based on multi-model fusion. Background Art

[0002] Traditional image processing and deep learning methods can achieve landslide detection and early warning using remote sensing images. However, there are still many problems:

[0003] (1) Landslides usually appear at different scales, ranging from a few meters to several kilometers in length. The change in scale will bring challenges to landslide recognition. Some landslide recognition models are mainly targeted at large-scale landslide images. However, for landslide images with complex backgrounds and different size ratios, especially small-scale landslide images, the missed detection rate and false detection rate are high. There are two main reasons for the low detection rate of small-scale landslide images. First, small-scale landslides belong to the details of the image. They have low resolution and occupy fewer pixels in the image. Especially after the pooling process, the features of small-scale landslides are more likely to be lost during the encoding process. For example, an image of size [512,512] is downsampled to size [256,256] after the pooling layer, which will lose some pixel information. This part of the lost pixel information is often the details of the image. Secondly, with the deepening of the number of network layers and the increase of the receptive field, the features of small-scale landslide images are more difficult to retain than those of large-scale landslide images. This means that general object detection models or semantic segmentation models are good at extracting medium-scale or large-scale landslide features, but largely ignore small-scale landslides.

[0004] (2) The background of mountain remote sensing images is complex, including houses, riverbeds and other features, and a remote sensing image may contain multiple irregular landslide areas, which will also interfere with landslide identification. For landslide identification with complex backgrounds, areas with spectral characteristics similar to those of the landslide area often have false detections. There are two main types of errors: one type occurs mainly in places with complex surface textures, such as terrace changes, natural erosion or ridge intersections. Since the different surface textures in these places may form a round chair or arc-shaped boundary morphology similar to the rear edge of the landslide, it may lead to false identification. The other type occurs mainly in areas containing multiple irregular landslides, where the landslide boundaries are crisscrossed, the landslides are of different shapes and sizes, and the irregular shapes can also lead to false detections.

[0005] Glossary:

[0006] SPP network: spatial pyramid pooling in English and spatial pyramid pooling in Chinese.

[0007] PANet: Path Aggregation Network in English, path aggregation network in Chinese. Its greatest contribution is the proposal of a top-down and bottom-up bidirectional fusion backbone network, while adding a "short-cut" between the bottom and top layers to shorten the path between layers. Summary of the invention

[0008] The purpose of the present invention is to provide a landslide detection and early warning method based on multi-model fusion, which can solve the above problems and improve the landslide detection accuracy under multi-scale and complex backgrounds, especially the landslide detection accuracy under small scale.

[0009] In order to achieve the above object, the technical solution adopted by the present invention is as follows: a landslide detection and early warning method based on multi-model fusion, comprising the following steps;

[0010] (1) Obtain a remote sensing image of the area to be measured and cut it into multiple sub-remote sensing images, each of which corresponds to a sub-area within the area to be measured;

[0011] (2) Classifying and identifying the sub-remote sensing images in turn as landslide images or non-landslide images, and only retaining the landslide images;

[0012] (3) outputting a detection frame for each landslide image, including steps (31)-(32);

[0013] (31) Use the target detection model to detect the landslide area in the landslide image;

[0014] (32) If the target is detected, output the detection frame; if the target is not detected, adjust the input size of the landslide image or the parameters of the target detection model until the target is detected;

[0015] (4) Enlarging the size of the detection frame and then filling the area outside the detection frame in the landslide image with black;

[0016] (5) performing semantic segmentation on the landslide image obtained in step (4) to obtain a semantic segmentation result map;

[0017] (6) Mark the semantic segmentation result map obtained in step (5) as a reference map;

[0018] (7) Monitor the sub-areas of each landslide image, one of which is monitored by;

[0019] (71) obtaining the remote sensing image of the sub-region corresponding to the landslide image again, processing it according to steps (3)-(5), and obtaining a semantic segmentation result image as a comparison image;

[0020] (72) Subtracting the comparison image from the reference image pixel by pixel to obtain a deformation image containing only the deformation area, and calculating the area value S of the deformation area;

[0021] (73) comparing the area value with a preset threshold;

[0022] If S < threshold, repeat steps (71)-(72);

[0023] If S ≥ threshold, an alarm is issued, and the current comparison image is marked as a reference image, and steps (71)-(72) are repeated.

[0024] Preferably, the step (2) performs classification and recognition through a classification model, and the classification model is a SwinTransformer classification model, a ResNet classification model or a VGG classification model.

[0025] Preferably, in the step (31), the target detection model is a YOLO target detection model or an improved target detection model, wherein the improved target detection model comprises a YOLO4 model, wherein the YOLO4 model is composed of a backbone network, an SPP network and a PANet, wherein the backbone network adopts a Desenet121 network.

[0026] Preferably, in step (32), the parameters of the target detection model include the model weight, the target score confidence, the threshold of non-maximum suppression, the alpha coefficient of the Desenet121 network, the probability of mosaic data enhancement, and / or the learning rate of the model.

[0027] Preferably, in step (4), the size of the detection frame is enlarged as follows:

[0028] The coordinates of the i-th vertex of the expanded detection box are calculated by the following formula:

[0029]

[0030]

[0031] Where i = 1 to 4, representing the i-th vertex of the four vertices of the detection box, X′ i and Y′ i are the horizontal and vertical coordinates of the i-th vertex of the enlarged detection box, respectively. i and Y i To expand the horizontal and vertical coordinates of the i-th vertex of the detection box, is the landslide area ratio, S L is the area of ​​the landslide region in the landslide image, S I is the area of ​​the landslide image; w is the zoom factor, and Input is the size of the sub-remote sensing image.

[0032] Preferably, in step (5), semantic segmentation is performed by an improved semantic segmentation model, wherein the improved semantic segmentation model comprises a Unet network, wherein the Unet network comprises an encoding part and a decoding part, wherein the encoding part adopts a resnet50 network, and a CBAM attention mechanism module and a PPM pyramid pooling module are sequentially arranged at an output end of the resnet50 network;

[0033] The resnet50 network includes 5 residual blocks from top to bottom, and corresponds to 5 feature outputs feat1-feat5. The feat5 is sent to the decoding part after passing through the CBAM attention mechanism module and the PPM pyramid pooling module;

[0034] The loss function FL(P t ) is calculated by the following formula:

[0035] FL(P t )=-(1-P t ) γ log(P t )

[0036] Where γ is the modulation coefficient, 0≤γ≤5, P t is the probability function, p is the probability of a pixel point in the landslide image being considered as a landslide;

[0037]

[0038] Preferably, the input size of the resnet50 network is [512,512,3], and the sizes of feat1-feat5 are [256,256,64], [128,128,256], [64,64,512], [32,32,1024], and [16,16,2048], respectively.

[0039] The idea of ​​the present invention is: first, a remote sensing image is obtained and cut into multiple sub-remote sensing images, and the sub-remote sensing images are classified and identified in turn, and only the sub-remote sensing images related to the landslide are retained as the landslide image, so as to facilitate the later key monitoring of the corresponding area of ​​the landslide image. Secondly, after the landslide image is subjected to target detection, detection frame enlargement, and semantic segmentation, the semantic segmentation result image is output. During subsequent monitoring, the remote sensing image of the corresponding area of ​​the landslide image is obtained again for processing. The two semantic segmentation result images are subtracted pixel by pixel to obtain a deformation map, and then statistically added pixel by pixel to obtain the area value of the deformation area caused by the landslide. If the area value is greater than the threshold, an early warning is issued, otherwise, the later remote sensing image is continued to be collected, and the above steps are repeated to achieve detection and early warning of landslide deformation at the same location.

[0040] Compared with the prior art, the advantages of the present invention are: a landslide detection and early warning method based on multi-model fusion is proposed, and in the process of processing the landslide image into a semantic segmentation result map, the present invention improves target detection, detection frame enlargement, and semantic segmentation.

[0041] For target detection: an improved target detection model is proposed, which uses the Desenet121 network to replace the original backbone network in the YOLO4 model. Compared with the traditional convolutional neural network at the same layer depth, it has better convergence effect and can enhance the feature extraction ability of the network, so that more abundant features can be learned for utilization and the accuracy of target detection can be improved.

[0042] Enlarging the detection frame: The present invention expands the corresponding range of different target detection frames that need to be output according to landslides of different scales to ensure that certain scene semantic information is retained without subjecting the model to complex interference, thereby improving the accuracy of semantic segmentation as a whole.

[0043] For semantic segmentation: an improved semantic segmentation model is proposed, and the main improvements are as follows:

[0044] (1) In view of the complex background of mountain remote sensing images, the traditional U-Net backbone feature extraction network was abandoned. Instead, a deep residual network structure of ResNet50 was designed to extract features, corresponding to 5 feature outputs feat1-feat5. This structure can improve the recognition accuracy of landslides in complex backgrounds.

[0045] (2) Adding the CBAM attention mechanism module: In order to make the network learn landslide features more efficiently, we added the CBAM attention mechanism module to the traditional U-Net network structure and subjected the obtained features to an adaptive attention feature extraction process. This allows the network to focus on important information about the landslide with high weights and ignore irrelevant information such as the background with low weights. This allows the network to pay more attention to the features it needs to pay most attention to and complete efficient feature learning.

[0046] (3) Adding PPM pyramid pooling module: To address the problem of large variations in landslide scales and difficulty in identifying small-scale landslides, we added a PPM pyramid pooling module that can aggregate contextual information from different regions and improve the ability to obtain global information. This can enhance the effect of multi-scale landslide feature fusion and improve the recognition accuracy of small-scale landslides.

[0047] (4) Add loss function to solve the problem of sample imbalance, where background pixels account for the majority and landslide pixels account for the minority. Background pixels can be considered as negative samples, while landslide pixels can be considered as positive samples. For small-scale landslide images, this is equivalent to a large number of negative samples and a small number of positive samples. To solve this problem, we use the standard cross entropy loss log(P t ) is based on the weight factor (1-P t ) γ , that is, balancing the cross entropy, which can reduce the proportion of negative sample loss in the total loss.

[0048] In summary, the present invention can improve the landslide detection accuracy under multi-scale and complex backgrounds, especially the landslide detection accuracy under small scale. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flow chart of the present invention;

[0050] Figure 2 A flowchart of processing a sub-remote sensing image into a semantic segmentation result image in the present invention;

[0051] Figure 3 This is a structural diagram of the improved target detection model;

[0052] Figure 4 This is the structure diagram of the improved semantic segmentation model;

[0053] Figure 5a Identify the detection frame in the large-scale landslide image and expand the comparison image after filling;

[0054] Figure 5b Comparison map after identifying the detection frame and expanding and filling it in the medium-sized landslide image;

[0055] Figure 5c Identify the detection frame in the small-size landslide image and expand the comparison image after filling;

[0056] Figure 6 It is a schematic diagram of obtaining a deformation diagram from a comparison diagram and a reference diagram of the present invention;

[0057] Figure 7a Comparison of two large-scale landslide images obtained using three methods;

[0058] Figure 7b Comparison of two medium-sized landslide images obtained using three methods;

[0059] Figure 7c Comparison of two small-scale landslide images obtained using three methods;

[0060] Figure 7dComparison of two landslide images with complex backgrounds obtained using three methods. DETAILED DESCRIPTION

[0061] The present invention will be further described below in conjunction with the accompanying drawings.

[0062] Example 1: See Figure 1-Figure 6 ,A landslide detection and early warning method based on multi-model fusion includes the following steps;

[0063] (1) Obtain a remote sensing image of the area to be measured and cut it into multiple sub-remote sensing images, each of which corresponds to a sub-area within the area to be measured;

[0064] (2) Classifying and identifying the sub-remote sensing images in turn as landslide images or non-landslide images, and only retaining the landslide images;

[0065] (3) outputting a detection frame for each landslide image, including steps (31)-(32);

[0066] (31) Use the target detection model to detect the landslide area in the landslide image;

[0067] (32) If the target is detected, output the detection frame; if the target is not detected, adjust the input size of the landslide image or the parameters of the target detection model until the target is detected;

[0068] (4) Enlarging the size of the detection frame and then filling the area outside the detection frame in the landslide image with black;

[0069] (5) performing semantic segmentation on the landslide image obtained in step (4) to obtain a semantic segmentation result map;

[0070] (6) Mark the semantic segmentation result map obtained in step (5) as a reference map;

[0071] (7) Monitor the sub-areas of each landslide image, one of which is monitored by;

[0072] (71) obtaining the remote sensing image of the sub-region corresponding to the landslide image again, processing it according to steps (3)-(5), and obtaining a semantic segmentation result image as a comparison image;

[0073] (72) Subtracting the comparison image from the reference image pixel by pixel to obtain a deformation image containing only the deformation area, and calculating the area value S of the deformation area;

[0074] (73) comparing the area value with a preset threshold;

[0075] If S < threshold, repeat steps (71)-(72);

[0076] If S ≥ threshold, an alarm is issued, and the current comparison image is marked as a reference image, and steps (71)-(72) are repeated.

[0077] The step (2) performs classification and recognition through a classification model, and the classification model is a Swin Transformer classification model, a ResNet classification model or a VGG classification model.

[0078] In the step (31), the target detection model is a YOLO target detection model or an improved target detection model, and the improved target detection model includes a YOLO4 model, and the YOLO4 model is composed of a backbone network, an SPP network and a PANet, wherein the backbone network adopts a Desenet121 network.

[0079] In the step (32), the parameters of the target detection model include the model weight, the target score confidence, the threshold of non-maximum suppression, the alpha coefficient of the Desenet121 network, the probability of mosaic data enhancement, and / or the learning rate of the model.

[0080] In the step (4), the size of the detection frame is enlarged as follows:

[0081] The coordinates of the i-th vertex of the expanded detection box are calculated by the following formula:

[0082]

[0083]

[0084] Where i = 1 to 4, representing the i-th vertex of the four vertices of the detection box, X′ i and Y′ i are the horizontal and vertical coordinates of the i-th vertex of the enlarged detection box, respectively. i and Y i To expand the horizontal and vertical coordinates of the i-th vertex of the detection box, is the landslide area ratio, S L is the area of ​​the landslide region in the landslide image, S I is the area of ​​the landslide image; w is the zoom factor, and Input is the size of the sub-remote sensing image.

[0085] In the step (5), semantic segmentation is performed by an improved semantic segmentation model, wherein the improved semantic segmentation model includes a Unet network, wherein the Unet network includes an encoding part and a decoding part, the encoding part adopts a resnet50 network, and a CBAM attention mechanism module and a PPM pyramid pooling module are sequentially set at the output end of the resnet50 network;

[0086] The resnet50 network includes 5 residual blocks from top to bottom, and corresponds to 5 feature outputs feat1-feat5. The feat5 is sent to the decoding part after passing through the CBAM attention mechanism module and the PPM pyramid pooling module;

[0087] The loss function FL(P t ) is calculated by the following formula:

[0088] FL(P t )=-(1-P t ) γ log(P t )

[0089] Where γ is the modulation coefficient, 0≤γ≤5, P t is the probability function, p is the probability of a pixel point in the landslide image being considered as a landslide;

[0090]

[0091] The input size of the resnet50 network is [512, 512, 3], and the sizes of feat1-feat5 are [256, 256, 64], [128, 128, 256], [64, 64, 512], [32, 32, 1024], and [16, 16, 2048], respectively.

[0092] In this embodiment, we set w=10, Input=416.

[0093] In addition, the present invention uses represents the landslide area ratio. We can classify the landslide images into large-scale landslides, medium-scale landslides and small-scale landslides according to the landslide area ratio. When , it can be considered as a large-scale landslide; when When , it can be considered as a medium-scale landslide; when It can be considered as a small-scale landslide.

[0094] See also Figure 5a-5c , respectively, gives the comparison pictures after the detection frame is identified and expanded and filled in the large, medium and small landslide images. Figure 5a The left picture shows that the remote sensing image recognizes two targets and outputs two detection frames. After enlarging them, two enlarged detection frames are obtained. Figure 5a The two enlarged detection frames overlap, so we merge them and fill the rest of the background area with black. Figure 5a As shown in the picture on the right.

[0095] Figure 5b , Figure 5cThere is a landslide area in each of them, so there is only one target, and a detection frame is output. After enlarging, the background area is filled with black.

[0096] Expanding the detection frame can retain more scene semantic information. This is because the output target detection frame sizes are different for landslides of different scales. A large target detection frame can retain more scene semantic information, while a small target detection frame will block out a lot of scene semantic information related to landslides, such as vegetation coverage, color difference between landslides and the environment, etc. These are very important context information features, and the semantic segmentation model also needs to identify landslides based on these scene semantic features. Therefore, we need to expand the corresponding range based on the different target detection frames output by landslides of different scales to ensure that certain scene semantic information is retained without causing the model to be subject to complex interference.

[0097] This paper proposes an improved target detection model, which replaces the original backbone network in the YOLO4 model with the Desenet121 network. Figure 3 , the input of the Desenet121 network structure is [416,416,3] in size, and then after convolution, pooling, activation and other operations, the feature layer of size [104,104,64] is obtained, and then after two Dense Block modules and two Transition Layer modules, the output of the first feature layer is [26,26,256] in size; after a Dense Block module and a Transition Layer module, the output of the second feature layer is [13,13,512] in size; finally, after a Dense Block module and pooling activation operations, the output of the third feature layer is [13,13,1024] in size, which completes the replacement of the YOLOv4 backbone feature extraction network. In fact, the input and output are not limited to the sizes described in this article and can be adjusted according to actual conditions.

[0098] Regarding the PPM pyramid pooling module, the specific location of this module is after the output feature layer of the CBAM attention mechanism. The specific process is to divide the acquired feature layer into 1×1, 2×2, 4×4, and 8×8 sub-regions, respectively, and perform average pooling in each sub-region. Then, upsample the obtained feature maps to restore them to the original feature map size, and finally splice them in the channel dimension to obtain the final composite feature map.

[0099] Example 2: See Figure 1-7d , we preset the experimental conditions for the experiment.

[0100] The hardware server configuration is: NVIDIA A4000 graphics card, INTEL Xeon 4210R ten-core processor and Samsung 32G memory, etc., and the operating environment is Windows 10 64-bit system. The remote sensing images used in the dataset are from the Bijie City dataset and the self-made dataset, with a total of 3103 remote sensing images. The dataset is divided according to the ratio of training set plus validation set: test set equal to 9:1. We divide the landslides into large-scale landslides, medium-scale landslides, small-scale landslides and landslides in complex background areas according to the ratio of pixels in the image and the complexity of the background. The ratio of pixels also indicates the percentage of the landslide area to the entire image area.

[0101] Large-scale landslides are images where the landslide area accounts for more than 50% of the total area;

[0102] Mesoscale landslides are images where the landslide area accounts for between 10% and 50% of the total area;

[0103] Small-scale landslides are defined as images where the landslide area is less than 10% of the total area;

[0104] The landslides in complex background areas are: those with more than two irregularly shaped landslides, interference from clouds and fog, interference from buildings and mountain terraces, interference from exposed rock and soil structures or other interference images.

[0105] In order to compare and analyze the differences, advantages and disadvantages of different methods in multi-scale complex landslide remote sensing image recognition, we use U-Net, PSP-Net and the method of the present invention as comparisons, and calculate the evaluation indicators of the three methods on the test set, as shown in Table 1 below:

[0106] Table 1 Comparison of model accuracy on multi-scale complex background test set

[0107]

[0108] In the table, Mean IoU is the average intersection-over-union ratio. From Table 1, we can see that the accuracy of landslide detection in complex and multi-scale backgrounds of the proposed method is better than that of the comparison method, especially in small-scale landslides, the Mean IoU is improved by 20.6% compared with that of U-Net, and the Mean IoU of landslides in complex backgrounds is improved by 2.08% compared with that of U-Net, with an average improvement of 9.91%, indicating the effectiveness of the proposed method in landslide detection in multi-scale and complex background areas. It also shows that the landslide detection method based on multi-model fusion proposed in this paper has obvious advantages over a single model in landslide detection.

[0109] In addition, in the process of testing with the three methods, we compared the part from input to semantic segmentation result map of remote sensing images, such as Figure 7a-7d As shown. Among them, Figure 7a-7d These are two comparison pictures of large-scale landslides, medium-scale landslides, small-scale landslides and landslides in complex background areas. Among them, the label is the most accurate landslide label for manual marking of landslide areas.

[0110] from Figure 7a-7d It can be seen that the prediction results of the U-Net model in various backgrounds have a certain area of ​​false detection, especially in the small-scale and complex backgrounds, the prediction results have a large area of ​​false detection; PSP-Net also has only a small area of ​​false detection; the prediction results of the proposed method in multi-scale and complex backgrounds are closer to the true value. The reason is that after the improved YOLOv4 model detects the landslide target, it expands the detection frame, intercepts the content in the detection frame, and fills the area outside the detection frame with black, which can reduce some background interference. On this basis, combined with the improved U-Net for semantic segmentation tasks, the prediction effect in multi-scale and complex backgrounds can be effectively improved.

[0111] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A landslide detection and early warning method based on multi-model fusion, characterized by: The steps include: (1) Obtain a remote sensing image of the area to be measured and cut it into multiple sub-remote sensing images, each of which corresponds to a sub-area within the area to be measured; (2) Classifying and identifying the sub-remote sensing images in turn as landslide images or non-landslide images, and only retaining the landslide images; (3) outputting a detection frame for each landslide image, including steps (31)-(32); (31) Use the target detection model to detect the landslide area in the landslide image; (32) If the target is detected, output the detection frame; if the target is not detected, adjust the input size of the landslide image or the parameters of the target detection model until the target is detected; (4) Enlarging the size of the detection frame and then filling the area outside the detection frame in the landslide image with black; (5) performing semantic segmentation on the landslide image obtained in step (4) to obtain a semantic segmentation result map; (6) Mark the semantic segmentation result map obtained in step (5) as a reference map; (7) Monitor the sub-areas of each landslide image, one of which is monitored by; (71) obtaining the remote sensing image of the sub-region corresponding to the landslide image again, processing it according to steps (3)-(5), and obtaining a semantic segmentation result image as a comparison image; (72) Subtracting the comparison image from the reference image pixel by pixel to obtain a deformation image containing only the deformation area, and calculating the area value S of the deformation area; (73) comparing the area value with a preset threshold; If S < threshold, repeat steps (71)-(72); If S ≥ threshold, an alarm is issued, and the current comparison image is marked as a reference image, and steps (71)-(72) are repeated.

2. The landslide detection and early warning method based on multi-model fusion according to claim 1 is characterized by: The step (2) performs classification and recognition through a classification model, and the classification model is a Swin Transformer classification model, a ResNet classification model or a VGG classification model.

3. The landslide detection and early warning method based on multi-model fusion according to claim 1 is characterized by: In the step (31), the target detection model is a YOLO target detection model or an improved target detection model, and the improved target detection model includes a YOLO4 model, and the YOLO4 model is composed of a backbone network, an SPP network and a PANet, wherein the backbone network adopts a Desenet121 network.

4. The landslide detection and early warning method based on multi-model fusion according to claim 1 is characterized by: In the step (32), the parameters of the target detection model include the model weight, the target score confidence, the threshold of non-maximum suppression, the alpha coefficient of the Desenet121 network, the probability of mosaic data enhancement, and / or the learning rate of the model.

5. The landslide detection and early warning method based on multi-model fusion according to claim 1 is characterized by: In the step (4), the size of the detection frame is enlarged as follows: The coordinates of the i-th vertex of the expanded detection box are calculated by the following formula: Where i = 1 to 4, representing the i-th vertex of the four vertices of the detection box, X′ i and Y i ′ are the horizontal and vertical coordinates of the i-th vertex of the enlarged detection box, respectively, i and Y i To expand the horizontal and vertical coordinates of the i-th vertex of the detection box, is the landslide area ratio, S L is the area of ​​the landslide region in the landslide image, S I is the area of ​​the landslide image; w is the zoom factor, and Input is the size of the sub-remote sensing image.

6. The landslide detection and early warning method based on multi-model fusion according to claim 1 is characterized by: In the step (5), semantic segmentation is performed by an improved semantic segmentation model, wherein the improved semantic segmentation model includes a Unet network, wherein the Unet network includes an encoding part and a decoding part, the encoding part adopts a resnet50 network, and a CBAM attention mechanism module and a PPM pyramid pooling module are sequentially set at the output end of the resnet50 network; The resnet50 network includes 5 residual blocks from top to bottom, and corresponds to 5 feature outputs feat1-feat5. The feat5 is sent to the decoding part after passing through the CBAM attention mechanism module and the PPM pyramid pooling module; The loss function FL(P t ) is calculated by the following formula: FL(P t )=-(1-P t ) γ log(P t ) Where γ is the modulation coefficient, 0≤γ≤5, P t is the probability function, p is the probability of a pixel point in the landslide image being considered as a landslide; 7. The landslide detection and early warning method based on multi-model fusion according to claim 6 is characterized by: The input size of the resnet50 network is [512, 512, 3], and the sizes of feat1-feat5 are [256, 256, 64], [128, 128, 256], [64, 64, 512], [32, 32, 1024], and [16, 16, 2048], respectively.

Images