Method for evaluating landslide susceptibility by combining remote sensing images and crowd-sourced data

Abstract

The application discloses a kind of remote sensing image combined with crowd source data to evaluate the method of landslide susceptibility, comprising the following steps: division slope unit, collect the spatial distribution data of landslide disaster-causing factor and carry out pretreatment, according to landslide disaster-causing factor is handled with state classification, with independence test screening;From crowd source data extraction landslide disaster point and carry out standardization processing, generate first class landslide training sample;YOLOv4 algorithm of light weight is used to extract the landslide disaster point of remote sensing image, generate second class landslide training sample;First class landslide training sample and second class landslide training sample are integrated and cleaned to obtain third class landslide training sample, the above-mentioned three kinds of landslide training sample is respectively through random forest classification method and carries out landslide susceptibility evaluation, to evaluate the danger of regional landslide.The remote sensing image combined with crowd source data to evaluate the method of landslide susceptibility of the application can generate landslide sample, to evaluate the danger of regional landslide quickly and accurately.

Description

Technical Field

[0001] This invention relates to a method for assessing landslide hazard risk, specifically, a method for assessing landslide susceptibility using remote sensing imagery combined with crowdsourced data. Background Technology

[0002] Landslide geological disasters are characterized by their wide distribution, high frequency of occurrence, strong concealment, and great destructiveness. They often cause catastrophic consequences and pose a huge threat to people's lives and property safety and social and economic development.

[0003] Landslide hazard point data is the fundamental reference data for landslide susceptibility assessment, and obtaining detailed landslide hazard point data is of great significance. Therefore, establishing a landslide sample database and efficiently and scientifically evaluating landslide susceptibility is of great importance for regional transportation planning and construction, as well as disaster prevention and mitigation.

[0004] Crowdsourced data, as a new type of generalized geographic big data, is primarily studied and applied in fields such as humanities, economics, and disaster emergency response. Its application in landslide susceptibility assessment is still in its early stages. Landslides and their harm to human life and property are caused by a variety of factors. Using remote sensing, geographic information systems, geology, meteorology, and other methods to obtain landslide-related information, and leveraging crowdsourced data fusion technology to study the uncertainty of landslide movement, scientifically describing the internal and external factors, and exploring a systematic theory and methodology for landslide risk analysis that includes both hazard and vulnerability aspects will help reduce potential losses from landslides. Crowdsourced data has a wide range of sources and acquisition methods, high timeliness, and rapid data updates, allowing for timely acquisition of disaster information. However, differences in crowdsourced data standards, the reliability and breadth of data sources, and data management lead to varying data quality standards. Furthermore, the collected landslide disaster data is often distributed along roads and rivers, necessitating consideration of the impact of landslide sample generation strategies on landslide sensitivity.

[0005] In view of this, there is a need to provide a method for assessing landslide susceptibility by combining remote sensing imagery with crowdsourced data, in order to solve or overcome the above-mentioned technical problems. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method for assessing landslide susceptibility by combining remote sensing imagery with crowdsourced data. This method can generate landslide samples, thereby quickly and accurately assessing the risk of landslides in a region.

[0007] To address the aforementioned technical problems, this invention provides a method for assessing landslide susceptibility using remote sensing images and crowdsourced data, comprising the following steps: (1) dividing the slope into units, collecting spatial distribution data of landslide-causing factors for preprocessing, and performing state classification processing based on the landslide-causing factors for selection using independence testing; (2) extracting landslide disaster points from the crowdsourced data and performing standardization processing to generate a first type of landslide training sample; (3) using the lightweight YOLOv4 algorithm to extract the landslide disaster points from the remote sensing images to generate a second type of landslide training sample; (4) integrating and cleaning the first type of landslide training sample and the second type of landslide training sample to obtain a third type of landslide training sample, and evaluating the landslide susceptibility of the first type of landslide training sample, the second type of landslide training sample, and the third type of landslide training sample using a random forest classification method to assess the risk of landslides in the region.

[0008] Preferably, the slope unit is generated by the automatic sliding slope unit program r slopeunits developed based on the GRASS GIS platform, which inputs elevation data and sets relevant input parameters.

[0009] Preferably, the spatial distribution data of the landslide-causing factors include elevation, slope, aspect, profile curvature, topographic humidity index, lithology, distance from fault, average annual rainfall, distance from road, distance from river, land use type, soil type, and landform type.

[0010] Preferably, the crowdsourced data includes open-source landslide disaster site databases, literature repositories, news portals, and social media.

[0011] Preferably, the standardization process in step (2) includes: (21) assigning latitude and longitude coordinates to the obtained landslide point address information; (22) unifying the coordinates of the landslide disaster points from different sources into coordinates under the same spatial reference coordinate system; and (23) cleaning and deduplicating the landslide disaster points to generate the first type of landslide training samples.

[0012] Preferably, the remote sensing image is a Sentinel-2 image.

[0013] Preferably, the model structure of the YOLOv4 algorithm includes a backbone network, a neck network, and a head network. The neck network includes SPP, PAN, and FPN. The head network consists of two convolutional layers. The three branches of the neck network are output and multi-scale predictions are performed to output the prediction results.

[0014] Preferably, the backbone network is a MobileNetv3 network, which includes CBH and bneck convolutional modules. The CBH includes depthwise separable convolutional layers, batch normalization layers, and h-swish activation functions. When the MobileNetv3 network extracts features from the input image, it can obtain feature layers through continuous convolution. The feature layers are used to identify landslides and generate second-type landslide training samples.

[0015] Preferably, the h-swish activation function is calculated as follows:

[0016] ReLU(6) = min(max(0,x),6)

[0017]

[0018] Where ReLU represents the activation function, and h-swish represents the modified activation function.

[0019] Preferably, the vector data of the slope unit and the data of the landslide disaster point are superimposed. If a landslide disaster point falls into a certain slope unit, the slope unit is determined to be an unstable slope.

[0020] Compared with the prior art, the present invention has at least the following beneficial effects:

[0021] The method for assessing landslide susceptibility using remote sensing imagery and crowdsourced data of the present invention includes the following steps: First, slope units are divided, spatial distribution data of landslide-causing factors are collected and preprocessed, and state classification is performed based on the landslide-causing factors. The landslide-causing factor data is then screened after independence testing. Next, landslide hazard points are extracted from the crowdsourced data and standardized to generate a first-class landslide training sample, achieving near-zero-cost and highly timely generation of landslide training samples using crowdsourced data. Finally, a lightweight YOLOv4 algorithm is used to extract landslide hazard points from the remote sensing imagery to generate a second-class landslide training sample, enabling rapid acquisition of large-area landslide data. The first type of landslide training sample is used to avoid areas inaccessible to humans. Finally, the two types of samples are integrated and deduplicated to generate a third type of landslide training sample, creating a more comprehensive landslide database. The first, second, and third types of landslide training samples are then used to assess landslide susceptibility using a random forest classification algorithm. The third type of landslide training sample achieves higher accuracy in landslide susceptibility assessment. Crowdsourced data-assisted remote sensing imagery can quickly generate landslide training samples, establishing a landslide database and enabling large-scale landslide susceptibility assessment, providing a reference for regional transportation planning and land use planning.

[0022] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0023] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the following detailed description to explain the invention, but do not constitute a limitation thereof. In the drawings:

[0024] Figure 1 This is a flowchart of the steps in the method for assessing landslide susceptibility using remote sensing imagery combined with crowdsourced data according to the present invention.

[0025] Figure 2 This is a schematic diagram of rivers, roads, and elevations of the area studied in one embodiment of the method for assessing landslide susceptibility using remote sensing imagery and crowdsourced data according to the present invention.

[0026] Figure 3 This is a flowchart of the steps for obtaining training samples for a second type of landslide in one embodiment of the method for assessing landslide susceptibility using remote sensing imagery and crowdsourced data according to the present invention.

[0027] Figure 4a This is a distribution map of the first type of landslide training samples in one embodiment of the method for assessing landslide susceptibility using remote sensing imagery and crowdsourced data according to the present invention.

[0028] Figure 4b This is a distribution map of the second type of landslide training samples in one embodiment of the method for assessing landslide susceptibility using remote sensing imagery and crowdsourced data according to the present invention.

[0029] Figure 4c This is a distribution map of the third type of landslide training samples in one embodiment of the method for assessing landslide susceptibility using remote sensing imagery and crowdsourced data according to the present invention.

[0030] Figure 5a In one embodiment of the method for assessing landslide susceptibility using remote sensing imagery and crowdsourced data of the present invention, the landslide susceptibility distribution map of the first type of landslide training samples is obtained through a random forest classification algorithm;

[0031] Figure 5b In one embodiment of the method for assessing landslide susceptibility using remote sensing imagery and crowdsourced data of the present invention, the landslide susceptibility distribution map of the second type of landslide training samples is obtained through a random forest classification algorithm;

[0032] Figure 5c In one embodiment of the method for assessing landslide susceptibility using remote sensing imagery and crowdsourced data of the present invention, a landslide susceptibility distribution map is obtained by using a random forest classification algorithm for the third type of landslide training samples.

[0033] Figure 6This is an embodiment of the method for assessing landslide susceptibility using remote sensing imagery and crowdsourced data according to the present invention, showing the ROC curves for landslide susceptibility evaluation obtained from three types of landslide training samples.

[0034] Figure 7 This is a landslide susceptibility statistical map obtained from three types of landslide training samples in one embodiment of the method for assessing landslide susceptibility using remote sensing imagery and crowdsourced data according to the present invention. Detailed Implementation

[0035] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0036] Reference Figure 1 This invention provides a method for assessing landslide susceptibility using remote sensing images and crowdsourced data, comprising the following steps: (1) dividing the slope into units, collecting spatial distribution data of landslide-causing factors for preprocessing, performing state classification processing based on the landslide-causing factors, and screening by independence test; (2) extracting landslide disaster points from the crowdsourced data and performing standardization processing to generate a first type of landslide training sample; (3) using the lightweight YOLOv4 algorithm to extract landslide disaster points from the remote sensing images to generate a second type of landslide training sample; (4) integrating and cleaning the first and second types of landslide training samples to obtain a third type of landslide training sample, and evaluating the landslide susceptibility of the first, second, and third types of landslide training samples respectively using the random forest classification method to assess the risk of landslides in the region.

[0037] The above-described basic embodiments of the present invention provide a method for assessing landslide susceptibility using remote sensing imagery combined with crowdsourced data. First, slope units are divided, and spatial distribution data of landslide-causing factors are collected and preprocessed. State classification is performed based on the landslide-causing factors, and the data is screened after independence testing. Then, landslide hazard points are extracted from the crowdsourced data and standardized to generate a first type of landslide training sample (first type sample), achieving near-zero cost and high timeliness in generating landslide training samples using crowdsourced data. Further, the lightweight target detection algorithm YOLOv4 is used to extract landslide hazard points from the remote sensing imagery to generate a second type of landslide training sample (second type sample). The system first generates a large-scale landslide training sample database by integrating and deduplicating the first and second types of landslide training samples to create a third type of landslide training sample. This creates a more comprehensive landslide database. The three types of samples are then used to assess landslide susceptibility using a random forest classification algorithm. The third type of landslide training sample achieves higher accuracy in landslide susceptibility assessment. Crowdsourced data-assisted remote sensing imagery can quickly generate landslide training samples, establishing a landslide database and enabling large-scale landslide susceptibility assessment. This provides a reference for regional transportation planning and land use planning.

[0038] Preferably, the automatic sliding slope unit program r slopeunits, developed based on the GRASS GIS platform, generates slope units by inputting elevation data and setting relevant input parameters.

[0039] Preferably, the spatial distribution data of landslide-causing factors include elevation, slope, aspect, profile curvature, topographic humidity index, lithology, distance from fault, average annual rainfall, distance from road, distance from river, land use type, soil type, and landform type.

[0040] Preferably, crowdsourced data includes open-source landslide hazard databases, literature repositories, news portals, and social media.

[0041] Preferably, the standardization process in step (2) includes: (21) assigning latitude and longitude coordinates to the obtained landslide point address information; (22) unifying the coordinates of landslide disaster points from different sources into coordinates under the same spatial reference coordinate system; and (23) cleaning and deduplicating the landslide disaster points to generate the first type of landslide training samples.

[0042] Preferably, the remote sensing image is a Sentinel-2 image.

[0043] Preferably, the YOLOv4 algorithm model structure includes a backbone network, a neck, and a head. The neck includes spatial pyramid pooling (SPP), path aggregation network (PAN), and feature pyramid network (FPN) structures. The head consists of two convolutional layers, outputting the three branches of the neck and performing multi-scale predictions, and outputting the prediction results.

[0044] Preferably, the backbone network is a MobileNetv3 network, which includes CBH (Conv2D-BN-hswish) and bneck (bottleneck) convolutional modules. CBH includes depthwise separable convolutional layers, batch normalization layers, and h-swish activation functions. When the MobileNetv3 network extracts features from the input image, it can obtain feature layers through continuous convolution. The feature layers are used to identify landslides and generate training samples for the second type of landslide.

[0045] Preferably, the h-swish activation function is calculated as follows:

[0046] ReLU(6) = min(max(0,x),6)

[0047]

[0048] Where ReLU represents the activation function, and h-swish represents the modified activation function.

[0049] Preferably, the vector data of the slope unit and the data of the landslide disaster point are superimposed. If a landslide disaster point falls into a certain slope unit, the slope unit is determined to be an unstable slope.

[0050] The following example illustrates the specific operation process of the method for assessing landslide susceptibility using remote sensing imagery combined with crowdsourced data according to the present invention, in order to better understand the technical solution of the present invention. The specific operation process is as follows:

[0051] It should be noted that this specific embodiment takes Milin County, Nyingchi Prefecture, Tibet Autonomous Region as the research area. See [link / reference]. Figure 2 A schematic diagram showing the rivers, roads, and elevation locations of the study area.

[0052] Step (1): Divide the slope units, which is based on the automatic sliding slope unit program rslopeunits developed on the GRASS GIS platform. Input the elevation data and set the relevant input parameters; The collected spatial distribution data of landslide disaster factors include elevation, slope, aspect, profile curvature, topographic humidity index, lithology, distance from fault, annual average rainfall, distance from road, distance from river, land use type, soil type, and landform type. Convert all factor data into raster data to facilitate data editing and calculation. Further discretize the continuous data and classify the state; In addition, the spatial reference of the collected factor data may not be uniform, so it is necessary to unify the spatial reference of each factor data. The geographic coordinate system used in this embodiment is the WGS84 coordinate system; Using the Pearson correlation coefficient, remove factors in the factor pairs with high correlation. Calculate the Pearson correlation coefficient value using the following formula:

[0053]

[0054] Where X and Y represent different related factors. When |r|>0.5, it means that the factors are strongly correlated and one of them needs to be removed. In the specific embodiment of the present invention, 10 factors are finally retained for landslide susceptibility evaluation: distance from road, land use type, average annual rainfall, distance from river, distance from fault, lithology, elevation, slope, aspect and landform type.

[0055] Step (2): Crowdsourced data includes open-source landslide disaster point databases, literature repositories, news portals, and social media. These diverse crowdsourced data are standardized to become usable data. The specific steps for standardization are: a) Data collected from various platforms is in the form of data tables or web text. The data tables or text information are linked to remote sensing images. The text information needs to be interpreted into disaster points with attribute and spatial information on a Geographic Information System (GIS). The obtained landslide point address information is assigned latitude and longitude coordinates; b) The coordinates of landslide disaster points from different sources are unified to the same spatial reference coordinate system (WGS84 coordinate system); c) The landslide disaster point data is cleaned and deduplicated. First, the "deduplication" tool in ArcGIS software is used to delete overlapping landslide disaster points. Since the effective units retained for slope units are all larger than 300m × 300m, the Euclidean distance between two points is calculated:

[0056]

[0057] For distances less than The landslide hazard points within the minimum slope unit (maximum distance) are cleaned, and only one valid landslide hazard point is retained within that distance; the processed landslide hazard points are considered as the first type of landslide training samples, such as... Figure 4a As shown.

[0058] Step (3): Use the lightweight YOLOv4 algorithm to extract landslide disaster points from remote sensing images and generate a second type of landslide training samples. The specific steps are as follows: a) Construct a public landslide dataset. Use Labelimg software to generate a label file corresponding to each landslide sample. The dataset for training the model is the Bijie City landslide dataset, which can be obtained from http: / / study.rsgis.whu.edu.cn / pages / downlo a) Landslide images were obtained publicly from the website and cropped out to a total of 770 images; b) Data augmentation was used to process the landslide sample data and corresponding label data. Data augmentation mainly involved combining enhancement processes such as brightening, darkening, blurring, adding noise, horizontal flipping, vertical flipping, and rotation on each image, expanding the 770 landslide samples and label data to 5770; c) A lightweight YOLOv4 model was built and trained. The lightweight YOLOv4 model uses MobileNetv3 to replace the original CSPDarkNet53 network as the feature extraction layer of the landslide recognition and detection network. The MobileNetv3 network mainly consists of CBH (Con v2D-BN-hswish) and bneck (bottleneck) convolutional modules, with the specific structure as follows: Figure 3 As shown, CBH consists of depthwise separable convolutional layers, batch normalization layers, and the h-swish activation function. The calculation formula for the h-swish activation function is as follows:

[0059] ReLU(6) = min(max(0,x),6)

[0060]

[0061] When extracting features from an input image, the MobileNetv3 network, given an input size of 416×416, can obtain feature layers of 52×52, 26×26, and 13×13 through continuous convolutions for landslide identification. Model precision, recall, and mAP are used to evaluate the theoretical performance of the model. Precision represents the ratio of correctly extracted landslides to the total number of correctly extracted landslides; recall represents the ratio of correctly extracted landslides to the total number of landslides in the dataset; and mAP represents the integral of precision and recall over the interval [0,1]. Higher mAP indicates better model performance. Table 1 shows the metrics of the pre-trained model.

[0062] Table 1

[0063]

[0064] Images obtained by fusing Sentinel-2 images from Milin County from May to August 2021 were input into the training model. The experimental results were imported into Google Earth Pro for visual interpretation and verification. 168 landslide hazard points were correctly extracted, and the extraction accuracy is shown in Table 2. The correctly extracted landslide samples were used as landslide samples generated from the remote sensing images, and these samples were also used as training samples for the second type of landslides. Figure 4b As shown.

[0065] Table 2

[0066]

[0067] Step (4): Integrate and deduplicate the first type of landslide training samples and the second type of training samples; merge the first type of landslide training samples and the second type of landslide training samples; use the "duplicate" tool in ArcGIS to delete overlapping landslide disaster points; treat the integrated and deduplicated landslide samples as the third type of landslide training samples, such as... Figure 4c As shown in Figure 5, the landslide susceptibility distribution maps were obtained by classifying the three types of landslide training samples using the random forest algorithm. Figure 5a It is a landslide susceptibility map generated using training samples from the first type of landslide. Figure 5b It is a landslide susceptibility map generated using training samples from the second type of landslides. Figure 5c It is a landslide susceptibility map generated using training samples of the third type of landslide.

[0068] The ROC curves of the three types of landslide training samples obtained by the random forest classification model are as follows: Figure 6 As shown, the AUC values ​​were 0.9202, 0.9187, and 0.9466, respectively. The AUC value of the third type of landslide training samples participating in the RF model (random forest model) was higher than that of the first and second types of landslide training samples participating in the RF model. This indicates that the landslide susceptibility mapping accuracy is better when using the proposed remote sensing imagery combined with crowdsourced data to generate landslide training samples. This landslide sample generation strategy can quickly generate landslide training samples for large areas, establish a landslide database, realize the assessment of landslide susceptibility in large areas, and provide a reference for regional transportation planning and construction and land use planning.

[0069] The susceptibility map of the study area is divided into five levels: extremely low susceptibility area, low susceptibility area, medium susceptibility area, high susceptibility area, and extremely high susceptibility area. Figure 7The distribution of each susceptibility level in the susceptibility zoning maps predicted by three sample generation strategies was statistically analyzed. In the prediction results of the first type of sample, the proportions of extremely low susceptibility areas, low susceptibility areas, medium susceptibility areas, high susceptibility areas, and extremely high susceptibility areas were 27.70%, 24.54%, 19.87%, 14.37%, and 13.52%, respectively. In the prediction results of the second type of sample, the proportions of extremely low susceptibility areas, low susceptibility areas, medium susceptibility areas, high susceptibility areas, and extremely high susceptibility areas were 32.76%, 25.27%, 16.92%, and 14.80%, respectively. The predicted landslide susceptibility rates for the three landslide sample generation strategies were as follows: 10.25% for extremely low susceptibility areas, 16.96% for low susceptibility areas, 14.96% for medium susceptibility areas, 10.38% for high susceptibility areas, and 8.88% for extremely high susceptibility areas. This indicates that the extremely low susceptibility areas were the most prevalent among the three landslide sample generation strategies, while the extremely high susceptibility areas were the least prevalent. This suggests that the prediction results of the three landslide sample generation strategies have a certain degree of reliability. Furthermore, the proportions of extremely low and low susceptibility areas predicted by the third type of sample were greater than those predicted by the first and second types of samples, while the proportions of extremely high and high susceptibility areas predicted by the third type of sample were also less than those predicted by the first and second types of samples. Therefore, this invention's method for generating landslide samples using remote sensing imagery combined with crowdsourced data to assess landslide susceptibility has higher sensitivity and is more suitable for landslide susceptibility mapping, providing valuable reference for large-scale regional transportation planning and land use planning.

[0070] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present invention, various simple modifications can be made to the technical solution of the present invention, and these simple modifications all fall within the protection scope of the present invention.

[0071] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the present invention will not describe the various possible combinations separately.

[0072] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed by the present invention.

Claims

1. A method for assessing landslide susceptibility using remote sensing imagery combined with crowdsourced data, characterized in that, Includes the following steps: (1) Divide the slope into units, collect spatial distribution data of landslide disaster-causing factors for preprocessing, and perform state classification processing based on the landslide disaster-causing factors for screening by independence test; (2) Extract landslide disaster points from crowdsourced data and perform standardization processing to generate the first type of landslide training samples; The standardization process includes the following steps: (21) Assign latitude and longitude coordinates to the obtained address information of the landslide disaster point; (22) Unify the coordinates of the landslide disaster points from different sources into coordinates under the same spatial reference coordinate system; (23) Clean and deduplicate the landslide disaster points to generate the first type of landslide training samples; (3) Use the lightweight YOLOv4 algorithm to extract the landslide disaster points from the remote sensing images and generate the second type of landslide training samples; (4) The first type of landslide training sample and the second type of landslide training sample are integrated and cleaned to obtain the third type of landslide training sample. The first type of landslide training sample, the second type of landslide training sample and the third type of landslide training sample are respectively evaluated for landslide susceptibility using the random forest classification method to assess the risk of landslides in the region. The crowdsourced data includes open-source databases of landslide disaster sites, literature repositories, news portals, and social media.

2. The method for assessing landslide susceptibility using remote sensing imagery and crowdsourced data according to claim 1, characterized in that, The automatic sliding slope unit program r slopeunits, developed based on the GRASS GIS platform, generates the slope unit by inputting elevation data and setting relevant input parameters.

3. The method for assessing landslide susceptibility using remote sensing imagery combined with crowdsourced data according to claim 1, characterized in that, The spatial distribution data of the landslide-causing factors include elevation, slope, aspect, profile curvature, topographic humidity index, lithology, distance from fault, average annual rainfall, distance from road, distance from river, land use type, soil type, and landform type.

4. The method for assessing landslide susceptibility using remote sensing imagery combined with crowdsourced data according to claim 1, characterized in that, The remote sensing image is a Sentinel-2 image.

5. The method for assessing landslide susceptibility using remote sensing imagery combined with crowdsourced data according to claim 1, characterized in that, The YOLOv4 algorithm's model structure includes a backbone network, a neck network, and a head network. The neck network includes SPP, PAN, and FPN. The head network consists of two convolutional layers. The algorithm outputs the three branches of the neck network and performs multi-scale predictions, outputting the prediction results.

6. The method for assessing landslide susceptibility using remote sensing imagery and crowdsourced data according to claim 5, characterized in that, The backbone network is a MobileNetv3 network, which includes CBH and bneck convolutional modules. The CBH includes depthwise separable convolutional layers, batch normalization layers, and h-swish activation functions. When the MobileNetv3 network extracts features from the input image, it can obtain feature layers through continuous convolution. These feature layers are used to identify landslides and generate training samples for the second type of landslide.

7. The method for assessing landslide susceptibility using remote sensing imagery combined with crowdsourced data according to claim 6, characterized in that, The h-swish activation function is calculated as follows: （1） （2） Where ReLU represents the activation function, and h-swish represents the modified activation function.

8. The method for assessing landslide susceptibility using remote sensing imagery combined with crowdsourced data according to any one of claims 1 to 7, characterized in that, The vector data of the slope unit and the data of the landslide disaster point are superimposed. If a landslide disaster point falls into a certain slope unit, the slope unit is determined to be an unstable slope.

Images