Method, device and system for identifying and predicting landslides

Abstract

Identify and predict landslides without each data point sharing landslide data. In the method, the second device performs multi-modal model fusion of multiple single-modal landslide identification and prediction models obtained from multiple first devices, trains the fused base model based on a multi-modal landslide dataset, obtains a landslide identification and prediction base model, and in response to a model customization request sent from the third device, clips the landslide identification and prediction base model, obtains a landslide identification and prediction customized model, and sends it to the third device. The third device locally trains the landslide identification and prediction customized model based on a local landslide dataset, obtains a target landslide identification and prediction model, uses the model to process landslide data in an area where the third device is located, obtains a landslide identification and prediction result, and the landslide identification and prediction result indicates the occurrence and type of landslide.

Description

[Technical field]

[0001] The present disclosure is in the field of computer systems and data processing based on a particular model, In particular, it relates to a method, apparatus and system for identifying and predicting landslides. [Background technology]

[0002] In landslide geological disasters, which are characterized by a large number of landslide points, a large surface area, and a high degree of danger, ,Landslides pose a major threat to human life safety and public infrastructure. Currently, landslide monitoring devices are used due to the complex geological and climatic conditions in which landslides occur. Each has its own advantages and limitations, and there are large differences in landslide data from region to region, so Different regions use different landslide identification and prediction methods and different landslide identification and prediction models. This requires that the company must make necessary changes to the way it operates, resulting in a waste of resources and some uncertainty. In some scenarios, the server aggregates landslide data from different regions and The landslide identification and prediction model is trained in a unified manner based on the data. To ensure the safety of the area, the region must agree to share landslide data. Landslide identification and prediction is difficult to achieve. Summary of the Invention

[0003] In response to the problems existing in the above-mentioned prior art, the present application provides a method and apparatus for identifying and predicting landslides. and systems, and provide a common base model for a wide area (including multiple regions) Design and further user end based landslide identification and prediction for a certain area The technical solution can be based on the model to customize the model. This allows landslide identification and prediction to be realized without sharing data, saving resources and improving landslide identification and prediction. The accuracy of predictions can be improved. In a first aspect, the present application provides a method for identifying and predicting landslides, the method comprising: Landslide identification and prediction model using multiple single-modes obtained from multiple first-stage instruments using two-stage instruments Based on the multi-mode model fusion, we performed a multi-mode landslide dataset. The multi-layered fusion model is trained to obtain a landslide identification and prediction base model. Modal model fusion is a method to fusion the feature maps of multiple single-mode landslide identification and prediction models. To stitch and process the stitched feature maps, we use a polynomial neural network. We construct a network and attention mechanism, where each single-mode landslide is identified by The prediction model was trained based on the single-mode landslide data set by each first device. The types of landslide data in the single-mode landslide data set of one first device are the same. and the second device, in response to a model customization request transmitted from a third device, said landslide identification and prediction based on said third device local landslide data type; The model is clipped and a customized model for landslide identification and prediction is obtained. Submitting said landslide identification and prediction customization model, said model customization request The third device includes a type of local landslide data, and the third device includes a type of local landslide data. Localize the landslide identification and prediction customized model based on the landslide data set The third device performs the above-mentioned process to obtain a target landslide identification and prediction model. The measurement model is used to process the landslide data in the area where the third device is located, and to estimate the landslide obtaining a landslide identification and prediction result; and determining whether a landslide has occurred and Indicates the type of slip. The second device is a device for detecting and predicting a plurality of single-mode landslide identification and prediction models obtained from a plurality of first devices. The multi-mode model fusion of the device includes: The feature maps of the slip identification and prediction model are stitched, and the stitched feature maps are We build a polynomial neural network and an attention mechanism to process the Here, the attention mechanism is a channel attention mechanism and / or or spatial attention mechanisms. In a possible embodiment, the second device comprises a plurality of unimodal landslide identification and prediction models Before stitching the feature maps, the landslide identification method provided by the present application The method of predicting further includes: a second device for detecting a plurality of single-mode landslides; Among the feature maps of the model, feature maps whose similarity exceeds a certain level are merged or overlapped. Eliminate and process. In a possible embodiment, the second device receives the model customization transmitted from the third device. Landslide identification and prediction in response to requests. Clipping of base model for landslide identification and prediction. Obtaining the predictive customization model includes: In response to the model customization request, the feature maps of the landslide identification base model are Customize landslide identification and prediction by clipping feature maps other than the target feature map A model is obtained, where the target feature map is the landslide data in the model customization request. The feature maps of one or more single-mode landslide identification and prediction models corresponding to the type of landslide are do. In a possible embodiment, the method for identifying and predicting landslides provided by the present application further comprises: The second device further includes receiving the plurality of updated single mode locations from the plurality of first devices. Receive slip identification and prediction models. In a second aspect, the present application provides a second device comprising the following modules: The module uses multiple single-mode landslide identification and prediction models obtained from multiple first-stage devices. are used for multi-mode model fusion, where each single mode landslide identification and prediction The measurement model is trained based on the single-mode landslide data set by each first device. The types of landslide data in the single-mode landslide data set of one first device are the same. The model training module is based on the multi-modal landslide dataset. The base model was then trained to obtain a landslide identification and prediction base model. The model clipping module responds to a model customization request sent from a third device. In response, the third device based on the type of local landslide data, landslide identification and prediction The model is then clipped and used to obtain a customized model for landslide identification and prediction. The transmitter / receiver module then transmits the customized landslide identification and prediction model to the third device. The model customization request is used to obtain the local landslide data of the third device. Types included. The model fusion module specifically integrates multiple single-mode landslide identification and prediction models. Stitching feature maps and applying polynomial neural networks and attention mechanisms The polynomial neural network and the attention mechanism are used to construct The algorithm is used to process the stitched feature map, where attention The attention mechanism is a channel attention mechanism and / or a spatial attention mechanism. Includes ism. In a possible embodiment, the second device further comprises a pre-treatment module, The rule is a set of feature maps of multiple single-mode landslide identification models that have a predetermined similarity level. It is used to merge or deduplicate more than one feature map. In a possible embodiment, the model clipping module specifically receives the model from a third device. In response to the model customization request sent by The feature maps other than the target feature map are clipped from the group, and the landslide identification and prediction customization is performed. The target feature map is used to obtain the model customization requirements. One or more unimodal landslide identification and prediction methods that correspond to the type of landslide data being sought This is a feature map of the model. In a possible embodiment, the second device receives a plurality of updated single mode signals from a plurality of first devices. and a transceiver module for receiving the landslide identification and prediction model. . In a third aspect, the present application provides a third device comprising the following modules: a transmitting and receiving module; The module sends a model customization request to the second device, and the second device receives the landslide identification and prediction data. Used to receive measurement customization models, where the model customization request is The third device includes a variety of local landslide data, and can be customized to identify and predict landslides. The model was obtained by clipping the landslide identification base model using the second device. The landslide identification base model is based on a plurality of the first devices received by the second device. A number of single-mode landslide identification and prediction models were obtained by multi-mode model fusion, and model training was performed. The module detects landslides received from the second device based on the local landslide data set. Localized training of customized landslide identification and prediction models to identify and predict landslides The prediction module employs the target landslide identification and prediction model, 3. To process the landslide data in the area where the device is located and obtain the landslide identification and prediction results The landslide identification and prediction results are used to determine whether a landslide has occurred and the type of landslide. Give instructions. In a possible embodiment of the first to third aspects, the type of landslide data is optical image data. data, synthetic aperture radar SAR image data, laser radar point cloud data or sensor The different types of data include at least one of the following monitoring data: This is also called data. In a possible embodiment of the first to third aspects, the landslide data further comprises geolocation information. Based on this, the above landslide identification and prediction results can be used to further indicate the location of the landslide point. Show. In a possible embodiment of the first to third aspects, the second device obtains the signal from the first device. The single-mode landslide identification and prediction model uses the model parameters of the single-mode landslide model. By transmitting model parameters, data transmission can be saved and efficiency can be improved. It is possible to do so. In a possible embodiment of the first to third aspects, the single-mode landslide identification and prediction model The model parameters are feature maps of the unimodal landslide identification and prediction model. In a possible embodiment of the first to third aspects, the plurality of updated single mode landslides are Prediction accuracy of the new identification and prediction model and prediction of the single-mode landslide identification and prediction model before updating The difference in accuracy is greater than a predetermined difference. In this way, the number of interactions between the first device and the second device is reduced. However, due to the difference in the calculation capacity of each first device, and the difference in landslide data and models, The time overhead required to complete training of the single-mode landslide identification and prediction model is different. This can reduce the time required to update the landslide identification and prediction base model. To some extent, it reduces the time overhead and improves the accuracy of the landslide identification and prediction base model. It is possible to do so. In a possible embodiment of the first to third aspects, the first device is The training method is to train a single-mode landslide identification and prediction model based on the set of supervised The first device end is a single-mode landslide identification and prediction model. To improve the training accuracy of the model, we combine it with self-supervised training to train the model without labels. This allows the first device to be independent of the second device to provide the label. The first device can independently complete the training task of a single-mode landslide identification and prediction model. This reduces the interaction between the first and second devices and saves time overhead. This can be done. In a fourth aspect, the present application provides a method for detecting a plurality of processors, the method comprising: a memory communicatively connected to a processor, wherein the memory includes instructions executable by at least one processor, the instructions including at least When executed by one or more processors, the at least one processor performs the first aspect and and performing the steps performed by the second device of any possible embodiment, or Executing the steps performed by the third device in one embodiment and any possible implementations It is possible. In a fifth aspect, the present application provides a method for producing a computer-implemented method for producing a computer-implemented method, the method comprising: to cause a computer to carry out the method according to the first aspect and any possible embodiment thereof A computer readable storage medium for use is provided. In a sixth aspect, the present application provides a computer program comprising computer program instructions. The computer program instructions executed on the computer provide a software product that The method according to one aspect and any possible implementation is carried out by a computer. In a seventh aspect, the present application relates to a landslide comprising a plurality of first devices, a second device and a third device. The present invention provides an identification and prediction system, each of which comprises an apparatus according to the first aspect and any possible embodiment thereof. Specifically, each of the first devices in the plurality of first devices is a single A single-mode landslide identification and prediction model is obtained based on the single-mode landslide data set. and transmitting the single-mode landslide identification and prediction model to the second device. The second device is used to identify and predict landslides based on multiple single-mode landslide identification and prediction models. The second and third devices are then used to obtain an identification and prediction-based model. , model customization, model clipping, model localization training, and landslide synchronization Complete configuration and forecasting. The embodiments of the present application provide a method, an apparatus and a system for identifying and predicting landslides, In order to realize landslide identification and prediction, it is necessary to predict landslides over a wide area and collect data from multiple information sources. We propose a landslide identification and prediction-based model that supports heterogeneous monitoring data from The second device uses the single-mode landslide identification and prediction model transmitted from multiple first devices. The base model obtained by multi-mode model fusion is a landslide identification and prediction model for different regions. The model can be used as the basis for the development of dedicated landslide identification and prediction models for different regions. This saves resources to obtain a base model for landslide identification and prediction without the need to separately train a new model. In addition, the second device can be customized to meet the customization needs of the third device. The third device may clip the base model based on local landslide data. The received model can be locally trained and used for landslide identification and prediction. The technical solution provided by the embodiments of the present application is that each data end shares landslide data. A common landslide identification and prediction model is obtained without using the existing landslide identification and prediction methods. It can save resources, and the fused base model can also be used to detect landslides with different modes. Capture complex interactions between data and key features for landslide identification and prediction This can strengthen the expressiveness of the model, improve the performance, and improve the accuracy of landslide identification and prediction. can be improved. [Brief description of the drawings]

[0004] [Figure 1] 1 is a flowchart of a method for training a single-mode landslide identification and prediction model provided by an embodiment of the present application. [Diagram 2] 1 is a flowchart of a method for identifying and predicting landslides provided by an embodiment of the present application. [Diagram 3] 2 is a flowchart of a method for identifying and predicting landslides provided by an embodiment of the present application. [Figure 4] FIG. 2 is a schematic diagram of a multi-modal model fusion process provided by an embodiment of the present application. [Diagram 5] 3 is a flowchart of a method for identifying and predicting landslides provided by an embodiment of the present application. [Figure 6] FIG. 2 is a schematic diagram showing the architecture of the interaction of each device in the method for identifying and predicting landslides provided by an embodiment of the present application. [Figure 7] FIG. 2 is a schematic diagram showing the structure of a first device provided by an embodiment of the present application. [Figure 8] FIG. 2 is a schematic diagram showing the structure of a second device provided by an embodiment of the present application. [Figure 9] FIG. 2 is a schematic diagram showing the structure of a third device provided by an embodiment of the present application. [Figure 10] 1 is a structural schematic diagram of an electronic device provided by an embodiment of the present application; DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0005] As used herein, a "model" refers to a process that processes inputs and provides corresponding outputs. Typically, the input layer and the output layer, as well as one between the input layer and the output layer, are In deep learning applications, the model used is A model typically contains many hidden layers, which increases the depth of the network. The model may be a neural network, the layers of the neural network being: The input layer is connected in a sequential manner so that the output of the previous layer is provided as the input to the next layer. The output layer receives the input of the neural network, and the output of the output layer is the final output of the neural network. In this specification, the terms "neural network," "network," and "neural The terms "neural network model" and "model" are used interchangeably. The technical solutions provided by the embodiments of the present application are used in landslide identification and prediction, The slide data is input to the trained landslide identification and prediction model, and the landslide identification and prediction model The data is processed to obtain landslide identification and prediction results, and it is possible to determine whether a landslide has occurred in a certain area and You can find out the name and type of landslide. Here, landslide data refers to data on the geological conditions at a certain point collected by a sampling device. For example, landslide data may be collected from several image data or sensors at a certain location. may be monitoring data measured by another monitoring device. In addition, when classified according to the volume of the landslide body, landslide types are small, medium-sized, and large. Please understand that this includes landslides, large landslides and very large landslides. When classified according to the sliding speed, landslide types are classified into peristaltic landslides, slow landslides, The landslide velocity includes medium and high velocity landslides. The degree of material composition of the landslide body and the landslide velocity are When classified according to the relationship between the landslide and the geological structure, the types of landslides are: overburden landslides, bedrock landslides, Landslides and special landslides are included. Of course, landslide types are also classified by other factors. Other results may be possible, and the embodiment of the present application does not limit the type of landslide. In the field of identification and prediction of geological disasters (e.g. landslides), The artificial intelligence model is an artificial intelligence model with a centralized architecture. The server side of Kucha collects landslide data from different regions and uses the landslide data as training data. It is used as a database to train models based on training data on the server side and to identify and predict landslides. Obtain the model and use the landslide identification and prediction model to perform landslide identification and prediction. The raw data (i.e., landslide data for each region) used in the survey includes high-precision map data and address data. These raw data are usually collected by different departments and These raw data may be related to national security, so data owned by different departments may be subject to There is reluctance to share data, and increased regulatory compliance is limiting the free flow of raw data. As a result, this model training method is prone to data privacy leaks. and server performance bottlenecks, and the security of landslide monitoring data The need to process landslide monitoring data that is continuously collected is not met. In light of these problems, the landslide monitoring data Predicting landslides while ensuring security and the absence of raw data is an urgent need. In addition, because the geological and climatic environments in which landslides occur are complex, it is necessary to collect landslide data for each region. Based on this, we propose a method for identifying and predicting landslides. In addition, different landslide identification and prediction methods and models were used for each study area. This leads to resource waste and uncertainty. There are also major challenges in generalizing slip identification and prediction models. In order to solve the above problems, the embodiments of the present application provide a method, an apparatus and a method for identifying and predicting landslides. To provide a system for landslide identification and prediction, the system is designed to cover a wide area. Landslide prediction and identification supporting heterogeneous monitoring data from multiple sources. We propose a prediction-based model (also called the Foundation model) The main content of the technical solution includes: the second device as a server side is connected to a plurality of first devices; The single-mode landslide identification and prediction model transmitted from the equipment is fused with a multi-mode model to Multi-modal model fusion involves polynomial-based stitching of feature maps. We introduce neural networks and attention mechanisms to develop a fused base model. The second device trains the landslide identification and prediction base model, and the second device uses the data sent from the third device. After receiving the model customization request, the landslide identification and prediction base model is The base model after clipping is transmitted to a third device, and the third device Localize training (or fine-tuning) of the received model based on local landslide data After the data is processed (also called training), it is used for landslide identification and prediction. The landslide identification and prediction database can be applied to a wide area without sharing landslide data among data centers. A base model can be obtained and the basis for landslide identification and prediction can be established according to local landslide data. We customized the landslide identification and prediction model based on the landslide model to perform landslide identification and prediction. Moreover, multiple single-mode landslide identification and prediction models can be implemented to save resources. The fused base model incorporates the complex interaction relationships between different landslide modes. The model can capture the relationship between the landslides and the important features for landslide identification and prediction. It can strengthen the current capacity, improve the performance, and increase the accuracy of landslide identification and prediction. Please note that the types of landslide data in different regions may differ, so the above multiple sources Heterogeneous monitoring data from can refer to landslide data that contain different data types. It is understood. In the embodiment of this application, the model used for landslide identification and prediction is already limited to a small area. Rather than a model that only supports landslide identification and prediction (a single-mode model), This is a base model that supports landslide identification and prediction in large areas (i.e., wide areas). The landslide identification method provided by the embodiments of the present application is applicable to a landslide identification and prediction system. The system adopts a distributed architecture of production-service-consumption mode, and multiple It includes several data ends, one server end, and one or more user ends. Here, The end is the producer, which collects single-mode landslide data and performs The model is then used to train a single-mode landslide identification and prediction model. The servicer side is a single-mode landslide identification and prediction model trained by each data end. Then, a multi-mode model fusion is performed to obtain a landslide identification and prediction base model. Then, the customization is applied to the user end in response to the customized needs of different user ends. It is used to return customized models to one or more user ends as consumers. and a sub-sample was used to request a landslide identification and prediction model suitable for local landslide data. Sends needs to the server side. For ease of explanation, in the embodiment of this application, the data in the landslide identification and prediction system is The data end is called the first device, the server end is called the second device, and the user end is called the The first devices are each connected to the second device so as to be able to communicate with each other, and the second device is is communicatively connected to a third device. Optionally, the landslide identification and prediction system is The fourth device may be communicatively connected to the third device, and Receive landslide identification and prediction results from the third device and display the landslide identification and prediction results to the user It is possible. Based on the above, the method for identifying and predicting landslides provided by the embodiments of the present application is The method is realized by the interaction of a plurality of devices, and the method includes the following steps: The stages may include: In the first stage, each primary device trains a single-mode landslide identification and prediction model; In the second stage, the second device fuses multiple single-mode landslide identification and prediction models to identify the landslide Obtain an identification and prediction-based model, Phase 3: The second device interacts with the third device, and the customization model is sent to the third device. Perform landslide identification and prediction. The present disclosure will now be described in detail in conjunction with the embodiments with reference to the accompanying drawings. First, take one first device as an example to see if the first device in the first stage is a single-mode landslide. The process of training the identification and prediction model will be briefly described. For example, the first device is Geological monitoring stations, monitoring points, and various resources are distributed throughout the area. These may be remote sensing aircraft, various weather balloons or various satellites. Referring to FIG. 1, FIG. 1 illustrates a single-mode landslide identification method provided by an embodiment of the present application. 1 is a flowchart of a method for training a predictive model, the method comprising the following steps: Step 101, construct a single-modal landslide dataset. The unimodal landslide dataset contains multiple training samples, each of which is a unimodal landslide. Includes landslide identification and prediction results of the single-mode landslide data and the single-mode landslide data Landslide data characterize the current geological conditions at the landslide location and are used to The determination and prediction results are data indicating whether a landslide has occurred and the type of landslide. The data supporting whether a landslide has occurred may be the probability of a landslide occurring, The data indicative of type may be the probability of occurrence of different types of landslides. Landslide Identification The prediction results are also called identification and prediction results or landslide result data, etc. does not limit the name. Here, the single-mode landslide data is collected by a first device, and the single-mode landslide data is The data is raw data. Different data types are sometimes called different modes, For example, a single data type is called single-mode and multiple data types are called multi-mode. After the first device collects the landslide data (if necessary, the first device The data was then pre-processed (cleaned, screened, etc.) and manually processed. Label the landslide data by experiment or other data labeling tools, and The landslide identification and prediction results corresponding to the received data are labeled (i.e., the landslide data is We label the corresponding locations for landslide occurrence and the landslide type. Make sure you get the pull. The type of landslide data can be selected from optical image data, synthetic aperture radar (SAR) image data, and At least one of the following data: data, laser radar point cloud data, or sensor monitoring data Contains at least one. Here, the optical image data is a photograph of the site where the landslide occurred taken by a camera. , satellite optical image data, or airborne optical image data. SAR image data is a type of data obtained by satellite SAR radar, ground-based synthetic aperture radar, etc. The data may be radar data or airborne synthetic aperture radar data. Laser radar point cloud data is obtained by an airborne laser scanner. It may be data. The sensor monitoring data are collected by surface sensors on the landslide body. The data may be monitoring data collected by an internal sensor of the ramp. In addition, it is possible to decide which type of landslide prevention facility to use in a given area based on actual needs and environmental conditions. It should be noted that some regions may choose to obtain data from The landslide location was covered by clouds, and synthetic aperture radar (SAR) images were used to identify such landslides. SAR images are suitable for collecting landslide data, e.g. Landslide locations in this area are often covered with vegetation, and the LiDAR point clouds The point cloud data is suitable for identifying such landslides. Collect as data. Step 102, a single mode preset based on the single mode landslide data set Train a landslide identification and prediction model. In addition, the model for identifying and predicting landslides based on single-mode landslide data is a single-mode model. Please understand that this model is sometimes called the landslide identification and prediction model. The model for identifying and predicting landslides based on multi-mode landslide identification data is multi-mode. Timothys Landslide Identification and Prediction Model, or Landslide Identification and Prediction Based Model, or Landslide This is sometimes called the basic model for error identification and prediction. Optionally, the single mode landslide identification and prediction model can be implemented by the server side (e.g., a second device). A single-mode landslide model is constructed based on the landslide model and corresponds to different types (or modes) of landslide data. The identification and prediction models may differ depending on the local data type. Then, the corresponding single-mode landslide identification and prediction model is requested from the second device, and the model is trained. Of course, the single-mode landslide identification and prediction model may be constructed independently by each first device. stomach. For example, for optical images, the single-mode landslide identification and prediction model uses the visual transform rmer neural network, and for synthetic aperture radar images, The landslide identification and prediction model may be a third-order convolutional neural network, For laser radar point clouds, the single mode landslide identification and prediction model is shown in Fig. It may be a neural network, and the sensor monitoring data (monitoring data The unimodal landslide identification and prediction model is a long-term With short-term memory networks and time-series neural networks such as Transformer It's fine. In the embodiments of the present application, the specific structure of the single-mode landslide identification and prediction model is This is not limited to the actual application, but may vary depending on the actual landslide data types and prediction results. Based on the characteristics (e.g., landslide type, etc.), a single-mode landslide identification and prediction model is developed. The structure can be designed. In the embodiment of this application, the process of training the single-mode landslide identification and prediction model includes the following steps: Includes: Input single mode landslide data as sample and convert it to single mode landslide data. The corresponding landslide identification and prediction results are output as desired samples, and the input landslide data is The results of the single-mode landslide identification and prediction model for the data (i.e., landslide identification and prediction The measured result is output as a predicted sample, and based on the predicted sample output and the desired sample output, Calculate the model loss of the single-mode landslide identification and prediction model using an appropriate loss function, Then, we optimize the model parameters based on the model loss and perform multiple iterations to obtain the final The cutoff condition for training is also set to a value that the model loss is reduced to a certain value. The model difference is smaller than a specified difference, and the number of training rounds exceeds a specified number. It includes various conditions such as training cutoff conditions, and appropriate training cutoff conditions can be selected according to the actual situation. . Based on the above embodiment, it is possible to selectively select the labeling difficulty of the landslide data ( On the other hand, the trained single mode In order to improve the accuracy of the landslide identification and prediction model, the first device is designed to identify single-mode landslides. Methods for training predictive models include supervised training and self-supervised training. Here, supervised training involves training raw training samples (the landslide data and the labeled results) The aim is to train the model using Self-supervised training constructs a teacher signal based on the landslide data, and compares the landslide data with the teacher signal. The training sample is a teacher signal, and the model is trained using the teacher signal as a training sample. The method is to mask some areas in the single mode landslide data and then How to predict masked regions based on the unmasked regions, or based on the single-mode regions of the previous moment? A method for predicting single mode landslide data at a later time based on landslide data, or other A similar method may be used. In the embodiment of the present application, the sample size of the training sample is increased by self-supervision. By doing so, the identification and prediction accuracy of the single-mode landslide identification and prediction model obtained by the final training was improved. In addition, by combining self-supervised training, To train the model, the first device does not need to rely on the second device to provide labels. The first device can independently complete the training task of the single-mode landslide identification and prediction model. This reduces the interaction between the first and second devices and saves time overhead. This can be done. Optionally, some landslide data further includes geolocation information, e.g. SAR Image data includes geographic location information for each pixel point. The geographic location information is longitude, latitude, and and altitude. In one embodiment, the landslide identification and prediction model is based on a geographic location corresponding to the landslide data. Based on the information, the location of the landslide point can also be predicted, i.e., the above landslide identification The prediction result further indicates the location of the landslide point, and the location of the landslide point is the longitude , latitude and altitude. Based on the above, the landslide identification and prediction model is used to predict the location of the landslide occurrence point. In this case, the landslide identification and prediction model will include a new It also includes a prediction module that can process the global network and landslide occurrence points. In this case, the loss of the landslide identification and prediction model is calculated by the landslide prediction loss (i.e., the landslide The prediction results include whether a landslide occurred and the type of landslide, and the location prediction loss. Note that the loss L of the model is: TIFF0007688807000002.tif628, L 1 indicates the landslide prediction loss, and L 2 denotes the location prediction loss, TIFF0007688807000003.tif43 and TIFF0007688807000004.tif43 is the landslide predicted loss L 1 and the location prediction loss L2 This is a parameter for balancing the above. Optionally, the neural network for processing geolocation information is graph neural It may be a network. In another embodiment, the landslide data may further include environmental information, Thus, the trained landslide identification and prediction model can predict landslides in conjunction with environmental information, It is possible to study the relationship between the occurrence of landslide disasters and environmental factors. Optionally, the environmental information may include weather information, load-bearing information of the landslide body, etc. By way of example, weather information may include, but is not limited to, factors such as temperature, precipitation, and wind speed. The load-bearing capacity information of the landslide body is not available from the outside, but it acts on landslides that are prone to landslides. It may include pressure, for example the pressure of water on the side of a dam. In the following, in terms of the interactions between the first, second and third devices, Thus, a landslide identification and prediction base model is obtained, and in the third stage, the landslide is detected by the third device. The process of completing slippage identification and prediction is described. For example, the second device is a multi-mode data It may be a remote server or a cloud server that provides a data fusion service, The third device is the server or the landslide identification and prediction server that needs to perform the landslide identification and prediction. It may be a personal user end that needs to use the service, e.g. 3 The equipment is small servers, laptops, tablets, mobile phones, smart TVs and other power sources. It may also be a child device. Referring to FIG. 2, FIG. 2 shows a method for identifying and predicting landslides provided by an embodiment of the present application. 1 is a flowchart of a method for detecting a cellular environment, the method comprising the steps of: In step 201, a second device extracts a plurality of single-mode landslide identification and prediction models from a plurality of first devices. Get Dell. Here, each single mode landslide identification and prediction model is a first device that performs the steps in the above embodiment. The first device is obtained by performing step 101 and step 102, i.e., the first device is a single mode device. Based on the landslide data set, a single-mode landslide identification and prediction model is obtained. Types of landslide data in the trained single-mode landslide dataset of one first device is the same. In the embodiment of the present application, after the first device trains a single-mode landslide identification and prediction model, Then, the single-mode landslide identification and prediction model is sent to the second device. Selectively, the second device is used to obtain a single mode landslide identification and prediction model from the first device. This can be the model itself or model parameters. Therefore, transmitting the model parameters can reduce data transmission. In one embodiment, the model parameters of the single-mode landslide identification and prediction model are The feature map of the landslide identification and prediction model is The feature map of the one-mode landslide identification and prediction model is used to extract the features in the model. It should be understood that the examples in this application reflect some of the structures and parameters of the Here, we mainly explain the case where the model parameters are feature maps. Step 202, the second device acquires a plurality of single-mode landslide coincidences obtained from a plurality of first devices. The constant and predictive models are fused into a multi-modal model. By data fusion of multiple single-mode landslide identification and prediction models, multi-mode A multi-mode landslide identification model suitable for processing landslide data can be obtained. do. Optionally, the specific process of multi-modal model fusion is shown in Figure 3 in conjunction with Figure 2. includes steps 2021 to 2022. In step 2021, a second device performs feature matching of a plurality of single-mode landslide identification and prediction models. Stitch the group together. Selectively, the second device can be used to set the model parameters of multiple single-mode landslide identification and prediction models. First, different single-mode landslide data are collected to facilitate model fusion. Align the model parameters of the different modes to align the differences introduced by It may be processed. The second device performs stitching on the feature maps of multiple single-mode landslide identification and prediction models. Then, we perform a contact operation to obtain the stitched feature map. For example, , multiple feature maps are F 1 , F 2 , ……, F n The characteristics after stitching are as follows: Let be a map and Z be a tensor. In some cases, the single-mode landslide data acquired by the multiple first devices may include the same type of There are many types of landslide data, and the first set of data is used for the same type of landslide data. Some of the single-mode landslide identification and prediction models trained by the same setup have similarities. In such a case, the second device scans the feature maps of multiple single-mode landslide identification and prediction models. Before the stitching process, multiple single-mode landslide identification and prediction models need to be preprocessed. It is necessary to reduce the amount of computation, improve efficiency, and simplify the base model after fusion. It can be made into Selectively preprocessing the above multiple single-mode landslide identification and prediction models can be performed in detail. Specifically, the second device includes: a feature map of a plurality of single-mode landslide identification models; Specifically, feature maps whose similarity exceeds a predetermined similarity are merged or de-duplicated. In step 2022, the second device uses a polynomial neural network and an attention mechanism. We construct a system to process the stitched feature maps. Optionally, the attention mechanism may be a channel attention mechanism and / or contains a spatial attention mechanism. So far, the feature maps of the base models obtained by multi-modal model fusion We use stitched feature maps, polynomial neural networks and attention Includes a locking mechanism. In addition, polynomial neural networks rk, PNN) is a special kind of neural network whose main feature is that it The objective is to perform polynomial mapping and nonlinear transformation on the data. It is used to process data modeling tasks. The core of PNN is a polynomial activation function. The PNN nonlinearly transforms the input features using a polynomial activation function, and extracts the high-level features in the data. It is possible to capture characteristic interaction relationships. In the embodiment of the present application, the method for obtaining the polynomial neural network is a residual network. Based on a network (e.g., a residual network like ResNet), the “addition” operation is replaced by “multiplication” ” operation, connect the residual modules between layers, and remove the activation function from the residual network. The realization method of the polynomial neural network is shown in the following formula. Been: TIFF0007688807000005.tif7576 where Z is the stitched feature map and Z′ is the polynomial neural network. is the feature map processed by the function, Conv() is the convolution operation, and Norm () is the normalization operation and "*" is the multiplication operation. In an embodiment of the present application, multi-modal feature fusion is performed using a polynomial neural network. We have completed the study and have developed a method to investigate the properties of nonlinear activation functions in polynomial neural networks. Properly handle complex nonlinear relationships in multimodal data (i.e., multiple feature maps). It is possible to capture the complex interactions between different modes and high-order feature interaction relationships. The expressive power of the model obtained by processing the polynomial neural network is enhanced. Therefore, the model can make more comprehensive use of multi-modal information, and the overall model This can improve performance. Furthermore, the feature map (Z') processed by the above polynomial neural network is This is handled by a tension mechanism, and the fused model can be obtained, which ,The model can capture important information from the input ,information. In an embodiment of the present application, spatial attention mechanisms and / or channel attention mechanisms are used. Here, the spatial self-attention mechanism (Spatial Self-Attention) is a method to detect spatially structured data, e.g. It is used to process sequence data, such as images or spatial arrays, and channel attenuation The interaction mechanism allows us to analyze the relationships between different locations (or spaces) in the data. For example, for each location or pixel, the spatial self-attention mechanism can To understand global contextual information, we calculate the correlation between the location and all other locations. This calculation helps the model capture the dependencies and importance between different locations in the data. It can be extended. Channel Self-Attention Mechanism n) are used to process the channel dimension of the data. For example, in image data In this case, the channel dimension usually represents different features or feature maps, and the channel self-assignment The collaboration mechanism is able to analyze the relationship and significance between different channels of data. By introducing the channel attention mechanism, the processing model can When processing data, different channels are used to improve the richness of the data representation and model performance. The information in the document can be effectively integrated and utilized. In one embodiment of the present application, the two self-attention mechanisms may be used alone. For example, when processing two-dimensional image data, spatial autocorrelation Using an attention mechanism, the spatial relationship between pixels is taken into account, and channel self-attention By using a scalar mechanism to handle the relationships between different feature channels, model performance is improved. can be improved. In addition, the model uses spatial attention mechanism and channel attention mechanism. The order of processing the two attention mechanisms is restricted when processing the feature maps of For example, the spatial attention mechanism first processes the image, and then the channel It may be handled by the attention mechanism, or it may be handled by the channel attention mechanism first. It may be processed by the spatial attention mechanism and then by the spatial attention mechanism. stomach. Step 203, the second device performs the fusion based on the multi-modal landslide data set. The base model is trained to obtain a landslide identification and prediction base model. Multi-mode landslide datasets are generated using multiple modes (all single landslides received by the second device). Multi-mode landslide data refers to a collection of landslide data (including modes corresponding to the multi-mode model) and The slip data set is also called the reference data set. In an embodiment of the present application, the second device performs melting based on a multi-modal landslide data set. The process of training the combined base model is as follows: the first device is a single-mode landslide identification and prediction model; The process is similar to that of training the model, so it will not be repeated here. and obtaining a landslide identification and prediction base model, and storing the model locally in the second device. After that, other devices (e.g., a third device at the user's side) need to perform landslide identification and prediction. In some cases, the landslide identification and prediction model can be customized to the second device. Referring to FIG. 4, for example, the second device receives three single-mode signals transmitted from the first device. Take the case of receiving feature maps from a boundary identification and prediction model as an example. The process of integrating slip identification and prediction-based models is explained as an example. Here, three The feature maps of the single mode landslide identification and prediction model are feature map a, feature map b and feature map c. In Fig. 4, the stitched feature map, polynomial The neural network and attention mechanism are the feature extraction module of the base model. The convolutional layer is the prediction module of the base model ( TIFF0007688807000006.tif2284), and the prediction module processes it to obtain the landslide identification and prediction results. The base model is continuously updated by seeding, and a base model for landslide identification and prediction is obtained. They are trained to do so. Optionally, the prediction module uses a convolutional layer Conv_(1×1)(Z) to predict landslides. It can be a model that realizes identification and prediction by using a multi-layer perceptron, Linear It may be a module that realizes landslide identification and prediction using r(Z). As described above, the second device converts multiple single-mode landslide identification and prediction models into multi-mode The process of model fusion and obtaining a landslide identification and prediction base model is a continuous process with lateral coordination training. The advantage of the base model is that it can be used to train landslides obtained from various landslide data. The base model is a fusion of landslide characteristic factors, and is useful for identifying and predicting landslides in different regions. This can be used as the basis for a measurement model, thus allowing the identification of landslides dedicated to different regions. This eliminates the need to train fixed and predictive models separately, saving resources. In some embodiments, the first device performs a single mode detection based on new training samples. The landslide identification and prediction model was updated, and the updated single-mode landslide identification and prediction model was used for the second Thus, the second device may receive multiple updated Receive single-mode landslide identification and prediction model and update landslide identification and prediction base model Specifically, the second device is an updated single-mode landslide identification and prediction model. The feature map of the original single-mode landslide identification and prediction model was replaced with the feature map of the original single-mode landslide identification and prediction model. The feature map of the other single-mode landslide identification and prediction models is not changed, and the second device The model is trained to obtain a new base model based on the reference dataset. Optionally, the second device can use the updated unimodal landslide identification and prediction model from the first device. After receiving, the second device may test the updated single mode landslide model, i.e. That is, the reference data from the second device is input to the updated single mode landslide model, and the value ( The performance evaluation indexes for each prediction accuracy such as accuracy, recall, etc. The prediction accuracy of the updated single-mode landslide model is compared with that of the single-mode landslide model before the update. If the prediction accuracy of the updated single-mode landslide model is better than that of the previous landslide model, the updated single-mode landslide model is The identification and prediction model is used to update the landslide identification and prediction base model, otherwise, the corresponding The updated single-mode landslide identification and prediction model is used as the basis for updating the landslide identification and prediction model. Do not use again. In one embodiment, the prediction accuracy of the updated unimodal landslide identification and prediction model is improved. A method to determine dolphins with better prediction accuracy than previous single-mode landslide identification and prediction models The results are as follows: the prediction accuracy of the updated single-mode landslide identification and prediction model and the Determine whether the difference in prediction accuracy between the previous and new single-mode landslide identification and prediction models is greater than a specified difference. The prediction accuracy of the updated single-mode landslide identification and prediction model is compared with that of the single-mode landslide identification and prediction model before the update. If the difference in prediction accuracy of the landslide identification and prediction model is greater than a specified difference, the updated This means that the prediction accuracy of the single-mode landslide identification and prediction model is higher. In the embodiment of the present application, the prediction accuracy of the updated unimodal landslide model is compared with that of the unupdated unimodal landslide model. The second device uses the base model only if the prediction accuracy is better than that of the modal landslide model. In this way, the number of interactions between the first device and the second device is reduced, and each first device end Single-mode landslide identification due to differences in the computational capabilities of the equipment, landslide data and models - Solves the problem of different training time overheads depending on the prediction model, To some extent, the time overhead required for updating slip identification and prediction-based models is reduced. It is possible. Step 204: The third device sends a model customization request to the second device. In addition, the landslide identification and prediction-based model trained by the second device has multiple modes of This is a basic model that combines landslide characteristic factors and is suitable for different types of landslide data in different regions. Therefore, for different regions, the third device can detect the local landslides from the second device to the third device. Submit a model customization request to customize a model suitable for processing your data. It is understood that one can trust. In an embodiment of the present application, the model customization request includes a local landslide of the third device. Target data type (also called target landslide data type or target landslide data mode) Select one or more types of local landslide data from the third device. may include multiple types. Step 205, the second device responds to the model customization request sent from the third device. landslide identification and prediction base model is clipped and landslide identification and prediction customization is performed. Get the model. In an embodiment of the present application, the second device clips the landslide identification and prediction based model The model obtained by the above is also a base model, and the base model is based on the area in which the third device is located. It is suitable for processing landslide data. Optionally, step 205 may be accomplished by step 2051, as shown in FIG. It is possible. Step 2051, the second device responds to the model customization request sent from the third device. Among the feature maps of the landslide identification base model, the feature maps other than the target feature map are Then, we obtain a customized landslide identification and prediction model, where the target feature match is A group consists of one or more units that correspond to the type of landslide data in the model customization request. Feature map of the one-mode landslide identification and prediction model. In an embodiment of the present application, the second device clips the landslide identification and prediction based model This effectively simplifies landslide identification and prediction using a single mode without local landslide data types. The feature map of the measured model is then removed, and the resulting landslide identification and prediction is customized. The model is better able to handle local landslide data from a third device. Step 206: The second device transmits the landslide identification and prediction customized model to the third device. do. The second device can be selected to input the model parameters of the landslide identification and prediction customization model to the third device. The model parameters may be feature maps of the model. In step 207, a third device performs landslide identification based on the local landslide data set. The customized predictive model is locally trained to obtain a target landslide identification and prediction model. The above localized training means that the third device is a customer-specific landslide identification and prediction system that can be implemented by transfer learning. The size model was then transferred to a domain appropriate for the local landslide data from the third device (i.e. Landslide identification and prediction Customized model fine-tuning and localized landslide identification and prediction The aim of the study is to obtain a model (i.e., a target landslide identification and prediction model). · The prediction model reuses the landslide characteristic factors of the landslide identification and prediction-based model, The target landslide identification and prediction model is trained using local landslide data from the third device. This is better than the previous model. In step 208, the third device uses the target landslide identification and prediction model to determine the location of the third device. The landslide data of the target area is processed to obtain the landslide identification and prediction results, and the landslide identification and prediction results are The results indicate whether a landslide has occurred and the type of landslide. In the embodiment of the present application, the target landslide identification and prediction obtained by the above step 207 The model outperformed a model trained using only local landslide data from the third device. Therefore, the prediction accuracy of landslide identification and prediction based on the target landslide identification and prediction model is also high. become. In some embodiments, the third device end is for detecting different types of landslide identification and prediction results. Generate different warning information based on the probability of landslide occurrence, and distribute the warning information to at least one of the first 4 devices, which can display the warning information to the user. 4. Device end is personal user end, such as TV, mobile phone, tablet, personal It may be a computer, a small server, or some other electronic device. Thus, in order to better understand the interaction process between the first device, the second device, and the third device, In addition, the embodiment of the present application makes the interaction process easier to understand and more intuitive through FIG. 1 shows a schematic flow diagram for the embodiment shown in FIG. As shown in Fig. 6, the first devices each collect landslide data in different modes. (e.g. optical imagery, SAR imagery, LiDAR point clouds, etc.) The second device is trained to obtain a single-mode landslide identification and prediction model from the first device. Landslide identification and prediction models are downloaded and linked for continuous learning (model fusion and model training). The third device is a landslide identification and prediction base model. Customized landslide identification and prediction Get customized models and local landslide data Based on the dataset, transfer learning is performed on the customized model to obtain a localized A landslide identification and prediction model is obtained. Selectable, localized landslide identification and prediction models are available for different landslide identification and prediction scenarios. Models for applications, e.g., used to rapidly identify landslides triggered by earthquakes The model for identifying and predicting landslides that may occur in the long term after an earthquake occurs. A localized post-earthquake landslide identification and prediction model used to predict potential landslides A localized ancient landslide identification and prediction model for predicting landslides that occurred many years ago The model includes a localized landslide hazard identification and prediction model for predicting landslides during deformation. That's fine. The embodiments of the present application provide a method, an apparatus and a system for identifying and predicting landslides, In general, it is necessary to predict landslides over a wide area and to use heterogeneous data from multiple sources for landslide identification and prediction. A landslide identification and prediction model based on monitoring data is proposed. The device converts the single-mode landslide identification and prediction model transmitted from multiple first devices into a multi-mode The base model obtained by integrating the two models is the basis for landslide identification and prediction models in different regions. Thus, it can be used as a landslide identification and prediction model dedicated to different regions. This saves resources to obtain a base model for landslide identification and prediction without the need to train each model separately. In addition, the second device can clip the base model to reduce the distortion of the third device. Meet the customization needs, the third device is received based on local landslide data The model is then locally trained and used for landslide identification and prediction. The estimation and prediction model has a more accurate landslide disaster identification capability. The proposed technical solution is to share landslide data among the common data centers without sharing the data. The landslide identification and prediction model can be obtained, realizing landslide identification and prediction and saving resources. Furthermore, multiple single-mode landslide identification and prediction models were fused and the fused base model was The model captures the complex interaction relationships between different modal landslide data and facilitates landslide identification and prediction. It can capture important features for prediction, strengthening the expressiveness of the model and improving its performance. This can improve the accuracy of landslide identification and prediction. An embodiment of the present application relates to a method for controlling a plurality of first devices, a second device, and one or more third devices. The present invention further provides a landslide identification and prediction system that consists of the above devices, each of which interacts with the others to implement the above. Each step in the example is carried out. Further referring to FIG. 7 to FIG. 9, as an implementation of the method shown in each of the above figures, the embodiment of the present application is The present invention provides a first device, a second device and a third device, respectively, for single-mode landslide identification. An example of the training process of the prediction model is shown in FIG. 7, which is performed by the first device to identify landslides. The embodiment of the process of obtaining the prediction-based model is performed by the second device shown in FIG. The example of the customization and localization process of the identification and prediction model is shown in FIG. 9. It is performed by. As shown in FIG. 7, the first device 700 provided by the embodiment of the present application includes an acquisition module. The acquisition module 701 may include a model training module 702. Executing step 201 in the method embodiment to construct a single-modal landslide data set The model training module 702 performs step 202 in the above method embodiment. Implemented a unimodal landslide identification and prediction model based on the unimodal landslide dataset. are trained to obtain As shown in FIG. 8, the second device 800 of the embodiment of the present application includes a model fusion module 801, A delta training module 802, a model clipping module 803 and a transmit / receive module The model fusion module 801 may include a model fusion module 804. 202 (specifically including steps 2021 and 2022) The model training module 802 is used to perform step 203 in the above method embodiment. The model clipping module 803 is used in step 205 in the above method embodiment. (specifically including step 2051), and the transceiver module 80 4 is used to perform steps 201 and 206 in the above method embodiment. do. Optionally, the second device 800 may be configured to generate a feature map of a plurality of unimodal landslide identification models. It is used to merge or de-duplicate feature maps whose similarity exceeds a certain threshold. The device further comprises a pre-processing module. As shown in FIG. 9, the third device 900 of the embodiment of the present application includes a transceiver module 901, a model The transmitting and receiving module may further include a train module 902 and a prediction module 903. The module 901 is used to perform step 204 in the above method embodiment, and the model The training module 902 is used to perform step 207 in the method embodiment above. The prediction module 903 is used to perform step 208 in the above method embodiment. can be. Each module of the first device 700, the second device 800 and the third device 900 further includes It can be used to perform other steps in the method embodiment, and all of the above method embodiments can be used to perform other steps in the method embodiment. The relevant contents may be cited in the functional description of the corresponding functional module, and are not repeated here. Don't repeat. Further, an embodiment of the present application further provides an electronic device, the electronic device comprising at least one a processor and a memory communicatively coupled to the at least one processor; wherein the memory stores instructions executable by at least one processor; When the instructions are executed by at least one processor, the at least one processor The processor performs the method steps performed by the first device, the second device and the third device in the above embodiment. You can run a backup. Furthermore, embodiments of the present application further provide a computer-readable storage medium, A readable storage medium has computer instructions stored thereon, the computer instructions being adapted to execute a program for a computer. When executed by the method, the method steps according to any of the above embodiments can be realized. An embodiment of the present application further provides a computer program product, The program product includes computer program instructions, the computer program instructions being a process When executed by the processor, the method steps according to any of the above embodiments can be realized. Cut. FIG. 10 is a schematic diagram of an exemplary electronic device 1000 that may be used to implement embodiments of the present application. The electronic device may be a laptop, a desktop computer, a workstation, or a Applications, personal digital assistants, servers, blade servers, mainframes Digital computers in various forms, such as computers and other suitable computers The term "electronic equipment" is intended to represent a wide range of electronic devices, including personal digital processing, mobile phones, and other electronic devices. , smartphones, wearable devices, and other similar computing devices. The term may refer to various forms of mobile devices, such as a PC, a portable USB computers, mobile electronic devices, centralized servers, distributed servers, image acquisition terminals , sensor terminals, artificial satellites, mobile phones, in-vehicle terminals, large screens, and other terminal devices. The first device, the second device, the third device and / or the fourth device are any of the electronic devices 1000. The present invention may be applied to any combination of the above-mentioned components. , their connections and relationships, and their functions are merely exemplary and are described herein. and / or are not intended to limit the embodiments of the present disclosure as claimed. As shown in FIG. 10, an electronic device 1000 includes a read only memory (ROM) 1002. A computer program stored in the memory unit 1008 or a random access memory Based on the computer program loaded in the RAM 1003, various appropriate It comprises a computing unit 1001 capable of performing operations and processing. Various programs and data necessary for the operation of the device 1000 are stored in the RAM 1003. The computing unit 1001, the ROM 1002, and the RAM 1003 are connected to each other via a bus 1004. 5 is also connected to bus 1004. A plurality of components in the electronic device 1000, i.e., an input unit 1 such as a keyboard and a mouse 006, various displays, output units such as speakers 1007, magnetic disks, optical Storage units 1008 such as disks, network cards, modems, wireless communication transceivers A communication unit 1009 such as a server is connected to the I / O interface 1005. The communication unit 1009 enables the device 1000 to communicate with a computer network of the Internet. and / or exchange information / data with other devices via various telecommunications networks. It is possible. The computing unit 1001 includes various general-purpose and The computing units 1001 may be dedicated processing components. Some examples are central processing units (CPUs), graphics processing units (GPUs), and various Dedicated artificial intelligence (AI) computing chips, which run various machine learning model algorithms a computing unit, a digital signal processor (DSP), and any suitable including, but not limited to, processors, controllers, microcontrollers, etc. The computing unit 1001 executes the various methods described above. For example, in some embodiments, landslide identification methods provided by embodiments of the present application include The method of predicting may be implemented as a computer software program and A readable medium, such as a memory unit 1008, may be included. In some embodiments, In addition, a part or the whole of the computer program may be written in the ROM 1002 and / or the communication unit. may be loaded and / or installed into the device 1000 via unit 1009. A computer program is loaded into the RAM 1003 to operate the computing unit. When executed by 1001, one or more steps of the method described above are performed. Optionally, in another embodiment, the computing unit 1001 may The method may be implemented in any other suitable manner (e.g., with the aid of firmware). The device may be configured to perform the above-mentioned operations. In the context of this disclosure, a machine-readable medium refers to a program executed by an instruction execution system, device, or apparatus. programs for use in or with an instruction execution system, device, or apparatus The machine-readable medium may be a tangible medium that can contain or store a signal. The machine-readable medium may be any of a variety of media, including electronic, magnetic, optical, electromagnetic, red, and / or magnetic recording media. External wiring, or semiconductor system, device, or equipment, or any suitable combination thereof More specific examples of machine-readable storage media include, but are not limited to, Electrical connections based on one or more wires, portable computer disks, hard disks Random Access Memory (RAM), Read Only Memory (ROM), Erasable Program Programmable read-only memory (EPROM or flash memory), optical fiber , convenient compact disc read-only memory (CD-ROM), optical storage devices, magnetic or any suitable combination of the foregoing. Each embodiment in this specification is described in a step-by-step manner, and the same and similar parts of each embodiment are not necessarily the same. It is sufficient to refer to each embodiment for the details, and each embodiment focuses on the differences from other embodiments. is doing. Finally, the above embodiments are only used to illustrate the technical solutions of the present invention. Please note that the present invention is not limited to the above examples. Although the technical solutions described in the above embodiments are described in detail, those skilled in the art may easily modify the technical solutions described in the above embodiments. or some of its technical features may be equivalently substituted, and these modifications or substitutions are The essence of the corresponding technical solution deviates from the spirit and scope of the technical solution of the embodiment of the present application. It is not something that can be done.

Claims

1. 1. A method for identifying and predicting landslides, comprising: A second device is used to obtain multiple single-mode landslide identification and prediction models obtained from multiple first devices. Multi-mode model fusion was performed based on the model and multi-mode landslide data set was used. A fusion base model is trained using the multi-layered model to obtain a landslide identification and prediction base model. Multi-mode model fusion is a method to fusion the feature maps of the multiple single-mode landslide identification and prediction models. We use polynomial neural networks to stitch and process the stitched feature maps. Building a neural network and attention mechanism, Here, each single-mode landslide identification and prediction model is a single-mode landslide identification and prediction model for each first device. The single-mode landslide data set of one first device may be trained based on the data set. The types of landslide data in the dataset are the same. The second device, in response to a model customization request sent from a third device, The landslide identification and prediction based model is based on the type of local landslide data of the device. and clipping the landslide identification and prediction customized model to the third device. a customized model for predicting and identifying a target object; The type of local landslide data of the third device is included; The third device performs the landslide identification and prediction customization based on a local landslide data set. To obtain a target landslide identification and prediction model, The third device employs the target landslide identification and prediction model to identify the landslide area where the third device is located. The landslide data of the area is processed to obtain the landslide identification and prediction results, and the landslide identification and prediction The measurement results indicate whether a landslide has occurred and the type of landslide; A method for identifying and predicting landslides, comprising:

2. The second device, in response to a model customization request sent from a third device, The landslide identification and prediction based model is based on the type of local landslide data of the device. Clipping and obtaining a customized landslide identification and prediction model includes: The second device responds to a model customization request sent from the third device by In the feature map of the landslide identification-based model, feature maps other than the target feature map are cleared. to obtain the customized landslide identification and prediction model, wherein the target feature map is The drop is one or more landslide data types corresponding to the model customization request. is a feature map of multiple single-mode landslide identification and prediction models. The method according to claim 1.

3. The types of landslide data include optical image data, synthetic aperture radar (SAR) image data, radar image data, and at least one of the radar point cloud data or the sensor monitoring data 3. The method according to claim 1, further comprising:

4. The landslide data further includes geolocation information, and the landslide identification and prediction results are 3. The method according to claim 1 or 2, further comprising indicating the location of the landslide point.

5. A single-mode landslide identification and prediction model obtained from the first device by the second device. are model parameters of the single mode landslide model.

3. The method according to claim 1 or 2.

6. The model parameters of the single-mode landslide identification and prediction model are 6. The method of claim 5, wherein the feature map is a feature map of an identification and prediction model.

7. The method further comprises: The second device is configured to select a feature map of the plurality of single-mode landslide identification models based on a similarity The present invention is characterized in that the feature maps exceeding a predetermined similarity are merged or de-duplicated. The method according to claim 6.

8. The method further comprises: The second device receives a plurality of updated single-mode landslide identification and prediction data from the plurality of first devices.

3. The method of claim 1, further comprising receiving a measurement model.

9. The prediction accuracy of the plurality of updated single-mode landslide identification and prediction models and the single-mode landslide identification and prediction models before the update are compared. The difference in prediction accuracy of the modal landslide identification and prediction model is larger than a specified difference. The method according to claim 8.

10. The first device is adapted to perform single-mode landslide data acquisition using a training method including supervised training and self-supervised training. training the single-mode landslide identification and prediction model based on a dataset. The method according to claim 1 or 2, wherein

Images