Associative learning and UNet-based methods, equipment, and media for assessing landslide probability.

Abstract

In the field of geological hazard risk analysis, we provide, in particular, collaborative learning and UNet-based methods, equipment, and media for evaluating the likelihood of landslides. [Solution] The method includes: step S1 of collecting data on multiple landslide influencing factors; step S2 of establishing a federative learning framework; step S3 of designing a UNet deep learning model, which involves feature extraction and segmentation; step S4 of training and parameterizing the federative learning model, which updates the UNet-based global deep learning model; step S5 of iterating through federative learning and optimizing the model, which optimizes the parameters of the UNet-based deep learning model; step S6 of outputting an evaluation result of the likelihood of landslide occurrence; and step S7 of providing a warning of landslide risk and a proposal for disaster prevention and mitigation.

Description

Technical Field

[0001] The present invention relates to the technical field of geological disaster risk analysis, and particularly relates to an evaluation method, device, and medium for landslide occurrence probability based on joint learning and UNet.

Background Art

[0002] The evaluation of landslide occurrence probability is an important part of geological disaster prevention and preparedness. In conventional evaluation methods, generally, judgments are made depending on on-site investigations and experience, which takes time and is costly. With the development of computer technology and data science, data-driven methods have gradually become an important research direction for the evaluation of landslide occurrence probability. These methods analyze geological environmental elements, landslide history data, etc. to predict the probability of future landslide occurrence and provide decision support.

[0003] Conventional evaluation methods for landslide occurrence probability, such as experimental models, analytic hierarchy process, and statistical models, can effectively predict landslide risks, but have limitations in processing the fusion of complex data and multi-source information. Especially when analyzing image data, conventional convolutional neural network (CNN) methods cannot perform pixel-level classification, resulting in the loss of detailed information and affecting the accurate identification of landslide areas. Moreover, CNN lacks skip connections and cannot effectively recover the high-resolution details of images. Also, when the data volume is small, the performance is poor and the efficiency is low.

Summary of the Invention

Problems to be Solved by the Invention

[0004] In order to solve the limitations of conventional evaluation methods, the present invention aims to provide an evaluation method for landslide occurrence probability that combines joint learning and deep learning technologies, effectively solving the problems existing in centralized training such as slow data processing speed, insufficient memory, and low security, and enabling a more efficient and secure evaluation of landslide occurrence probability.

Means for Solving the Problems

[0005] This invention provides a method, apparatus, and medium for evaluating landslide probability based on associative learning and UNet.

[0006] A computer-based federated learning and UNet-based method for evaluating the likelihood of landslides, The acquisition of landslide influencing factor data, comprising step S1 of collecting data on multiple landslide influencing factors and extracting data on multiple influencing factor layers in a landslide-prone area, Establishing a federative learning framework, step S2 involves establishing a federative learning framework among multiple data holders, The design of a UNet deep learning model, comprising step S3 of designing a UNet-based deep learning model, using it to evaluate the likelihood of landslide occurrence, inputting influencing factor layer data into the model, and performing feature extraction and segmentation. The process involves training and integrating parameters of a federative learning model, specifically step S4, which involves training a UNet-based deep learning model using a landslide sample dataset within a federative learning framework, integrating model parameters using a federative averaging algorithm, and updating the UNet-based global deep learning model. The process involves iterative associative learning and model optimization, specifically step S5, which optimizes the UNet-based deep learning model parameters through multi-round iterations of associative learning. The process involves generating a feature map of landslide potential and performing probabilistic regression, comprising step S6: generating a feature map of landslide potential regions using an optimized UNet-based deep learning model, performing probabilistic regression using a Sigmoid layer, and outputting an evaluation result of landslide potential; The verification of results and proposals for disaster prevention and mitigation include step S7, which involves verifying and assessing the results of the evaluation of the likelihood of landslides occurring, and providing warnings about landslide risks and proposals for disaster prevention and mitigation based on the evaluation results.

[0007] Selectively, the multiple landslide influencing factor data in step S1 include topographic elements, geological elements, hydrological elements, land cover elements, and climatic elements, where, The aforementioned topographic elements include slope direction, slope degree, elevation, and curvature. The aforementioned geological elements include lithology, The aforementioned hydrological elements include the average annual precipitation and the distance from the river system. The aforementioned land cover elements include enhanced vegetation index, normalized differential vegetation index, surface reflectance, and land use. The aforementioned climatic elements include the annual average temperature and the ground surface temperature.

[0008] Selectively, the multiple influencing factor layer data in step S1 include elevation, slope, slope direction, curvature, lithology, distance from a water system, and land use.

[0009] Selectively, the construction of the aforementioned landslide sample dataset is The data collection process includes the steps of obtaining landslide history data and landslide risk point data, The construction of a landslide sample dataset involves constructing a landslide sample dataset based on landslide history data and landslide risk point data, determining the number of landslide rasters within a rectangular area, marking the area as a landslide point if the number of landslide rasters exceeds 50%, marking it as a non-landslide point if the number is less than 50%, and defining the marked landslide sample dataset as an experimental dataset. This involves dividing the experimental dataset into a training set and a validation set in a random ratio of 8:2, where landslide samples are labeled as 1 and non-landslide samples are labeled as 0. Data augmentation, which includes the step of performing data augmentation on constructed landslide sample data.

[0010] Selectively, the data augmentation is, Identification of landslide feature rasters, comprising the steps of identifying rasters of landslide points in a landslide sample dataset, Application of Mixup data augmentation technology, wherein Mixup data augmentation technology is used to randomly combine rasters of identified landslide points to generate augmented samples, Adding an extended sample to an experimental dataset, comprising the steps of adding the generated extended sample to the experimental dataset.

[0011] Selectively, the parameter updates of the UNet-based global deep learning model are performed. JPEG0007906342000004.jpg17170JPEG0007906342000005.jpg42170

[0012] Selectively, the training of the UNet-based deep learning model is performed as follows: The activation function is set to LeakyReLU, which is used as the activation function for each convolutional layer in the UNet-based deep learning model. JPEG0007906342000006.jpg16170JPEG0007906342000007.jpg6170 Model training, applying weight-based learning in the convolutional layer, JPEG0007906342000008.jpg10170JPEG0007906342000009.jpg18170JPEG0007906342000010.jpg22170JPEG0007906342000011.jpg15170JPEG0007906342000012.jpg7170JPEG0007906342000013.jpg10170JPEG0007906342000014.jpg8170Regularization by Dropout, which uses Dropout technology to prevent overfitting, JPEG0007906342000015.jpg24170 Reflection padding, processing the convolution boundary using reflection padding, JPEG0007906342000016.jpg10170JPEG0007906342000017.jpg25170 Model validation, using a binary cross-entropy loss function, takes into account the difference between the probability distribution predicted by the model and the actual distribution. JPEG0007906342000018.jpg16170JPEG0007906342000019.jpg25170

[0013] Selectively, the generation of a feature map of landslide probability in S6 and the regression of probabilities are performed. Optimization and application of a UNet model, step S61 of using an optimized UNet-based deep learning model to make predictions for areas prone to landslides, and extracting features of the landslide areas through multilayer convolution and manipulation, JPEG0007906342000020.jpg17170JPEG0007906342000021.jpg22170JPEG0007906342000022.jpg15170JPEG0007906342000023.jpg16170A calculation of the likelihood of landslide occurrence, comprising step S64 of generating a landslide likelihood map by calculating probability results using a sigmoid layer and predicting the corresponding probability value of landslide occurrence in the image of each pixel point.

[0014] Selectively, the aforementioned landslide risk warnings and disaster prevention / mitigation proposals are, JPEG0007906342000024.jpg49170 The formulation of disaster prevention and mitigation proposals includes the steps of formulating appropriate disaster prevention and mitigation proposals for different risk warning areas, proposing enhanced monitoring, personnel evacuation, and restrictions on development activities for red warning areas, proposing the implementation of construction and maintenance measures and enhanced patrols for yellow warning areas, and proposing the deployment of regular monitoring and periodic assessments for green warning areas.

[0015] An electronic device including a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein when the processor executes the computer program, the computer executes federated learning and a method for evaluating the likelihood of landslides based on UNet is realized.

[0016] A computer-readable storage medium storing a computer program, wherein when this program is executed by a processor, the computer executes federated learning and a method for evaluating the likelihood of landslides based on UNet is realized.

Advantages of the Invention

[0017] In the present invention, by adopting federated learning technology, each participant can ensure data privacy and security by performing federated learning with a locally trained model without sharing the original data. This technology breaks through the bottlenecks of traditional centralized data storage and computing, enables multiple data holders to share knowledge in a distributed environment without exposing sensitive data. With the introduction of federated learning, this evaluation method has a high privacy protection ability and is suitable for fields with high data privacy requirements, such as government agencies and scientific research departments.

[0018] The present invention uses a deep learning model based on UNet, which has excellent feature extraction ability and can effectively improve the identification efficiency and accuracy of landslide occurrence probability areas. The UNet model extracts the features of multiple influencing factors such as terrain and geology through multi-layer convolution and pooling operations to more accurately assess the likelihood of landslides. By combining these advanced technologies, a more efficient and accurate landslide occurrence probability assessment can be realized, and the model can be updated immediately to respond to new data changes, having high application value and social benefits.

Brief Description of the Drawings

[0019] [Figure 1] This is a flowchart of the evaluation method according to an embodiment of the present invention. [Figure 2] This is a schematic diagram showing the geographical location and landslide distribution of the research area in Beichuan Qiang Autonomous County according to an embodiment of the present invention. [Figure 3] This is a schematic diagram showing the geographical location and landslide distribution of the research area in Wenchuan County according to an embodiment of the present invention. [Figure 4] This is a schematic diagram of the elevation map of Wenchuan County according to an embodiment of the present invention. [Figure 5] This is a schematic diagram of the elevation map of Beichuan Qiang Autonomous County according to an embodiment of the present invention. [Figure 6] This is a schematic diagram of the gradient drawing for Wen River County according to an embodiment of the present invention. [Figure 7] This is a schematic diagram of the slope drawing in Beichuan Qiang Autonomous County according to an embodiment of the present invention. [Figure 8] This is a schematic diagram of the slope direction drawing in Bunsen County according to an embodiment of the present invention. [Figure 9] This is a schematic diagram of the slope direction drawing in Beichuan Qiang Autonomous County according to an embodiment of the present invention. [Figure 10] This is a schematic diagram of the curvature plot of Wen River County according to an embodiment of the present invention. [Figure 11] This is a schematic diagram of the curvature drawing of Beichuan Qiang Autonomous County according to an embodiment of the present invention. [Figure 12] This is a schematic diagram of the lithology of Wenchuan County according to an embodiment of the present invention. [Figure 13] This is a schematic diagram of the lithology of Beichuan Qiang Autonomous County according to an embodiment of the present invention. [Figure 14] This is a schematic diagram illustrating the distance from the river system of Wenchuan County according to an embodiment of the present invention. [Figure 15] This is a schematic diagram illustrating the distance from the river system of Beichuan Qiang Autonomous County according to an embodiment of the present invention. [Figure 16] This is a schematic diagram of land use mapping in Wenchuan County according to an embodiment of the present invention. [Figure 17] This is a schematic diagram of land use mapping in Beichuan Qiang Autonomous County according to an embodiment of the present invention. [Figure 18]This is a schematic diagram of the likelihood of landslide occurrence based on distributed training according to an embodiment of the present invention. [Figure 19] This is a schematic diagram of the likelihood of landslide occurrence based on non-distributed training according to an embodiment of the present invention. [Figure 20] This is a schematic diagram of the likelihood of landslide occurrence based on distributed training according to an embodiment of the present invention. [Figure 21] This is a schematic diagram of the likelihood of landslide occurrence based on non-distributed training according to an embodiment of the present invention. [Modes for carrying out the invention]

[0020] As shown in Figures 1 to 21, a computer-based federated learning and UNet-based method for evaluating landslide probability includes the following steps:

[0021] Step 1: Extract landslide spatial data for the research area.

[0022] In this study, we selected areas in southwestern China prone to frequent landslides as our research subjects, focusing on Wenchuan County and Beichuan Qiang Autonomous County for methodological validation. These regions are known for their complex topography and geological conditions, and frequent seismic activity and torrential rains lead to frequent landslide disasters. Since the Wenchuan earthquake, the combined effect of geological structural changes and rainfall has further increased the risk of landslides.

[0023] As shown in Figures 2 and 3, landslide sample data has been collected from Wenchuan County and Beichuan Qiang Autonomous County.

[0024] Step 2: Extract information on influencing factors.

[0025] Influence factor information is extracted from the multi-source data in Step 1 via the ArcGIS platform. The influence factors originate from four aspects: topographic elements include slope direction (as shown in Figures 8 and 9), slope degree (as shown in Figures 6 and 7), elevation (as shown in Figures 4 and 5), and curvature (as shown in Figures 10 and 11); geological elements include lithology (as shown in Figures 12 and 13); hydrological elements include distance from water systems (as shown in Figures 14 and 15); and land cover elements include land use (as shown in Figures 16 and 17).

[0026] Step 3: Build a landslide sample dataset.

[0027] We obtain landslide history data and landslide risk point data. Based on landslide history data and landslide risk point data, a landslide sample dataset will be constructed. Specifically, the number of landslide rasters within a rectangular area will be determined. If the number of landslide rasters within this area exceeds 50%, the area will be marked as a landslide point; if it is less than 50%, it will be marked as a non-landslide point. The experimental dataset was randomly divided into a training set and a validation set in an 8:2 ratio, where landslide samples were labeled 1 and non-landslide samples were labeled 0.

[0028] Step 4: Perform data augmentation on the landslide sample dataset.

[0029] The specific steps for augmenting the landslide sample dataset are as follows:

[0030] In the landslide sample dataset, identify the rasters containing landslide features. JPEG0007906342000025.jpg27170JPEG0007906342000026.jpg9170JPEG0007906342000027.jpg15170 The generated extended samples are added to the experimental dataset to improve the generalization ability and robustness of the model.

[0031] Step 5: Launch and train the UNet deep learning model.

[0032] We design a UNet-based deep learning model to ensure better extraction of landslide image features. Specifically, we use the LeakyRelU activation function, and its formula is as follows: JPEG0007906342000028.jpg14170JPEG0007906342000029.jpg136170JPEG0007906342000030.jpg23170

[0033] We process the convolution boundary using reflection padding, and the formula for reflection padding is: JPEG0007906342000031.jpg27170

[0034] The binary cross-entropy loss function is used to account for the difference between the probability distribution predicted by the model and the actual distribution. Specifically, the loss function is: JPEG0007906342000032.jpg16170JPEG0007906342000033.jpg25170

[0035] JPEG0007906342000034.jpg21170

[0036] Step 6: Establish a secure federated learning framework to ensure data privacy and security.

[0037] Specifically, a secure communication protocol is established between the two data holders, federated learning and training are performed using landslide data to obtain updated local UNet gradient parameters, and during the training process, the equation for updating the gradient parameters of the global UNet model is: JPEG0007906342000035.jpg17170JPEG0007906342000036.jpg49170 Throughout the entire process, the servers in Wenchuan County and Beichuan Qiang Autonomous County will always retain the original data and transmit only encrypted model parameters and gradient information to ensure data privacy and security.

[0038] Step 7: Applying the model and outputting the results. JPEG0007906342000037.jpg84170

[0039] Table 1 statistically shows the distribution of each level in the regional probability map predicted by the two training methods. In the prediction results from distributed training in Wenchuan County and Beichuan Qiang Autonomous County, distributed training showed a certain advantage in assessing landslide probability. In particular, in Beichuan County, distributed training significantly improved the prediction accuracy in areas with low probability and reduced misjudgments in areas with medium to high probability. This is because distributed training can better utilize the data features of each sub-domain and improve the generalization ability of the model, thereby improving the overall accuracy and reliability of the predictions.

[0040] JPEG0007906342000038.jpg56170

[0041] Step 8: Create a probability diagram.

[0042] The specific criteria for forming a distribution map of the likelihood of initial landslides occurring, based on probability values, are as follows:

[0043] (1) Probability value < 50%, representing a low level of likelihood of occurrence. (2) 50% < probability value < 70%, representing a moderate level of probability. (3) A probability value > 70% represents a region with a high probability of occurrence. JPEG0007906342000039.jpg67170

[0044] An electronic device including memory, a processor, and a computer program stored in the memory and operable by the processor, wherein when the processor executes the computer program, it realizes a federated learning and UNet-based landslide probability assessment method performed by the computer.

[0045] A computer-readable storage medium in which a computer program is stored, and when this program is executed by a processor, it realizes a federated learning and UNet-based landslide probability evaluation method executed by the computer.

Claims

1. A computer-based federated learning and UNet-based method for evaluating the likelihood of landslides, The acquisition of landslide influencing factor data involves step S1, which involves collecting data on multiple landslide influencing factors and extracting data on multiple influencing factor layers in a landslide-prone area. Establishing a federative learning framework, step S2 involves establishing a federative learning framework among multiple data holders, The design of a UNet deep learning model, comprising step S3 of designing a UNet-based deep learning model, using it to evaluate the likelihood of landslides, inputting influencing factor layer data into the model, and performing feature extraction and segmentation, The process involves training and integrating parameters of a federative learning model, specifically step S4, which involves training a UNet-based deep learning model using a landslide sample dataset within a federative learning framework, integrating model parameters using a federative averaging algorithm, and updating the UNet-based global deep learning model. The process involves iterative associative learning and model optimization, specifically step S5, which optimizes the parameters of a UNet-based deep learning model through multi-round iterations of associative learning. Step S6 is a feature map generation and probabilistic regression of landslide-prone areas, which involves generating a feature map of landslide-prone areas using an optimized UNet-based deep learning model, performing probabilistic regression using a Sigmoid layer, and outputting an evaluation result of the landslide-prone area. Specifically, Optimization and application of a UNet model, step S61 of making predictions for areas prone to landslides using an optimized UNet-based deep learning model, and extracting features of the landslide area by multilayer convolution and manipulation, Step S6 includes a calculation of the likelihood of landslide occurrence, which involves calculating probability results using a sigmoid layer and predicting the corresponding probability value of landslide occurrence in the image of each pixel point, thereby generating a landslide likelihood map. Step S7 involves verifying the results and proposing disaster prevention and mitigation measures, which include verifying and assessing the results of the landslide probability evaluation, and providing a warning of landslide risk and proposals for disaster prevention and mitigation based on the evaluation results. Includes, The data of multiple landslide influencing factors in step S1 include topographic elements, geological elements, hydrological elements, land cover elements, and climatic elements, where, The aforementioned topographic elements include slope direction, slope degree, elevation, and curvature. The aforementioned geological elements include lithology, The aforementioned hydrological elements include the average annual precipitation and the distance from the river system. The aforementioned land cover elements include enhanced vegetation index, normalized differential vegetation index, surface reflectance, and land use. A method for evaluating the likelihood of landslides using federated learning and UNet, characterized in that the aforementioned climate elements include the annual average temperature and the ground surface temperature.

2. The method for evaluating the likelihood of a landslide occurring, according to claim 1, characterized in that the multiple influencing factor layer data in step S1 include elevation, slope, slope direction, curvature, lithology, distance from a water system, and land use.

3. The construction of the aforementioned landslide sample dataset is as follows: The data collection process includes the steps of obtaining landslide history data and landslide risk point data, The construction of a landslide sample dataset involves constructing a landslide sample dataset based on landslide history data and landslide risk point data, determining the number of landslide rasters within a rectangular area, marking the area as a landslide point if the number of landslide rasters exceeds 50%, marking it as a non-landslide point if the number is less than 50%, and defining the marked landslide sample dataset as an experimental dataset. The distinction between training and validation sets involves randomly dividing the experimental dataset into training and validation sets in an 8:2 ratio, where landslide samples are labeled as 1 and non-landslide samples are labeled as 0. This is a data augmentation process, which involves the step of performing data augmentation on the constructed landslide sample data. The method for evaluating the likelihood of a landslide occurring, characterized by comprising associative learning and UNet-based learning according to claim 2.

4. The aforementioned data augmentation is, Identification of landslide feature rasters, comprising the steps of identifying rasters of landslide points in a landslide sample dataset, An application of Mixup data augmentation technology, in which the rasters of identified landslide points are randomly combined using Mixup data augmentation technology to generate augmented samples, Adding an extended sample to an experimental dataset, comprising the steps of adding the generated extended sample to the experimental dataset and The method for evaluating the likelihood of a landslide occurring, characterized by comprising associative learning and UNet-based learning according to claim 3.

5. The parameter updates for the aforementioned UNet-based global deep learning model are performed as follows:

6. The training of the aforementioned UNet-based deep learning model is as follows: Setting the activation function, selecting LeakyReLU as the activation function and using it for each convolutional layer in a UNet-based deep learning model. Model training involves applying weight-based learning in the convolutional layers. Regularization by Dropout, which uses Dropout technology to prevent overfitting. Reflection padding, which processes the convolution boundary using reflection padding, This is a model validation that uses a binary cross-entropy loss function to consider the difference between the probability distribution predicted by the model and the actual distribution.

7. The aforementioned warnings about landslide risks and proposals for disaster prevention and mitigation are, The process involves formulating disaster prevention and mitigation proposals, which include formulating appropriate proposals for different risk warning areas, proposing enhanced monitoring, personnel evacuation, and restrictions on development activities for red warning areas, proposing the implementation of construction and maintenance measures and enhanced patrols for yellow warning areas, and proposing the deployment of regular monitoring and periodic assessments for green warning areas. The method for evaluating the likelihood of a landslide occurring, characterized by comprising the associated learning and UNet-based methods described in claim 1.

8. An electronic device comprising memory, a processor, and a computer program stored in the memory and operable by the processor, wherein when the processor executes the computer program, it realizes the federated learning and UNet-based landslide probability evaluation method described in any one of claims 1 to 7.

9. A computer-readable storage medium in which a computer program is stored, characterized in that when the program is executed by a processor, it realizes the federated learning and UNet-based landslide probability evaluation method described in any one of claims 1 to 7.

Images