A full-automatic disaster-prone evaluation method and system based on PU learning
By using a fully automated disaster susceptibility assessment method based on PU learning, buffer thresholds and model parameters are automatically determined, solving the problem of relying on manual experience in existing technologies and realizing the automation and high efficiency of geological disaster susceptibility assessment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA UNIV OF GEOSCIENCES (WUHAN)
- Filing Date
- 2026-02-02
- Publication Date
- 2026-05-12
AI Technical Summary
Existing geological hazard susceptibility assessment schemes rely on manual experience and repeated experiments, which are complicated and inefficient.
A fully automated disaster susceptibility assessment method based on PU learning is adopted. By constructing a disaster susceptibility assessment model, balancing positive and unlabeled sample sets, and combining Bayesian optimization and PU learning algorithms, the buffer threshold and model parameters are automatically determined to achieve fully automated assessment.
It has achieved automation and high efficiency in the assessment of geological hazard susceptibility, reduced human intervention, avoided the experience setting and repeated experiments in traditional methods, and improved the stability and accuracy of the model.
Smart Images
Figure CN121638683B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of disaster assessment technology, specifically to a fully automated disaster susceptibility assessment method and system based on PU learning. Background Technology
[0002] Geological disasters such as landslides, collapses, and debris flows are often accompanied by enormous destructive power, causing significant socio-economic losses. Geological disasters are characterized by their high frequency, wide distribution, and difficulty in prevention; therefore, conducting geological disaster susceptibility assessments (i.e., assessing the risk level of geological disasters in a given area) is extremely important. Geological disaster susceptibility assessment refers to the comprehensive analysis of the impact of various disaster-causing factors on the occurrence of geological disasters under specific geological environmental conditions, thereby determining the likelihood and tendency of geological disasters to occur. Existing geological disaster susceptibility assessment schemes mostly rely on manual experience and repeated experiments, which are cumbersome and inefficient. Summary of the Invention
[0003] The main objective of this invention is to provide a fully automated disaster susceptibility assessment method and system based on PU learning, which aims to solve the problem that existing geological disaster susceptibility assessment schemes rely heavily on manual experience and repeated experiments, resulting in complicated steps and low efficiency.
[0004] The technical solution proposed in this invention is as follows:
[0005] A fully automated disaster susceptibility assessment method based on PU learning is applied to a fully automated disaster susceptibility assessment system based on PU learning; the system includes a processing module and a storage module that are communicatively connected to each other; the method includes:
[0006] The processing module obtains raw geological disaster data from the storage module;
[0007] The processing module obtains a positive sample set and an unlabeled sample set based on the original geological disaster data;
[0008] The processing module balances the positive sample set and the unlabeled sample set, and divides them into a training set and a test set;
[0009] The processing module is based on the PU learning concept to construct and train disaster susceptibility assessment models that combine different artificial intelligence models with positive samples and unlabeled learning.
[0010] The processing module adaptively fine-tunes the disaster susceptibility assessment model and tests the model using a test set to obtain the optimal disaster susceptibility assessment model.
[0011] The processing module outputs a raster result of the disaster susceptibility level of the target area based on the sample data of the target area and the optimal disaster susceptibility evaluation model.
[0012] Preferably, the original geological disaster data includes historical geological disaster events and basic geological environmental factors, and the original geological disaster data is obtained through satellite remote sensing interpretation and field investigation.
[0013] Preferably, the historical geological disaster events include landslides, collapses, debris flows, and ground subsidence, including one or more of landslides, collapses, debris flows, ground subsidence, ground fissures, or ground subsidence; the basic geological environmental factors include elevation, slope, aspect, curvature, topographic roughness, average annual precipitation, normalized difference vegetation index, engineering rock group, roads, faults, water systems, and land use types.
[0014] Preferably, the processing module obtains a positive sample set and an unlabeled sample set based on the original geological disaster data, including:
[0015] The processing module constructs a geospatial dataset point vector file;
[0016] The processing module extracts historical geological disaster events and basic geological environmental factors into geospatial dataset point vector files.
[0017] Preferably, the processing module extracts historical geological disaster events and basic geological environmental factors into a geospatial dataset point vector file, and then further includes:
[0018] The processing module reads the geospatial dataset point vector file. The attribute table of the geospatial dataset point vector file contains a column of basic geological environment factors. In order to distinguish between geological disaster points and non-disaster points, the factor of geological disaster points is assigned a value of 1, and the factor of non-disaster points is assigned a value of 0.
[0019] The processing module constructs a buffer for geological hazard points in the geospatial dataset point vector file, marks all samples in the buffer as positive sample sets, and marks the samples in the remaining areas as unlabeled sample sets.
[0020] Preferably, the processing module balances the positive sample set and the unlabeled sample set, and divides them into a training set and a test set, including:
[0021] The processing module normalizes the basic environmental factors, uses Pearson correlation analysis to remove basic environmental factors with a correlation greater than or equal to a preset value, and outputs a correlation heatmap.
[0022] The processing module uses an undersampling method to balance the ratio of positive sample set to unlabeled sample set, and then divides the positive sample set and unlabeled sample set into training set and test set according to a preset ratio.
[0023] Preferably, the processing module adaptively fine-tunes the disaster susceptibility assessment model and tests the model using a test set to obtain the optimal disaster susceptibility assessment model, including:
[0024] The processing module introduces a Bayesian optimization algorithm to adaptively fine-tune the parameters of the disaster susceptibility assessment model. The optimization parameters in the adaptive fine-tuning process vary depending on the model type: for positive sample and unlabeled ensemble learning models, the optimization parameters include the buffer threshold, the relevant parameters of the PU-Bagging positive sample and unlabeled ensemble learning algorithm, and the structure and training parameters of the corresponding base learner; for deep learning models based on prior probability PU loss, positive samples and unlabeled samples are directly processed through end-to-end training, and the optimization parameters include the buffer threshold, PU learning prior probability, and network structure and training parameters.
[0025] After a preset number of iterations, the processing module outputs the optimal parameter configuration for the disaster susceptibility assessment model.
[0026] The processing module selects the best-performing disaster susceptibility assessment model by comparing the weighted comprehensive scores of the 5-fold cross-validation of the disaster susceptibility assessment models, and uses it as the optimal disaster susceptibility assessment model.
[0027] Preferably, the processing module outputs a raster result of the disaster susceptibility level of the target area based on sample data of the target area using the optimal disaster susceptibility assessment model, including:
[0028] The processing module inputs the sample data of the target area into the optimal disaster susceptibility evaluation model, first obtaining the susceptibility probability point vector result of the target area, and then dividing it into different susceptibility levels according to the natural breakpoint method, which is the final output disaster susceptibility evaluation result. The susceptibility levels include extremely high susceptibility, high susceptibility, medium susceptibility, low susceptibility and extremely low susceptibility.
[0029] The present invention also proposes a fully automated disaster susceptibility assessment system based on PU learning, which applies a fully automated disaster susceptibility assessment method based on PU learning; the system includes a processing module and a storage module that are communicatively connected to each other.
[0030] The above technical solution can achieve the following beneficial effects:
[0031] The fully automated disaster susceptibility assessment method based on PU learning proposed in this invention addresses the problem that existing geological disaster susceptibility assessment schemes often rely on manual experience-based settings and repeated trials, resulting in cumbersome and inefficient procedures. Firstly, it enables adaptive optimization of the buffer threshold. In grid-based susceptibility assessment, the selection of the buffer threshold is incorporated into a Bayesian optimization mechanism, automatically determining the optimal buffer size. This ensures that the points within the buffer fully encompass the disaster area while avoiding the introduction of excessive boundary samples. Furthermore, it achieves full automation of the disaster susceptibility assessment. Reducing manual workload and eliminating the need for human intervention, it replaces the manual parameter tuning process relying on experience-based settings and repeated trials in traditional methods. After data input, it automatically completes factor screening, buffer construction, sample balancing, optimal parameter adjustment, model training, and optimal model selection, ultimately directly outputting the susceptibility assessment results.
[0032] Furthermore, the proposed solution overcomes the errors caused by the random selection of negative samples in traditional supervised methods. It employs a semi-supervised strategy based on the PU learning concept to automatically identify pseudo-negative samples from unlabeled samples and improves model stability through multiple rounds of adaptive sampling. Specifically, it trains various base classifiers T covering ensemble learning and deep learning methods, and adopts corresponding PU learning strategies for different types of classifiers: for Bagging and Boosting ensemble learning methods, it constructs multi-round sub-models within the PU-Bagging algorithm framework to identify reliable pseudo-negative samples from unlabeled samples; for deep learning methods, it constructs a PU loss function based on prior probability and directly processes positive and unlabeled samples through end-to-end training. Based on the above strategies, PU-T models based on different base classifiers T are constructed. Attached Figure Description
[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.
[0034] Figure 1 This is a flowchart illustrating the first embodiment of a fully automated disaster susceptibility assessment method based on PU learning proposed in this invention. Detailed Implementation
[0035] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.
[0036] This invention proposes a fully automated disaster susceptibility assessment method and system based on PU learning.
[0037] As attached Figure 1 As shown, in the first embodiment of the fully automated disaster susceptibility assessment method based on PU learning proposed in this invention, this fully automated disaster susceptibility assessment method based on PU learning is applied to a fully automated disaster susceptibility assessment system based on PU learning; the system includes a processing module and a storage module that are communicatively connected to each other; this embodiment includes the following steps:
[0038] Step S110: The processing module obtains the original geological disaster data from the storage module.
[0039] Step S120: The processing module obtains a positive sample set and an unlabeled sample set based on the original geological disaster data.
[0040] Specifically, based on geographic spatial distance, a buffer is set for positive samples to obtain buffer positive samples. The Bayesian optimization algorithm is used to search for the optimal threshold of the buffer within a set range, so that the points in the buffer can both contain the disaster range as much as possible and avoid introducing too many boundary samples to reduce noise.
[0041] Step S130: The processing module balances the positive sample set and the unlabeled sample set and divides them into a training set and a test set.
[0042] Specifically, a buffer is constructed, where all points within the buffer are considered positive samples and all points outside the buffer are considered unlabeled samples. Based on the PU (Positive-Unlabeled) learning concept, pseudo-negative samples are identified from the unlabeled samples.
[0043] Specifically, the disaster susceptibility assessment model here is the PU-T model; the PU model, or Positive-and-Unlabeled-Learning model, is a semi-supervised binary classification method in machine learning; the PU-T model is a PU model that includes a base classifier T. The base classifier T covers the mainstream architectures of machine learning and deep learning, specifically including the tree-based PU-RF model, the gradient boosting-based PU-XGB model, the neural network-based PU-MLP model, and the convolutional neural network-based PU-CNN model.
[0044] The PU-bagging algorithm is a PU learning method based on the ensemble concept. It adaptively samples and identifies pseudo-negative samples on an unlabeled dataset U. It aims to build a robust classification model in the absence of explicit negative samples by introducing a bootstrap sampling (Bagging) mechanism.
[0045] The main process of the PU-bagging algorithm is as follows:
[0046] S131: Input positive sample set P, unlabeled sample set U;
[0047] S132: Randomly select a subset from U Combined with P to form the training set ;
[0048] S133: Assuming all For negative samples, a base learner is trained. In the Bagging-PU-learning framework, the base learner has generality. The Bagging-PU-learning framework is a semi-supervised learning framework that combines the Bagging ensemble method with Positive-Unlabeled (PU) learning.
[0049] S134: Repeat the above process B times (80 times) to obtain several base learners. ;
[0050] S135: Vote or average the prediction results of multiple base learners to obtain the final prediction probability: .
[0051] S136: If some unlabeled samples are judged as negative samples by most base learners, they can be regarded as pseudo-negative samples, and the remaining samples remain unlabeled.
[0052] Step S140: The processing module constructs and trains disaster susceptibility assessment models based on the PU learning concept, combining different artificial intelligence models that combine positive samples and unlabeled learning.
[0053] Specifically, for Bagging (Bootstrap-Aggregating), Boosting (a machine learning algorithm that can be used to reduce bias in supervised learning), ensemble learning and MLP (Multilayer-Perceptron) neural network methods, the PU-bagging algorithm is used to randomly draw samples from unlabeled samples U. Multi-round sub-models are constructed with the help of BaggingPUClassifier. For deep learning models, the PU loss function is constructed. The target value of positive samples is 1, which means that they are sure to be positive samples. The target value of unlabeled samples is the prior probability (the default value is 0.1). Unlabeled samples have a 10% probability that they are potential positive samples.
[0054] Specifically, the PU-bagging algorithm is an ensemble learning method specifically designed to handle Positive-Unlabeled (PU) learning problems. The BaggingPUClassifier mentioned above is a PU learning (Positive-Unlabeled-learning) classifier based on the Bagging ensemble learning framework.
[0055] Step S150: The processing module adaptively fine-tunes the disaster susceptibility assessment model and tests the disaster susceptibility assessment model through a test set to obtain the optimal disaster susceptibility assessment model.
[0056] Step S160: The processing module outputs the disaster susceptibility level raster result of the target area based on the sample data of the target area through the optimal disaster susceptibility evaluation model.
[0057] Specifically, the aforementioned disaster susceptibility probability vector points represent the probability of geological disaster occurrence at each vector point in the target area. The processing module converts the disaster susceptibility probability points in the target area into raster results with the same resolution as the basic geological environmental factors, and divides these raster results into five levels based on the natural breakpoint method: extremely high susceptibility, high susceptibility, moderate susceptibility, low susceptibility, and extremely low susceptibility, ultimately obtaining the geological disaster susceptibility results for the target area.
[0058] The fully automated disaster susceptibility assessment method based on PU learning proposed in this invention addresses the problem that existing geological disaster susceptibility assessment schemes often rely on manual experience-based settings and repeated trials, resulting in cumbersome and inefficient procedures. Firstly, it enables adaptive optimization of the buffer threshold. In grid-based susceptibility assessment, the selection of the buffer threshold is incorporated into a Bayesian optimization mechanism, automatically determining the optimal buffer size. This ensures that the points within the buffer fully encompass the disaster area while avoiding the introduction of excessive boundary samples. Furthermore, it achieves full automation of the disaster susceptibility assessment. Reducing manual workload and eliminating the need for human intervention, it replaces the manual parameter tuning process relying on experience-based settings and repeated trials in traditional methods. After data input, it automatically completes factor screening, buffer construction, sample balancing, optimal parameter adjustment, model training, and optimal model selection, ultimately directly outputting the susceptibility assessment results.
[0059] Furthermore, the proposed solution overcomes the errors caused by the random selection of negative samples in traditional supervised methods. It employs a semi-supervised strategy based on a Bagging ensemble PU learning algorithm to automatically identify pseudo-negative samples from unlabeled samples and improves model stability through multiple rounds of adaptive sampling. Specifically, it trains various base classifiers T encompassing ensemble learning and deep learning methods, and adopts corresponding PU learning strategies for different types of classifiers: for Bagging and Boosting ensemble learning methods, multiple rounds of sub-models are constructed within the PU-Bagging algorithm framework to identify reliable pseudo-negative samples from unlabeled samples; for deep learning methods, a PU loss function based on prior probability is constructed, and positive and unlabeled samples are directly processed through end-to-end training. Based on the above strategies, PU-T models based on different base classifiers T are constructed respectively.
[0060] In the second embodiment of the fully automated disaster susceptibility assessment method based on PU learning proposed in this invention, based on the first embodiment, the original geological disaster data includes historical geological disaster events and basic environmental factors, and the original geological disaster data is obtained through satellite remote sensing interpretation and field investigation.
[0061] In the third embodiment of the fully automated disaster susceptibility assessment method based on PU learning proposed in this invention, based on the second embodiment, the historical geological disaster events include one or more of landslides, collapses, debris flows, ground subsidence, ground fissures, or ground settlement; the basic environmental factors include elevation, slope, aspect, curvature, topographic roughness, annual precipitation, normalized vegetation index, engineering rock group, roads, faults, water systems, and land use rate.
[0062] In the fourth embodiment of the fully automated disaster susceptibility assessment method based on PU learning proposed in this invention, based on the third embodiment, step S120 includes the following steps:
[0063] Step S410: The processing module constructs a geospatial dataset point vector file.
[0064] Step S420: The processing module uses the multi-value extraction to point tool of ArcGIS software to extract historical geological disaster events and basic environmental factors into a geospatial dataset point vector file.
[0065] Specifically, historical geological disaster events and basic environmental factors were graded and assigned values in ArcGIS software. It should be noted that a total of 131 historical disaster sites were obtained through a combination of satellite remote sensing interpretation and field surveys. The digital elevation model (DEM) data came from the Geospatial Data Cloud website, the vegetation cover was calculated from Landsat 8 data downloaded from the Geospatial Data Cloud website, the stratigraphic and structural data came from 1:50,000 geological maps, the land use data came from the 30-meter resolution annual land cover data of China from 1985 to 2024 published by Professors Yang Jie and Huang Xin of Wuhan University, the road and water system data came from the OSM website, and the annual average rainfall came from meteorological station observation data.
[0066] In the fifth embodiment of the fully automated disaster susceptibility assessment method based on PU learning proposed in this invention, based on the fourth embodiment, after step S420, the following steps are further included:
[0067] Step S510: The processing module reads the geospatial dataset point vector file. The attribute table of the geospatial dataset point vector file contains a basic geological environment factor column. In order to distinguish between geological disaster points and non-disaster points, the factor of geological disaster points is assigned a value of 1, and the factor of non-disaster points is assigned a value of 0.
[0068] Step S520: The processing module constructs a buffer for geological hazard points in the geospatial dataset point vector file, marks all samples in the buffer as positive sample sets, and marks the samples in the remaining areas as unlabeled sample sets.
[0069] In the sixth embodiment of the fully automated disaster susceptibility assessment method based on PU learning proposed in this invention, based on the second embodiment, step S130 includes the following steps:
[0070] Step S610: The processing module normalizes the basic environmental factors, removes basic environmental factors with a correlation greater than or equal to a preset value (e.g., 0.8) using Pearson correlation analysis, and outputs a correlation heatmap (specifically a 600dpi PNG image).
[0071] The processing module normalizes the basic environmental factors and uses Pearson correlation analysis to remove basic environmental factors with a correlation greater than or equal to a preset value (e.g., 0.8).
[0072] Specifically, the geological hazard susceptibility assessment model is trained by retaining only the basic environmental factors that are relatively independent.
[0073] Step S620: The processing module uses an undersampling method to balance the ratio of the positive sample set to the unlabeled sample set, and then divides the positive sample set and the unlabeled sample set into a training set and a test set according to a preset ratio (e.g., 8:2).
[0074] In the seventh embodiment of the fully automated disaster susceptibility assessment method based on PU learning proposed in this invention, based on the first embodiment, step S150 includes the following steps:
[0075] Step S710: The processing module introduces a Bayesian optimization algorithm to adaptively fine-tune the parameters of the disaster susceptibility assessment model. The optimization parameters in the adaptive fine-tuning process vary depending on the model type: For ensemble learning models based on positive samples and unlabeled ensemble learning algorithms, the optimization parameters include the buffer threshold, the relevant parameters of PU-Bagging positive samples and unlabeled ensemble learning algorithms, and the structure and training parameters of the corresponding base learners. For deep learning models based on prior probability PU loss, positive samples and unlabeled samples are directly processed through end-to-end training, and the optimization parameters include the buffer threshold, PU learning prior probability, and network structure and training parameters.
[0076] Specifically, on the one hand, for ensemble learning models based on positive samples and unlabeled ensemble learning algorithms (PU-Bagging), including PU-RF (a machine learning method combining random forests within the framework of positive unlabeled learning (PU learning), PU-XGB (an efficient machine learning algorithm based on the gradient boosting framework), and PU-MLP (a machine learning model combining PU learning and multilayer perceptrons), the parameters of the buffer threshold, relevant parameters of PU-Bagging positive samples and unlabeled ensemble learning algorithms, and the structure and training parameters of the corresponding base learners are optimized. On the other hand, for deep learning models based on prior probability PU loss, positive samples and unlabeled samples are directly processed through end-to-end training, and the parameters of the buffer threshold, PU learning prior probability, network structure, and training parameters are optimized.
[0077] Specifically, the above method calculates the composite score (AUC+F1) / 2 using 5-fold cross-validation, and uses "cross-validation weighted score = mean of composite score - 0.2 × standard deviation of composite score" as the optimization objective.
[0078] Step S720: After a preset number of iterations (50 times), the processing module outputs the optimal parameter configuration of the disaster susceptibility assessment model. The optimal parameter configuration includes the buffer threshold, PU model parameters, and hyperparameters of the four weak classifiers.
[0079] Step S730: The processing module selects the best-performing disaster susceptibility assessment model by comparing the weighted comprehensive scores of the 5-fold cross-validation of the disaster susceptibility assessment models, and uses it as the optimal disaster susceptibility assessment model.
[0080] Specifically, it should be noted that the Bayesian optimization algorithm (BO) in step S810 is mainly used for hyperparameter tuning of machine learning and deep learning models. The algorithm's goal is to find the global optimum of the objective function. It finds the optimal combination of hyperparameters by constructing a probabilistic surrogate model of the objective function with as few evaluations as possible.
[0081] Bayesian optimization, as the core module, realizes integrated automatic optimization of buffer threshold selection, PU learning parameter configuration and base learner parameter tuning. The core optimization process adopts the Bayesian optimization algorithm gp_minimize based on Gaussian process to approximate the objective function, and selects the next set of parameter combinations according to the acquisition function in each iteration, thereby obtaining the optimal result in 50 iterations.
[0082] In the composite score (AUC+F1) / 2:
[0083] ,
[0084] ,
[0085] ,
[0086] ,
[0087] The meanings of each parameter in the formula are shown in Table 1 below:
[0088] Table 1. Meaning of each hyperparameter in adaptive fine-tuning
[0089]
[0090] It should be noted that the buffer threshold corresponds to the hyperparameter buffer_dist. Different models have different Bayesian optimization hyperparameters, and the meanings of each Bayesian optimization hyperparameter are shown in Table 2 below:
[0091] Table 2. Meaning of each Bayesian optimization hyperparameter
[0092]
[0093] Continued from Table 2
[0094]
[0095] Finally, based on the optimal PU-T model and the corresponding optimal buffer threshold and classifier parameters, the disaster susceptibility level raster results are directly output to complete the spatial prediction of regional geological disaster susceptibility.
[0096] In the eighth embodiment of the fully automated disaster susceptibility assessment method based on PU learning proposed in this invention, based on the seventh embodiment, step S160 includes the following steps:
[0097] Step S810: The processing module inputs the sample data of the target area into the optimal disaster susceptibility assessment model. First, it obtains the susceptibility probability point vector result of the target area. Then, it converts the susceptibility probability point vector result into a raster result with the same resolution as the geological environmental factors, and divides it into different susceptibility levels according to the natural breakpoint method. This is the final output disaster susceptibility assessment result, where the susceptibility levels include extremely high susceptibility, high susceptibility, moderate susceptibility, low susceptibility, and extremely low susceptibility.
[0098] Custom drawings will then be created manually based on the requirements.
[0099] The present invention also proposes a fully automated disaster susceptibility assessment system based on PU learning, which applies a fully automated disaster susceptibility assessment method based on PU learning; the system includes a processing module and a storage module that are communicatively connected to each other.
[0100] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.
[0101] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.
Claims
1. A fully automated disaster susceptibility assessment method based on PU learning, characterized in that, An application to a fully automated disaster susceptibility assessment system based on PU learning; the system includes a processing module and a storage module that are communicatively connected to each other; the method includes: The processing module obtains raw geological disaster data from the storage module; The processing module obtains a positive sample set and an unlabeled sample set based on the original geological disaster data. Based on the geographic spatial distance, a buffer is set for the positive samples to obtain buffer positive samples. The optimal threshold of the buffer is searched within a set range using a Bayesian optimization algorithm. The processing module balances the positive sample set and the unlabeled sample set, and divides them into a training set and a test set; The processing module, based on the PU learning concept, constructs and trains different artificial intelligence models that combine positive samples and unlabeled learning. Specifically, a buffer is constructed, treating all points within the buffer as positive samples and all points outside the buffer as unlabeled samples. Based on the PU learning concept, false negative samples are identified from the unlabeled samples. The disaster susceptibility assessment model is a PU-T model. The PU-T model is a PU model that includes a base classifier T. The base classifier T specifically includes a tree-based PU-RF model, a gradient boosting-based PU-XGB model, a neural network-based PU-MLP model, and a convolutional neural network-based PU-CNN model. The processing module adaptively fine-tunes the disaster susceptibility assessment model and tests the model using a test set to obtain the optimal disaster susceptibility assessment model. The processing module outputs a raster result of the disaster susceptibility level of the target area based on the sample data of the target area and through the optimal disaster susceptibility evaluation model; The processing module adaptively fine-tunes the disaster susceptibility assessment model and tests it using a test set to obtain the optimal disaster susceptibility assessment model, including: The processing module introduces a Bayesian optimization algorithm to adaptively fine-tune the parameters of the disaster susceptibility assessment model. The optimization parameters in the adaptive fine-tuning process vary depending on the model type: for positive sample and unlabeled ensemble learning models, the optimization parameters include the buffer threshold, the relevant parameters of the PU-Bagging positive sample and unlabeled ensemble learning algorithm, and the structure and training parameters of the corresponding base learner; for deep learning models based on prior probability PU loss, positive samples and unlabeled samples are directly processed through end-to-end training, and the optimization parameters include the buffer threshold, PU learning prior probability, and network structure and training parameters. After a preset number of iterations, the processing module outputs the optimal parameter configuration for the disaster susceptibility assessment model. The processing module selects the best-performing disaster susceptibility assessment model by comparing the weighted composite scores of the 5-fold cross-validation of the disaster susceptibility assessment models. The formula for calculating the weighted composite score of the 5-fold cross-validation is: the mean of the composite scores. 0.2 × the standard deviation of the composite score. The formula for calculating the composite score is: (AUC+F1) / 2.
2. The fully automated disaster susceptibility assessment method based on PU learning according to claim 1, characterized in that, The original geological disaster data includes historical geological disaster events and basic geological environmental factors, and the original geological disaster data is obtained through satellite remote sensing interpretation and field investigation.
3. The fully automated disaster susceptibility assessment method based on PU learning according to claim 2, characterized in that, The historical geological disaster events include one or more of the following: landslides, collapses, debris flows, ground subsidence, ground fissures, or ground settlement; the basic geological environmental factors include elevation, slope, aspect, curvature, topographic roughness, annual precipitation, normalized difference vegetation index, engineering rock group, roads, faults, water systems, and land use types.
4. The fully automated disaster susceptibility assessment method based on PU learning according to claim 3, characterized in that, The processing module obtains a positive sample set and an unlabeled sample set based on the original geological disaster data, including: The processing module constructs a geospatial dataset point vector file; The processing module extracts historical geological disaster events and basic geological environmental factors into geospatial dataset point vector files.
5. The fully automated disaster susceptibility assessment method based on PU learning according to claim 4, characterized in that, The processing module extracts historical geological disaster events and basic geological environmental factors into a geospatial dataset point vector file, and then further includes: The processing module reads the geospatial dataset point vector file. The attribute table of the geospatial dataset point vector file contains a column of basic geological environment factors. In order to distinguish between geological disaster points and non-disaster points, the factor of geological disaster points is assigned a value of 1, and the factor of non-disaster points is assigned a value of 0. The processing module constructs a buffer for geological hazard points in the geospatial dataset point vector file, marks all samples in the buffer as positive sample sets, and marks the samples in the remaining areas as unlabeled sample sets.
6. The fully automated disaster susceptibility assessment method based on PU learning according to claim 2, characterized in that, The processing module balances the positive sample set and the unlabeled sample set, and divides them into a training set and a test set, including: The processing module normalizes the basic environmental factors, uses Pearson correlation analysis to remove basic environmental factors with a correlation greater than or equal to a preset value, and outputs a correlation heatmap. The processing module uses an undersampling method to balance the ratio of positive sample set to unlabeled sample set, and then divides the positive sample set and unlabeled sample set into training set and test set according to a preset ratio.
7. The fully automated disaster susceptibility assessment method based on PU learning according to claim 1, characterized in that, The processing module outputs a raster result of the disaster susceptibility level of the target area based on sample data of the target area and through the optimal disaster susceptibility assessment model, including: The processing module inputs the sample data of the target area into the optimal disaster susceptibility assessment model. First, it obtains the susceptibility probability point vector results of the target area. Then, it converts the susceptibility probability point vector results into raster results with the same resolution as the geological environmental factors, and divides them into different susceptibility levels according to the natural breakpoint method. This is the final output disaster susceptibility assessment result, where the susceptibility levels include extremely high susceptibility, high susceptibility, moderate susceptibility, low susceptibility, and extremely low susceptibility.
8. A fully automated disaster susceptibility assessment system based on PU learning, characterized in that, The system employs the fully automated disaster susceptibility assessment method based on PU learning as described in any one of claims 1-7; the system includes a processing module and a storage module that are communicatively connected to each other.