Geological disaster prediction method and system based on optimized pu-xgboost model
By optimizing the PU-XGBoost model, combining InSAR deformation data and the Spy sample strategy to screen negative samples, and using the BTO optimization algorithm to optimize hyperparameters, the problems of data acquisition difficulties and negative sample contamination in traditional geological disaster prediction are solved, and high-precision geological disaster susceptibility prediction and classification are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
- Filing Date
- 2026-04-07
- Publication Date
- 2026-07-31
AI Technical Summary
Traditional methods for assessing geological hazard susceptibility rely on field surveys and single statistical or machine learning models. These methods suffer from difficulties in data acquisition, low reliability of sample labels, weak model generalization ability, and negative sample contamination, making it difficult to achieve high-precision susceptibility prediction.
A geological hazard prediction method based on an optimized PU-XGBoost model is adopted. By introducing InSAR deformation data and historical geological hazard point data, combining the PU-Learning strategy of Spy samples to screen negative samples, and using the BTO optimization algorithm to optimize the XGBoost hyperparameters, a high-precision geological hazard prediction model is constructed.
It achieves high-precision identification of geological hazard hazard areas and pixel-by-pixel susceptibility classification mapping, providing a scientific basis for risk zoning and disaster prevention and mitigation decision-making, reducing the risk of negative sample contamination, and improving the model's predictive stability and accuracy.
Smart Images
Figure CN121980199B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of geological disaster monitoring and risk assessment technology, specifically to a geological disaster prediction method and system based on an optimized PU-XGBoost model. Background Technology
[0002] Traditional methods for assessing geological hazard susceptibility mainly rely on field geological surveys and single statistical or conventional machine learning models. However, these methods have limitations such as long data acquisition cycles, low sample label reliability, and weak model generalization ability. Field surveys require significant human and material resources, making it difficult to achieve large-scale, all-weather real-time monitoring, and can only obtain discrete point data. Single assessment models (such as logistic regression and support vector machines) often simply default to "non-landslide areas" (negative samples) when constructing training samples, ignoring the existence of potential hazard points in unknown areas, leading to a serious "negative sample contamination" problem. In addition, traditional methods focus on static environmental factors (such as slope and lithology), often failing to comprehensively consider the driving role of dynamic factors such as surface deformation in hazard development, making it difficult to achieve high-precision susceptibility prediction.
[0003] In recent years, with the development of InSAR (Interferometric Synthetic Aperture Radar) technology and artificial intelligence, existing methods still have the following problems in multi-source data fusion and model optimization: First, there is a lack of effective sample cleaning mechanisms, making it difficult to accurately identify reliable negative samples from unlabeled samples, resulting in bias in model training; second, it is difficult to optimize the hyperparameter combination of commonly used machine learning models, and traditional grid search or random search is inefficient and difficult to find the global optimum, resulting in insufficient accuracy and stability of prediction results; therefore, there is an urgent need for a geological disaster prediction method and system based on the optimized PU-XGBoost model to solve the above problems. Summary of the Invention
[0004] The purpose of this invention is to provide a geological hazard prediction method and system based on an optimized PU-XGBoost model to solve the problems existing in the background art.
[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: a geological hazard prediction method based on an optimized PU-XGBoost model, comprising the following steps:
[0006] Acquire multi-source data for the study area, including at least InSAR deformation monitoring data and historical geological hazard data, and preprocess the multi-source data;
[0007] Based on the preprocessed multi-source data, the positive sample set P is taken from historical geological disaster points and InSAR significant deformation points in InSAR deformation monitoring data, and the remaining data in the study area is taken as the unlabeled sample set U; the PU sample system is constructed by combining the positive sample set P and the unlabeled sample set U.
[0008] Based on the PU sample system, the negative sample set RN is selected from the unlabeled sample set U by introducing the PU-Learning strategy of Spy samples;
[0009] An XGBoost susceptibility evaluation model was established, and the BTO optimization algorithm was introduced to optimize and search for XGBoost hyperparameters. The optimal hyperparameter combination was obtained and the optimal model was trained using the positive sample set P and the negative sample set RN.
[0010] Based on the optimal model, the study area is used to predict and classify the susceptibility to geological disasters.
[0011] Preferably, the multi-source data further includes DEM data, rainfall and hydrological data, land cover and human engineering activity data; the preprocessing includes: performing SBAS-InSAR processing on the study area to output deformation rate and deformation accumulation raster; performing topographic factor derivation calculations on the DEM data; performing raster resampling on the rainfall data; and
[0012] All factors should be subjected to at least a unified coordinate system, unified resolution, and study area cropping.
[0013] Preferably, for the InSAR deformation time series in the InSAR deformation monitoring data, robust statistics are used to construct derived factors, and the derived factors include at least one of the following:
[0014] Mean deformation The calculation formula is as follows:
[0015] ;
[0016] Deformation standard deviation The calculation formula is as follows:
[0017] ;
[0018] absolute value of deformation rate The calculation formula is as follows:
[0019] ;
[0020] in, For the first Cumulative deformation over time For the number of phases, The mean, Standard deviation, and The start and end times of the monitoring period, This represents the absolute value of the deformation rate.
[0021] Preferably, the positive sample set P includes cataloged historical geological disaster points and InSAR significant deformation points located in slope areas with an annual deformation rate absolute value greater than 15 mm / year, monitored based on InSAR technology, and their target labels are assigned to 1; the unlabeled sample set U consists of all other sample data except for the positive sample set P, and all data in the unlabeled sample set U are uniformly assigned an initial label of 0.
[0022] Preferably, the PU-Learning strategy that introduces Spy samples is as follows:
[0023] Randomly select Spy samples from the positive sample set P, temporarily relabel them as negative samples, and mix them into the unlabeled sample set. In the process, a hybrid dataset is formed. ;
[0024] Predict the class probability distribution of all Spy samples using a classifier, and calculate the class probability distribution. quantiles as a screening threshold ;
[0025] Traverse the unlabeled sample set Unlabeled sample set Medium sample prediction probability The samples retained constitute the negative sample set RN.
[0026] Preferably, according to a preset ratio Randomly draw Spy samples from the positive sample set P;
[0027] A preset ratio of positive sample set P to negative sample set RN is used, and the predicted probability is based on this ratio. Select the sample with the lowest probability The samples constitute the negative sample set RN.
[0028] Preferably, the model parameters are optimized by minimizing the objective function that includes a regularization term; the BTO optimization algorithm is introduced to optimize the XGBoost hyperparameters, specifically as follows:
[0029] Map the XGBoost hyperparameter combinations to the walls of a barrel.
[0030] Define the i-th hyperparameter combination in the population. The adaptability is the water holding capacity of the wooden barrel. :
[0031] ;
[0032] Calculate the normalized values for each dimension to identify the bottleneck dimension limiting the current water capacity. :
[0033] ;
[0034] In the formula, For the first The individual The possible values for the dimension; and The populations at the 1st Maximum and minimum boundary values in a dimension; For dimension number, A function for calculating fitness;
[0035] Focusing only on the shortcomings Perform targeted repair, and calculate the function based on fitness. The individual with the best index found in the current iteration is taken as the global optimal solution. Using the global optimal solution By guiding and introducing random cooperative perturbations, the repair formula satisfies:
[0036] ;
[0037] In the formula, As a model evaluation metric, , For learning and collaboration coefficients, , It is a random number. For the corresponding dimension values of random reference individuals, This represents the number of iterations. For the first Individuals in the dimension of shortcomings The value that can be taken on.
[0038] Preferably, the geological hazard susceptibility of the study area is predicted based on the optimal model, and a geological hazard susceptibility probability grid of the study area is output; the susceptibility probability grid is classified into high susceptibility, medium susceptibility, low susceptibility and low susceptibility areas according to the natural breakpoint method.
[0039] This invention also discloses a geological hazard prediction system based on an optimized PU-XGBoost model, comprising:
[0040] Data acquisition module: used to download multi-source data, including InSAR deformation monitoring data, historical geological disaster point data, DEM data, rainfall and hydrological data, land cover and human engineering activity data;
[0041] Preprocessing module: Used to preprocess multi-source data, including unifying coordinate system and resolution, cropping the study area, handling invalid and missing values, and standardization.
[0042] Spy-PU sample construction module: Constructs a positive sample set P and an unlabeled sample set U, and introduces Spy samples to dynamically determine the screening threshold. It filters out the negative sample set RN from the unlabeled sample set U, and establishes the PU sample system by combining the positive sample set P and the negative sample set RN.
[0043] Model training and optimization module: The XGBoost susceptibility assessment model is trained based on the PU sample system, and the hyperparameters are optimized and searched using the BTO optimization algorithm to obtain the optimal model;
[0044] Output module: Used to output the probability raster and hierarchical zoning results of susceptibility.
[0045] Preferably, the BTO optimization object in the model training and optimization module includes at least one or more of the following: XGBoost learning rate, tree depth, subsampling rate, column sampling rate, minimum leaf sample weight, regularization coefficient, and number of base learners.
[0046] Beneficial effects: This invention introduces InSAR deformation results as dynamic prior information and fuses them with static environmental factors. It also employs a PU-Learning strategy that incorporates Spy samples to dynamically determine the screening threshold, thereby reducing the risk of negative sample contamination by screening reliable negative samples from unlabeled samples. This solves the problems of negative sample contamination and label uncertainty caused by traditional random sampling of negative samples. Combined with the BTO optimization algorithm, it performs efficient global optimization of XGBoost hyperparameters, achieving high-precision identification and pixel-by-pixel susceptibility classification mapping of geological hazard areas. This provides a scientific basis for geological hazard risk zoning and disaster prevention and mitigation decisions in the study area, and has good engineering application value and scalability. Attached Figure Description
[0047] Figure 1 The flowchart of the geological disaster prediction method based on the optimized PU-XGBoost model of the present invention is shown in the figure.
[0048] Figure 2 This is a schematic diagram illustrating the principle of reliable negative sample screening for PU-Learning based on the Spy mechanism of the present invention.
[0049] Figure 3 The graphs show the comparison of the prediction performance of different models in the geological hazard susceptibility assessment task in this invention, where (a) is a comparison graph of ROC curves and (b) is a comparison graph of PR curves. Detailed Implementation
[0050] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.
[0051] Example, reference Figure 1 As shown, a geological hazard prediction method based on an optimized PU-XGBoost model is provided, including the following steps:
[0052] Step 1: SBAS-InSAR processing is performed on the study area to obtain deformation data, and historical geological disaster points, topographic data, rainfall and hydrological data, land cover and human engineering activity data are acquired and preprocessed.
[0053] In one embodiment, multi-temporal Sentinel-1 radar imagery data (IW mode, VV polarization) covering the study area is downloaded and processed using SARscape software for SBAS-InSAR. The specific processing flow includes: connection map construction, interferometric workflow, first inversion (optimization and re-flattening, estimation of rate and terrain residuals, estimation of average intensity map and accuracy, and secondary unwrapping), second inversion (optimization and re-flattening, estimation of deformation components, removal of atmospheric components, and coherence and other calculations) and geocoding.
[0054] To reduce the impact of noise and outliers in SAR data on the deformation factor, robust statistics are used to construct a derived factor, which includes at least one of the following:
[0055] Mean deformation : Reflects the long-term overall deformation trend of the region;
[0056] ;
[0057] Deformation standard deviation The standard deviation reflects the dispersion and instability of surface deformation; the larger the standard deviation, the more intense the geological activity.
[0058] ;
[0059] absolute value of deformation rate : Reflects the degree of deformation.
[0060] ;
[0061] In the formula, For the first Cumulative deformation over time For the number of phases, The mean, Standard deviation, and The start and end times of the monitoring period, This represents the absolute value of the deformation rate.
[0062] Simultaneously, historical geological disaster data, DEM data (for extracting slope, aspect, curvature, and topographic relief), meteorological and hydrological data (annual average rainfall, distance from water systems), land cover data (NDVI, land use type), and human engineering activity data (distance from roads, distance from faults) were collected for the study area. SBAS-InSAR processing was performed on the study area to output deformation rate and cumulative deformation raster data. Topographic factors such as slope, aspect, curvature, and relief were derived from the DEM data. Low-resolution data, such as rainfall data, were resampled. All factors were standardized in coordinate system and resolution, and the study area was truncated. Data standardization was also performed: all multi-source data were resampled to a unified spatial resolution (10m × 10m) and a unified projected coordinate system was established.
[0063] Step 2: Based on the preprocessed multi-source data, historical geological disaster points and InSAR significant deformation points are used as the positive sample set P, and the remaining data in the study area are used as the unlabeled sample set U; the PU sample system is constructed by combining the positive sample set P and the unlabeled sample set U.
[0064] In one embodiment, the PU sample system is constructed by: setting a positive sample set... The set of pixels is defined as two categories: first, cataloged historical landslide points; and second, significantly deformed pixels located in slope areas with an absolute annual deformation rate greater than 15 mm / year, detected using InSAR technology. In addition to the positive samples mentioned above, all remaining raster cells within the study area are classified into the unlabeled sample set. They are all uniformly labeled 0 and used as background or potential negative samples in model training.
[0065] Step 3: Based on the PU sample system, by introducing the PU-Learning strategy using Spy samples, reliable negative samples RN with high confidence are accurately identified from the unlabeled sample set U. Combined with the positive sample set P, a clean positive and negative sample dataset is constructed for supervised learning, such as... Figure 2 As shown; this strategy aims to solve the "label noise" problem caused by traditional random negative sample selection; the specific operation is as follows:
[0066] Spy sampling and set construction: according to a preset sampling ratio (In this embodiment, the percentage is set to 15%) From the positive sample set Randomly select a subset As a "spy" sample, its label was temporarily reset to negative and mixed into the unlabeled sample set. In the process, a hybrid dataset is formed. The remaining positive samples are denoted as .
[0067] ;
[0068] ;
[0069] Initial discriminator output probability: using and Train an initial classifier ,right Predict all samples and obtain samples Posterior probability of belonging to the positive class .
[0070] ;
[0071] In the formula, A status label indicating that the sample belongs to the positive class (i.e., a geological hazard point); This is the initial discriminator model obtained through training; In the model Below, the known sample features are The probability that it is a positive class.
[0072] Threshold dynamic determination: statistical set The predicted probability distribution of all Spy samples; since Spy samples are essentially positive samples, their predicted probability should be significantly higher than that of true negative samples; calculate the probability distribution of this distribution. Quantiles (set to 10th quantile in this example) are used as the threshold for reliable negative samples. .
[0073] ;
[0074] Reliable negative sample screening: Traversing the unlabeled sample set Predicting probability The samples are considered to be in a very low susceptibility region (reliable negative samples); based on the set positive-to-negative sample ratio of 1: (In this embodiment, the ratio is 1:1), the sample with the lowest probability is selected from those that meet the conditions. These samples serve as the final reliable negative sample set. .
[0075] ;
[0076] ;
[0077] In the formula, For the first The individual The possible values for the dimension; and The populations at the 1st Maximum and minimum boundary values in a dimension; Dimension number;
[0078] Step 4: Train the XGBoost susceptibility assessment model based on the clean sample set constructed in Step 3, and introduce the BTO optimization algorithm to optimize and search for the XGBoost hyperparameters to obtain the optimal hyperparameter combination and train the optimal model; use the optimal model to output the geological hazard susceptibility probability grid of the study area.
[0079] Using a clean dataset of positive and negative samples Construct and train an XGBoost susceptibility evaluation model with the objective function... Including the loss function and regularization term, the formula is as follows:
[0080] ;
[0081] ;
[0082] in, This is the predicted output for the i-th sample. K represents the true labels, and K represents the total number of tree models. For the k-th regression tree, This is the function space of the corresponding tree; A differentiable convex loss function to measure the difference between predicted and actual values; T is the number of leaf nodes. Here, w is the Log-Loss loss function, and w is the weight vector of the leaf nodes. The penalty coefficient for splitting leaf nodes. is the L2 regularization coefficient for the weights of the leaf nodes;
[0083] The BTO (Block-The-Board) optimization algorithm is introduced to optimize the key parameters of XGBoost. Based on the barrel theory, the BTO algorithm maps the combination of XGBoost hyperparameters to the walls of a barrel, and searches for the global optimum by iteratively repairing the weakest dimension that limits the accuracy of the model. The optimization logic is as follows:
[0084] The optimization targets include the learning rate (eta), the maximum depth of the tree (max_depth), the subsample rate, the column sample rate (colsample_bytree), and the regularization weights (lambda), etc.
[0085] Define the first in the population Hyperparameter combinations fitness is (i.e., the model's AUC value).
[0086] Calculate the normalized values for each dimension and identify the bottleneck dimension index that limits the current water capacity. This allows for the identification of weaknesses:
[0087] ;
[0088] Focusing only on the shortcomings Perform targeted patching, and use the individual with the best metric found in the current iteration as the global optimum. Using the global optimal solution Introduce and generate random perturbations:
[0089] ;
[0090] In the formula, As a model evaluation metric, , For learning and collaboration coefficients, , It is a random number. For the corresponding dimension values of random reference individuals, This represents the number of iterations. For the first Individuals in the dimension of shortcomings The value that can be taken on.
[0091] The globally optimal hyperparameter combination is obtained through BTO iterative search and then input into the XGBoost susceptibility assessment model to perform pixel-by-pixel prediction for the entire region.
[0092] refer to Figure 3 As shown, the ROC curves and PR curves of different models on the test set are compared. Table 1 shows the comparison results of the prediction performance of each model.
[0093] Table 1 Comparison of prediction performance of different models
[0094]
[0095] Step 5: Classify the susceptibility probability raster into high-susceptibility, medium-susceptibility, low-susceptibility, and non-susceptibility zones using the natural breakpoint method; set natural breakpoints based on the susceptibility probability values of different areas to classify the degree of susceptibility, so as to facilitate accurate prediction and assessment of geological disaster susceptibility; using ArcGIS's reclassification tool, the probability values are divided into four levels using the natural breakpoint method and a geological disaster susceptibility classification map is output. The geological disaster susceptibility zoning statistics obtained in this step are shown in Table 2.
[0096] Table 2. Geological hazard susceptibility zoning statistics based on BTO-PU-XGBoost
[0097]
[0098] In summary, by enhancing the dynamic identification capability of disasters through InSAR deformation priors, reducing the risk of negative sample contamination by screening reliable negative samples from unlabeled samples through the Spy-PU strategy, and using BTO for efficient global optimization of XGBoost, we can finally achieve high-precision output of the geological disaster susceptibility probability grid and hierarchical zoning results in the study area. This approach has good engineering application value and is worthy of promotion.
[0099] This application also provides a geological hazard prediction system based on an optimized PU-XGBoost model, including:
[0100] Data acquisition module: used to download multi-source data, including InSAR deformation monitoring data, historical geological disaster point data, DEM data, rainfall and hydrological data, land cover and human engineering activity data;
[0101] Preprocessing module: Used to preprocess multi-source data, including unifying coordinate system and resolution, cropping the study area, handling invalid and missing values, and standardization.
[0102] Spy-PU sample construction module: Constructs a positive sample set P and an unlabeled sample set U, and introduces Spy samples to dynamically determine the screening threshold. It filters out the negative sample set RN from the unlabeled sample set U, and establishes the PU sample system by combining the positive sample set P and the negative sample set RN.
[0103] Model training and optimization module: The XGBoost susceptibility assessment model is trained based on the PU sample system, and the hyperparameters are optimized and searched using the BTO optimization algorithm to obtain the optimal model;
[0104] Output module: Used to output the probability raster and hierarchical zoning results of susceptibility.
[0105] The BTO optimization objects in the model training and optimization module include at least one or more of the following: XGBoost learning rate, tree depth, subsampling rate, column sampling rate, minimum leaf sample weight, regularization coefficient, and number of base learners.
[0106] Among them, data preprocessing, BTO optimization, XGBoost training and raster mapping are automated through scripted workflows to realize the data interface, specifically including: reading InSAR deformation data and multi-source factor raster, completing sample extraction and Spy-PU screening, triggering BTO optimization and XGBoost training, and outputting the susceptibility probability raster and hierarchical mapping of the study area.
[0107] The embodiments of the present invention have been described in detail above with reference to the examples. However, the present invention is not limited to the above embodiments. For those skilled in the art, after learning the contents described in the present invention, several equivalent changes and substitutions can be made without departing from the principle of the present invention. These equivalent changes and substitutions should also be considered to fall within the protection scope of the present invention.
Claims
1. A geological hazard prediction method based on an optimized PU-XGBoost model, characterized in that: Includes the following steps: Acquire multi-source data for the study area, including at least InSAR deformation monitoring data and historical geological hazard data, and preprocess the multi-source data; Based on the preprocessed multi-source data, the positive sample set P is taken from historical geological disaster points and InSAR significant deformation points in InSAR deformation monitoring data, and the remaining data in the study area is taken as the unlabeled sample set U; the PU sample system is constructed by combining the positive sample set P and the unlabeled sample set U. Based on the PU sample system, the negative sample set RN is selected from the unlabeled sample set U by introducing the PU-Learning strategy of Spy samples; An XGBoost susceptibility evaluation model was established, and the BTO optimization algorithm was introduced to optimize and search for XGBoost hyperparameters. The optimal hyperparameter combination was obtained and the optimal model was trained using the positive sample set P and the negative sample set RN. The model parameters are optimized by minimizing the objective function that includes a regularization term; the BTO optimization algorithm is introduced to optimize the XGBoost hyperparameters, specifically: Map the XGBoost hyperparameter combinations to the walls of a barrel. Define the i-th hyperparameter combination in the population. The adaptability is the water holding capacity of the wooden barrel. : ； Calculate the normalized values for each dimension to identify the bottleneck dimension limiting the current water capacity. : ； In the formula, For the first The individual The possible values for the dimension; and The populations at the 1st Maximum and minimum boundary values in a dimension; For dimension number, A function for calculating fitness; Focusing only on the shortcomings Perform targeted repair, and calculate the function based on fitness. The individual with the best index found in the current iteration is taken as the global optimal solution. Using the global optimal solution By guiding and introducing random cooperative perturbations, the repair formula satisfies: ； In the formula, As a model evaluation metric, , For learning and collaboration coefficients, , It is a random number. For the corresponding dimension values of random reference individuals, This represents the number of iterations. For the first Individuals in the dimension of shortcomings The value of ; Based on the optimal model, the study area is used to predict and classify the susceptibility to geological disasters.
2. The geological hazard prediction method based on the optimized PU-XGBoost model according to claim 1, characterized in that: The multi-source data also includes DEM data, rainfall and hydrological data, land cover and human engineering activity data; the preprocessing includes: performing SBAS-InSAR processing on the study area to output deformation rate and cumulative deformation raster; performing topographic factor derivation calculations on the DEM data; performing raster resampling on the rainfall data; and All factors should be subjected to at least a unified coordinate system, unified resolution, and study area cropping.
3. The geological hazard prediction method based on the optimized PU-XGBoost model according to claim 2, characterized in that: For InSAR deformation time series in InSAR deformation monitoring data, robust statistics are used to construct derived factors, which include at least one of the following: Mean deformation The calculation formula is as follows: ； Deformation standard deviation The calculation formula is as follows: ； absolute value of deformation rate The calculation formula is as follows: ； in, For the first Cumulative deformation over time For the number of phases, The mean, Standard deviation, and The start and end times of the monitoring period, This represents the absolute value of the deformation rate.
4. The geological hazard prediction method based on the optimized PU-XGBoost model according to claim 1, characterized in that: The positive sample set P includes cataloged historical geological disaster points and InSAR significant deformation points located in slope areas with an absolute annual deformation rate greater than 15 mm / year, monitored based on InSAR technology, and their target labels are assigned a value of 1; the unlabeled sample set U consists of all other sample data except for the positive sample set P, and all data in the unlabeled sample set U are uniformly assigned an initial label of 0.
5. The geological hazard prediction method based on the optimized PU-XGBoost model according to claim 1, characterized in that: The PU-Learning strategy that introduces Spy samples is as follows: Randomly select Spy samples from the positive sample set P, temporarily relabel them as negative samples, and mix them into the unlabeled sample set. In the process, a hybrid dataset is formed. ; Predict the class probability distribution of all Spy samples using a classifier, and calculate the class probability distribution. quantiles as a screening threshold ; Traverse the unlabeled sample set Unlabeled sample set Medium sample prediction probability The samples retained constitute the negative sample set RN.
6. The geological hazard prediction method based on the optimized PU-XGBoost model according to claim 5, characterized in that: According to the preset ratio Randomly draw Spy samples from the positive sample set P; A preset ratio of positive sample set P to negative sample set RN is used, and the predicted probability is based on this ratio. Select the sample with the lowest probability The samples constitute the negative sample set RN.
7. The geological hazard prediction method based on the optimized PU-XGBoost model according to claim 1, characterized in that: Based on the optimal model, the geological disaster susceptibility of the study area is predicted and a geological disaster susceptibility probability grid is output. The susceptibility probability grid is classified into high susceptibility, medium susceptibility, low susceptibility and low susceptibility areas according to the natural breakpoint method.
8. A geological hazard prediction system based on an optimized PU-XGBoost model, implementing the geological hazard prediction method based on an optimized PU-XGBoost model as described in any one of claims 1-7, characterized in that: include: Data acquisition module: used to download multi-source data, including InSAR deformation monitoring data, historical geological disaster point data, DEM data, rainfall and hydrological data, land cover and human engineering activity data; Preprocessing module: Used to preprocess multi-source data, including unifying coordinate system and resolution, cropping the study area, handling invalid and missing values, and standardization. Spy-PU sample construction module: Constructs a positive sample set P and an unlabeled sample set U, and introduces Spy samples to dynamically determine the screening threshold. It filters out the negative sample set RN from the unlabeled sample set U, and establishes the PU sample system by combining the positive sample set P and the negative sample set RN. Model training and optimization module: The XGBoost susceptibility assessment model is trained based on the PU sample system, and the hyperparameters are optimized and searched using the BTO optimization algorithm to obtain the optimal model; Output module: Used to output the probability raster and hierarchical zoning results of susceptibility.
9. The geological hazard prediction system based on the optimized PU-XGBoost model according to claim 8, characterized in that: The BTO optimization objects in the model training and optimization module include at least one or more of the following: XGBoost learning rate, tree depth, subsampling rate, column sampling rate, minimum leaf sample weight, regularization coefficient, and number of base learners.