Biological disaster prediction method and system

Abstract

The invention relates to a biological disaster prediction method and system, and the method comprises the following steps: S100, integrating multi-source remote sensing and environment data and a field investigation sample, and constructing an initial data set; s200, based on the initial data set, a space prediction model set and a time sequence prediction model are trained, and the space prediction model set comprises at least two anomaly detection models; s300, selecting the anomaly detection model with the optimal performance in the spatial prediction model set; and S400, outputting a prediction result based on the anomaly detection model with the optimal performance and the time sequence prediction model with the optimal performance. Through the arrangement, the timeliness, precision and adaptability of locust disaster prediction are remarkably improved, and the locust disaster prediction system is widely applied to agricultural disaster prevention, grassland ecological monitoring and disaster early warning platform construction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning, in particular to a biological disaster prediction method and system. BACKGROUND

[0002] As one of the important biological disasters in the global agricultural field, locust disaster has the characteristics of strong outbreak, fast spread, and difficult monitoring, which seriously threatens food security and ecological system stability. The formation and migration behavior of locust population are influenced by a variety of natural and human factors such as climate change, land use, and vegetation coverage, and the spatial distribution has high uncertainty and dynamics. Therefore, it is of great importance to accurately and timely predict the occurrence area and development trend of locust disaster for the protection of food production and ecological safety.

[0003] At present, the traditional locust disaster prediction method mainly relies on ground manual monitoring and experience judgment, which has the problems of limited monitoring range, poor precision and timeliness, high labor cost, and low efficiency. SUMMARY

[0004] In order to solve the above defects, the present application provides a biological disaster prediction method and system. In view of the problems of insufficient data utilization, lag response to dynamic changes, and poor model generalization ability in current locust disaster monitoring and prediction, the present application provides a comprehensive prediction method and system which integrates multi-source remote sensing, time series modeling, improved anomaly detection algorithm, and spatial prediction visualization, breaks through the limitations of existing technology in time series modeling, multi-model collaboration, and high-precision prediction, and has stronger universality, robustness, and practical value.

[0005] The technical scheme adopted by the present application is a biological disaster prediction method, comprising the following steps: S100, integrating multi-source remote sensing and environmental data, and field investigation samples to construct an initial data set; S200, training a spatial prediction model set and a time series prediction model based on the initial data set, the spatial prediction model set comprising at least two anomaly detection models; S300, selecting the anomaly detection model with the best performance in the spatial prediction model set; S400, outputting a prediction result based on the anomaly detection model with the best performance and the time series prediction model.

[0006] Preferably, the multi-source remote sensing and environmental data specifically comprises at least one of soil properties, meteorological factors, vegetation indices, terrain data, or land use types.

[0007] Preferably, the training of the spatial prediction model set and the time series prediction model based on the initial data set specifically comprises: S210, acquiring a static data set and a time series data set based on the initial data set; S220, training the spatial prediction model set based on the static data set; training the time series prediction model based on the time series data set.

[0008] Preferably, the spatial prediction model set specifically comprises at least two of support vector machines, isolation forests, one-class support vector machines, or local outlier factor; and / or The time series prediction model specifically comprises one of long short-term memory networks, bidirectional long short-term memory networks, gated recurrent units, or temporal convolutional networks.

[0009] Preferably, the acquiring the time series data set based on the initial data set specifically comprises: for dynamic variables in the initial data set, extracting time series by using a sliding time window to construct a time series data set.

[0010] Preferably, the S300 specifically comprises: S310, calculating a probability value of each model in the spatial prediction model set that the predicted positive sample score is higher than the negative sample score; S320, selecting the anomaly detection model with the highest probability value as the anomaly detection model with the best performance.

[0011] Preferably, the S400 further comprises: S500, generating a risk spatial distribution map and a dynamic time series atlas based on the prediction result.

[0012] Preferably, the S500 specifically comprises: S510, calculating a percentage of the prediction correct result in the total sample, and / or a harmonic mean of precision and recall of the prediction result; S520, when the percentage and / or the harmonic mean is less than a threshold value, adjusting the parameters of the model in the S400, and repeating the S500; when the percentage and / or the harmonic mean is greater than or equal to the threshold value, generating the risk spatial distribution map and the dynamic time series atlas.

[0013] The application also discloses a biological disaster prediction system, comprising: A data preprocessing module is configured to collect and preprocess multi-source remote sensing and environmental data and field investigation samples, and generate a static data set and a time series data set; A static multi-model training module is configured to train at least two anomaly detection models by using the static data set; A dynamic time series modeling module is configured to train a time series prediction model by using the time series data set; The model selection module is configured to evaluate the multiple abnormality detection models output by the static multi-model training module and select the abnormality detection model with the optimal performance according to the evaluation result. The model verification and optimization module receives the prediction result of the abnormality detection model with the optimal performance and the time series prediction model, verifies the prediction result by using real-time data, adjusts the model parameters, and optimizes the prediction performance.

[0014] Preferably, the method further comprises a visualization module configured to generate a risk space distribution map and a dynamic time series atlas based on the prediction result of the optimized model.

[0015] Compared with the prior art, the method has the following beneficial effects: 1. The method integrates soil, meteorological, terrain, land use and other multi-source remote sensing and environmental data, combines with field investigation samples, and constructs a comprehensive input feature system. Various algorithms such as support vector machine, isolated forest, local outlier factor and One-Class SVM are used for prediction modeling, and sliding time window and LSTM / GRU neural network are introduced to realize time series modeling of dynamic environmental factors such as NDVI and precipitation, so as to accurately identify and predict the distribution change of locust habitat in different time and regions.

[0016] 2. The method optimizes the optimal prediction model by model integration and precision evaluation system, outputs the spatial risk distribution map and dynamic evolution atlas, significantly improves the timeliness, precision and adaptability of locust disaster prediction, and is widely applied to agricultural disaster prevention, grassland ecological monitoring and disaster warning platform construction. BRIEF DESCRIPTION OF DRAWINGS

[0017] The application will be described in detail below with reference to the embodiments and the accompanying drawings, in which: Figure 1 is a biological disaster prediction method flow chart; Figure 2 is an initial data set schematic diagram; Figure 3 is a spatial prediction model set schematic diagram; Figure 4 is an AUC-ROC curve of four learning methods; Figure 5 is a schematic diagram of locust disaster prediction results and precision in three different periods. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the present application clearer, the embodiments of the present application will be further described in detail below with reference to the drawings. The examples of the embodiments are shown in the drawings, wherein the same or similar reference signs represent the same or similar components or components having the same or similar functions throughout. The embodiments described below by reference to the drawings are exemplary and are only used to explain the present application, but cannot be explained as a limitation of the present application.

[0019] There are some technical bottlenecks in the disaster prediction method and system, such as the difficulty of multi-source remote sensing data fusion, how to effectively process remote sensing information of different sensors and different wave bands and realize comprehensive utilization of data. The feature selection and optimization are insufficient, the multi-dimensional features related to the locust disaster cannot be fully mined, and the prediction accuracy of the model is low. The real-time early warning capability for large-scale area is poor, there is delay in wide range monitoring and rapid prediction, and it cannot meet the demand of real-time decision. Therefore, an improved prediction method and system capable of fusing multi-source remote sensing, dynamic modeling and multi-algorithm integration are needed to improve the intelligent identification and spatial distribution prediction capability of locust disaster.

[0020] The locust disaster prediction method based on multi-source remote sensing data and machine learning can solve the deficiencies of the existing method in precision, efficiency and timeliness through data fusion, feature extraction and selection, real-time prediction and other technical means, and has important research value and application prospect.

[0021] In one embodiment, a biological disaster prediction method, referring to Figure 1 , comprises the following steps: S100, integrating multi-source remote sensing and environmental data and field investigation samples to construct an initial data set, referring to Figure 2 . The field investigation samples are, for example, locust disaster and egg area sample points. The remote sensing data (such as vegetation index, terrain, weather), environmental data (soil, land use) and field investigation samples are integrated to construct an initial data set containing spatial features and time series, solving the problem of single data of traditional methods.

[0022] S200, training a set of spatial prediction models and a time series prediction model based on the initial data set, the set of spatial prediction models comprising at least two anomaly detection models. The set of spatial prediction models: containing at least two anomaly detection models (such as isolated forest, support vector machine), used to identify the spatial area where the locust disaster may occur; training the time series prediction model (such as LSTM), capturing the time evolution law of environmental factors (such as precipitation, temperature), predicting the disaster development trend.

[0023] S300, select the anomaly detection model with the best performance in the spatial prediction model set. Evaluate the spatial model set by Accuracy, F1-Score, ROC-AUC and other indicators, and select the anomaly detection model with the best performance to ensure the accuracy of spatial prediction.

[0024] S400, output the prediction result based on the anomaly detection model with the best performance and the time series prediction model. Specifically, based on the anomaly detection model with the best performance (spatial prediction model) and the time series prediction model, the joint modeling of spatial and temporal information is realized, and the prediction result is output. The results of the optimal spatial model and the time series model are combined to output the spatial distribution and spatio-temporal evolution prediction result of the locust disaster, providing comprehensive decision support for early warning.

[0025] In one specific embodiment, the spatial prediction model and the time series prediction model can be fused, and the integrated framework is adopted to take the spatial anomaly detection result as one of the input features of the time series model, realizing the joint modeling of spatial and temporal information. The specific process is as follows: the anomaly score output by the spatial model is taken as one of the additional input features of the time series model to form a new input vector: Finally, the fusion probability is output: Where σ is the Sigmoid function, the weight w1=0.6, w2=0.4. When P≥0.5, it is determined as a high-risk pixel.

[0026] Generate a spatial risk distribution map ( Figure 5 ), which shows that the high-risk area is mainly concentrated in the shrub grassland along the Amu Darya Valley. The comparison of the actual measurement shows that the prediction AUC of the fusion model is 0.93, and the F1 is 0.92; compared with Isolation Forest (AUC=0.88) and LSTM (AUC=0.86) alone, the accuracy is improved by 5-7%; the early warning time is extended from 7 days to 14 days.

[0027] The present application utilizes automatic data processing and model training to reduce manual intervention and reduce monitoring cost; multi-source data fusion improves the generalization ability of the model and has strong adaptability; remote sensing data is used to realize global / regional scale automatic monitoring, and the data update frequency is high, which can capture the disaster dynamics in real time; at least two anomaly detection models are fused, and the prediction stability is improved through result fusion and index optimization, which can adapt to diversified regional sample distribution and label scarcity problem, and enhance the practicability and generalization ability of the algorithm. The introduction of the time series prediction model effectively extracts the time series features of the initial data set, significantly improves the perception ability of the disaster outbreak trend and migration change, and realizes more forward-looking dynamic prediction. ​

[0028] In one embodiment, the multi-source remote sensing and environmental data specifically includes at least one of the following: soil properties, meteorological factors, vegetation index (NDVI), topographic data (DEM), or land use type (LULC). Soil properties include bulk density and sand content, while meteorological factors include precipitation, ground temperature, and soil moisture.

[0029] In one embodiment, data acquisition and preprocessing.

[0030] Remote sensing data selection: Based on the study area and required spatial resolution, select appropriate remote sensing satellite platforms and data sources. Use available product data as the data source and download the required imagery through official data portals and third-party service platforms (such as USGS Earth Explorer, Copernicus Open Access Hub, etc.). Downloads include: Land Use / Land Cover LULC, Digital Elevation Model (DEM), Precipitation (PREC), Normalized Difference Vegetation Index (NDVI), Land Surface Temperature (LST), etc. These satellites can provide multi-source, multi-temporal spatial distribution imagery. The table below shows the data sources and their spatiotemporal resolutions in this embodiment.

[0031] Among them, CHIRPS is data from climate disaster infrared precipitation stations, ERA5-Land is high-resolution land reanalysis data based on ERA5 from the European Centre for Medium-Range Weather Forecasts (ECMWF), MOD09GA is a daily surface reflectance data product, MOD11A1 is a daily L3 global 1 km sinusoidal grid data product for land surface temperature and emissivity, ISRIC SoilGrids is soil grid data from the International Soil Reference Information Centre (ISRIC), CGLS-LC100 is a 100-meter resolution land cover data product from the Copernicus Global Land Service (CGLS), and NASADEM is data from NASA's digital elevation model.

[0032] Time series planning: Time series planning of data is performed based on the sampling time of the locust infestation to determine the time interval for data acquisition and ensure that changes before and after the infestation can be captured. For seasonal crops, more frequent data updates (once a week) are required.

[0033] Download the data source, including: Land Use / Cover Data (LULC): Obtain the latest LULC classification maps to identify different land surface types (forest, grassland, farmland, etc.), which helps in understanding the environmental impact on pests. Digital Elevation Model (DEM): Download globally available DEM products such as the Space Shuttle Radar Topographic Mapping Mission (SRTM) and the Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model (ASTER GDEM) to obtain topographic information, as factors such as altitude and slope influence pest occurrence. Meteorological Data: Utilize global reanalysis datasets such as the European Centre for Medium-Range Weather Forecasts (ECMWF) Generation 5 Global Atmospheric Reanalysis Data (ERA5) and Modern Retrospective Analysis Research and Applications, Edition 2 (MERRA-2) to supplement areas not covered by ground stations and provide continuous time-series data. Climate Model Predictions: Consider future climate change scenarios and utilize climate model outputs such as the 6th Coupled Model Intercomparison Project (CMIP6) to provide a reference for long-term pest risk assessment. Vegetation indices (such as Normalized Difference Vegetation Index (NDVI)): Vegetation indices are calculated from multispectral imagery. NDVI = (NIR - Red) / (NIR + Red), where NIR represents the near-infrared band and Red represents the red band. Soil data: Detailed soil type distribution maps are obtained to understand soil texture, organic matter content, and other characteristics. Soil moisture monitoring: Real-time monitoring of soil moisture is conducted using passive microwave remote sensing technology (SMAP satellite).

[0034] Field sampling: Compile past pest outbreaks, including their location, time, and scale, to create a basic database. Construct time series of vegetation indices to monitor dynamic changes during the vegetation growth cycle and the occurrence of anomalies.

[0035] Based on locust plague sampling sites in northwestern Tajikistan in 2017, NDVI time series were extracted, and growth rate GRt, cumulative vegetation index CVI, and standardized anomaly index SAIt were calculated. Areas where SAIt < -1 showed a spatial overlap of 82% with ground locust plague records, indicating that this method can effectively identify early anomalies in disasters.

[0036] Among them, the dynamic change monitoring formula is: growth rate ,in Let be the vegetation index at time t. For time intervals; Cumulative vegetation index Anomaly monitoring formula: Vegetation Anomaly Index ,in, The vegetation index is the multi-year average. This indicates that the vegetation growth is worse than average. Standardized Vegetation Anomaly Index ,in, and Representing the mean and standard deviation for the same period over many years, respectively, when When the value is significantly low, it indicates that there may be abnormalities such as drought or pests.

[0037] Data spatial registration: Collinearity equation matching algorithms are used to ensure consistency between map projections and coordinate systems from different sources. Adjusting the timestamps of each dataset to ensure they are aligned at the same time point is crucial for analyzing causal relationships.

[0038] Among them, the formula for the collinearity equation is: in: Image coordinates; Principal point coordinates; focal length; Ground point coordinates; Projection center coordinates; : Direction cosine of the rotation matrix of the exterior orientation element.

[0039] In one embodiment, training the spatial prediction model set and the temporal prediction model based on the initial dataset specifically includes: S210. Based on the initial dataset, obtain a static dataset and a time-series dataset. Separate the static dataset (single-temporal spatial features, such as land use, topography, and vegetation cover) and the time-series dataset (time series of dynamic variables, such as sliding window sequences of NDVI, precipitation, and temperature) from the initial dataset to meet the needs of spatial detection and time-series prediction, respectively.

[0040] S220. Based on the static dataset, train the spatial prediction model set; based on the temporal dataset, train the temporal prediction model. The spatial prediction model set (e.g., isolated forest, LOF, etc. anomaly detection models) is trained based on the static dataset, focusing on spatial anomaly identification of locust plague occurrences (e.g., anomalous coupling regions between high vegetation cover areas and arid areas); the temporal prediction model (e.g., LSTM) is trained based on the temporal dataset, capturing the temporal evolution of locust plague probability through historical environmental factor sequences (e.g., the high incidence trend of disasters 1-2 weeks after the rainy season).

[0041] The solution in this embodiment significantly improves data adaptability. Static data helps to accurately locate geographical environments prone to locust plagues, while time-series data increases the early warning time. It achieves professional division of labor among models, with spatial models accurately delineating regions and time-series models accurately predicting the occurrence time. It also enhances the collaborative prediction effect and model generalization ability. Compared with a single model, the final prediction accuracy is improved, the prediction stability in different climate zones is significantly improved, and data processing is more efficient while training time is reduced.

[0042] Existing methods often rely on a single classifier (SVM or RF), which is prone to overfitting or failure under different regions and data volumes. In one embodiment, see... Figure 3 The spatial prediction model set specifically includes at least two of the following: Support Vector Machine (SVM), Isolation Forest, One-Class SVM, or Local Outlier Factor (LOF). An ensemble prediction architecture is constructed, integrating multiple models such as Support Vector Machine (SVM), Isolation Forest, One-Class SVM, and Local Outlier Factor (LOF), to achieve accurate identification of high-risk locust areas through ensemble learning.

[0043] In one embodiment, the time-series prediction model specifically includes one of the following: a long short-term memory network, a bidirectional long short-term memory network, a gated recurrent unit, or a temporal convolutional network.

[0044] Most existing technologies are based on single-phase NDVI or meteorological data, lacking modeling of variable evolution trends, making it difficult to achieve dynamic prediction of locust plagues. In one embodiment, obtaining a time-series dataset based on the initial dataset specifically includes: extracting time series data from the dynamic variables in the initial dataset using a sliding time window to construct a time-series dataset. For example, the time series of key variables such as NDVI and precipitation can be extracted using a sliding time window method, introducing time-series modeling to characterize the dynamic evolution of locust plague distribution, and using LSTM (Long Short-Term Memory) network to model its evolution trend, thereby enhancing the model's ability to perceive and predict dynamic changes in locust habitats and achieving more forward-looking dynamic predictions.

[0045] In one embodiment, S300 specifically includes: S310, calculating the probability value that the predicted positive sample score of each model in the spatial prediction model set is higher than the negative sample score; S320, selecting the anomaly detection model with the highest probability value as the anomaly detection model with the best performance.

[0046] This embodiment selects the optimal spatial prediction model by quantifying the model's ability to distinguish between positive and negative samples. Specifically, it first calculates the probability that each model in the spatial prediction model set predicts a higher score for positive samples than for negative samples. This probability directly reflects the model's ability to correctly sort locust-affected areas (positive samples) and normal areas (negative samples). The higher the probability, the more accurately the model can identify anomalous locust-affected areas. Subsequently, the anomaly detection model with the highest probability value is selected as the best-performing model, ensuring that the model used for locust-affected spatial prediction has the strongest sample discrimination performance.

[0047] The scheme in this embodiment can objectively evaluate model performance, avoid subjective experience judgment, and the probability-based screening method is more suitable for locust plague prediction scenarios with imbalanced samples than traditional indicators such as accuracy. It can accurately locate high-quality models, effectively improve spatial prediction accuracy, and the selected optimal model can significantly improve the consistency between the predicted spatial distribution of locust plagues and the actual measurements. It can enhance prediction reliability, lay a solid foundation for the subsequent output of comprehensive prediction results by combining time series models, and ensure the effectiveness of disaster early warning and the scientific nature of prevention and control decisions.

[0048] In one embodiment, step S400 is followed by step S500: generating a risk spatial distribution map and a dynamic time-series map based on the prediction results. The risk spatial distribution map includes, for example, a risk spatial distribution map. In this embodiment, after completing the locust plague prediction, the abstract prediction data is transformed into intuitive and visual results by generating a risk spatial distribution map and a dynamic time-series map based on the prediction results. The risk spatial distribution map visually presents the risk level of locust plagues in different regions in map form, helping users quickly locate high-risk areas; the dynamic time-series map shows the changing trend of locust plague risk over time, combining the dynamic evolution of environmental factors to demonstrate the development trend of the disaster.

[0049] In one embodiment, S500 specifically includes: S510. Calculate the percentage of correctly predicted results out of the total sample, and / or the harmonic mean of precision and recall. The percentage of correctly predicted results out of the total sample can specifically be accuracy; the harmonic mean of precision and recall can specifically be F1-score.

[0050] S520. When the percentage and / or harmonic mean is less than the threshold, adjust the parameters of the spatial model and temporal model in S400, such as the learning rate and window size, and re-execute S500 for verification. When the percentage and / or harmonic mean is greater than or equal to the threshold, generate the risk spatial distribution map and dynamic temporal map.

[0051] Traditional methods lack a systematic evaluation of model performance under scenarios involving spatial consistency and imbalanced data. This invention combines a comprehensive evaluation system with multiple indicators such as F1-score and AUC, which can comprehensively reflect the performance of prediction results in terms of classification accuracy and spatial distribution consistency. At the same time, it generates spatial heat maps and dynamic evolution maps, providing intuitive and quantitative support for disaster management, making it more practical.

[0052] By monitoring both accuracy and F1-Score, model parameters are optimized in a targeted manner. For example, the sliding window length of the LSTM time series model is adjusted (from 14 days to 21 days) or the number of trees in the isolated forest is increased (from 100 to 150 trees), thereby improving prediction accuracy. A closed-loop optimization mechanism avoids the subjectivity of manual parameter tuning based on experience. For instance, if the initial F1-Score for a prediction is 0.82, it can be improved to 0.92 after three parameter iterations, meeting the warning standard. Thresholds can be flexibly set according to application scenarios, adapting to different needs.

[0053] In the above embodiments, the anomaly detection model with the best performance in the spatial prediction model set is selected in S300, and the parameters of the model in S400 are adjusted in S500. Multiple indicators such as F1-score, AUC, Kappa coefficient, and IOU can also be used. For example, by combining multiple indicators such as F1-score, AUC, Kappa coefficient, and IOU, the prediction accuracy of each model can be evaluated, and the model with the best performance can be selected to output a spatial distribution map and dynamic time-series map of locust risk, providing technical support for regional disaster early warning and precise governance. By introducing multi-dimensional indicators such as F1-score and ROC-AUC to comprehensively evaluate the prediction effect, the applicability and interpretability of the model in scenarios with sparse and unevenly distributed locust samples can be improved.

[0054] In one specific embodiment, feature extraction and machine learning model construction.

[0055] Extracting specific feature variables: Vegetation index changes: Calculate and analyze changes in vegetation indices such as the Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI) over different time periods to reflect changes in plant health over time. Abnormally low vegetation indices may be early warning signals of pests. Climate change models: Collect long-term meteorological data (temperature, precipitation, humidity), establish time series of climate factors, and explore their relationship with the frequency of locust infestations. Consider using time series analysis tools (ARIMA models) to predict future climate trends. Soil moisture: Monitor soil moisture levels using passive microwave remote sensing data (SMAP satellite products), as both excessively dry and excessively wet soil conditions can lead to outbreaks of certain types of pests. Other relevant features: Consider incorporating more biophysical parameters (leaf area index LAI, canopy height) and socioeconomic factors (crop planting structure, agricultural management practices) as additional input features.

[0056] Training dataset. The selection of training data is crucial to ensuring the effectiveness and reliability of the model: Historical pest data: This includes, but is not limited to, information on the time, location, and intensity of pest occurrences. This data should be as detailed as possible and cover different seasons and years to capture long-term patterns of pest occurrence. Feature data: Combining the preprocessing steps mentioned above, use cleaned and standardized vegetation index (NDVI), meteorological data (temperature, humidity), and soil data (moisture, texture) as input features. Ensure that all features have a reasonable correlation with the target variable (pest occurrence).

[0057] Model training. Based on the characteristics of the feature data and the problem requirements, the following common machine learning algorithms can be selected: Support Vector Machine (SVM): Suitable for classification tasks, especially performing well in high-dimensional spaces. It can adapt to different data distributions by adjusting the kernel function (linear, multinomial, RBF, etc.). Isolation Forest: Designed specifically for anomaly detection, it can quickly and effectively identify data points far from the majority of sample points, making it ideal for finding anomalous areas in locust plagues. Local Outlier Factor (LOF): Also used for anomaly detection, but it focuses more on data points with low local density, thus being more sensitive in complex environments. SGD One-Class SVM: Particularly suitable for single-class classification problems, performing anomaly detection when there are no explicit negative samples.

[0058] Training process: Training / Validation / Test Set Partitioning: Allocate data in a reasonable proportion (70% training, 15% validation, 15% test), ensuring the data distribution in each subset is as consistent as possible. Feature Standardization: Normalize numerical features to give them similar scales, which helps improve model training efficiency. Parameter Optimization: Grid Search / Random Search: Define a parameter space and try different combinations of hyperparameters to find the optimal configuration. Bayesian Optimization: Use a probabilistic model to guide the search process, gradually approaching the optimal parameter settings, especially suitable for expensive model training. Cross-Validation: Use K-fold cross-validation (K=5 or 10) to evaluate model performance and prevent overfitting.

[0059] Time series modeling. This invention targets dynamic variables such as NDVI, precipitation, and ground temperature, extracting time series inputs using a sliding time window approach and employing an LSTM (Long Short-Term Memory) model for modeling. A historical value sequence (two weeks of NDVI) is established for each pixel: The model outputs the locust plague habitat probability, reflecting the spatial-temporal evolution trend of the disaster; it trains the prediction ability based on time changes to achieve disaster risk assessment in the future.

[0060] Model selection. After training is complete, the model must be comprehensively evaluated for classification problems: accuracy, precision, and AUC-ROC curve.

[0061] One-Class SVM: The high AUC (close to 1) of the ROC curve of One-Class SVM indicates that the model can effectively distinguish between positive and negative samples in classification tasks. Isolation Forest: Isolation Forest is commonly used for anomaly detection. If its ROC curve AUC is close to 1, it means that it can effectively distinguish anomalies from normal samples, especially performing well on large-scale datasets. LOF: A high AUC of the ROC curve of LOF usually means that it can effectively identify local anomalies (points with low local density), while a low AUC indicates that the model performs poorly when dealing with data of uneven density. SGD One-Class SVM: A high AUC value of the ROC curve of SGD One-Class SVM indicates that it can effectively distinguish between normal and anomalies, but a low AUC indicates that it is not sensitive enough to the identification of anomalies.

[0062] Figure 4 These are the AUC-ROC curves for the four learning methods in this embodiment. In this embodiment, Isolation Forest is selected as the machine learning method for predicting the distribution of locust plagues.

[0063] In one specific embodiment, locust plague distribution prediction is performed. A machine learning model trained using an isolation forest model is used to predict the future occurrence areas and intensity of locust plagues. Based on the temporal variations of remote sensing and meteorological data, the spatial distribution and spatiotemporal trends of locust plagues are predicted. Locust plague warnings are issued based on the prediction results, providing decision support. The actual occurrence of locust plagues is evaluated using real-time remote sensing and ground monitoring data to verify the accuracy of the prediction model and adjust model parameters to improve prediction performance.

[0064] Figure 5 This paper presents the prediction results of locust plagues at three different time periods and the corresponding accuracy evaluation metrics. Each set of images consists of a prediction result map and two evaluation metrics: Accuracy Score and F1 Score. The evaluation results are as follows: the third period has an Accuracy Score of 0.91 and an F1 Score of 0.92, exhibiting the best prediction performance. The locust plague distribution closely matches the actual measurements, making it suitable for practical early warning deployment and refined management. With model training iterations and the introduction of temporal features, the accuracy and F1 score of locust plague predictions significantly improve, indicating that the method of this invention has good spatiotemporal generalization ability and practical application potential.

[0065] In one embodiment, a biological disaster prediction system can execute the biological disaster prediction method in the above embodiments, and the system includes: The data preprocessing module is used to collect and preprocess multi-source remote sensing and environmental data, as well as field survey samples, to generate static datasets and time-series datasets. Specifically, it can integrate satellite remote sensing (NDVI, LST, etc.), environmental data (meteorology, soil), and field survey samples to generate static feature sets (such as topography and land use types) and time-series feature sets (such as precipitation and temperature time series).

[0066] The static multi-model training module trains at least two anomaly detection models using a static dataset. It trains multiple anomaly detection models (such as Isolation Forest, KNN, and OC-SVM) in parallel to capture spatial distribution patterns of disasters. The dynamic temporal modeling module uses the aforementioned temporal dataset to train a temporal prediction model. Based on models such as LSTM, it predicts the temporal evolution of disaster occurrences.

[0067] The model selection module is used to evaluate the various anomaly detection models output by the static multi-model training module and select the anomaly detection model with the best performance based on the evaluation results. The performance of each model is evaluated using metrics such as ROC-AUC and F1-Score, and the optimal spatial prediction model (such as the Isolation Forest model with AUC>0.9) is selected.

[0068] The model validation and optimization module receives the prediction results from the optimal anomaly detection model and the time-series prediction model, validates the prediction results using real-time data, and adjusts model parameters to optimize prediction performance. By using real-time data to validate the prediction results and dynamically adjusting model parameters (such as the LSTM time step size and anomaly detection threshold), a closed loop of "prediction-validation-tuning" is formed.

[0069] In one embodiment, the biological disaster prediction system also includes a visualization module that generates a risk spatial distribution map and a dynamic time series map based on the prediction results of the optimized model. The optimized prediction results are transformed into a risk spatial distribution map (visually displaying high-risk areas) and a dynamic time series map (predicting disaster spread paths and intensity changes).

[0070] This system achieves an upgrade from "experience-based judgment" to "data-driven precise early warning" in locust disaster prediction through a full-chain innovation of "data-model-verification-visualization," and is particularly suitable for disaster prevention and control decisions in food security and ecologically fragile areas.

[0071] Existing methods for predicting locust plagues rely solely on static, temporal remote sensing data, neglecting the temporal evolution of variables such as NDVI and precipitation. To address the problem of underutilization of temporal information from multi-source remote sensing and meteorological data, resulting in poor responsiveness of prediction models to the dynamic evolution of locust plagues, this invention proposes constructing a sliding time window and introducing sequence models such as LSTM / GRU to extract evolutionary features from multi-temporal remote sensing data, forming a temporal habitat map and improving the timeliness and accuracy of disaster prediction.

[0072] To address the issue that single anomaly detection models suffer from limited predictive stability and generalization ability, and are prone to prediction failures across different regions or years, this invention proposes a fusion of multiple anomaly detection models (One-Class SVM, IsolationForest, and LOF) and introduces an ensemble learning strategy (Stacking) to improve the model's adaptability and robustness in multi-temporal and spatiotemporal scenarios.

[0073] One-Class SVM, Isolation Forest, and LOF are trained separately, with each model trained on input sample 𝑥 𝑖 The system outputs an "outlier score". One-Class SVM fits the normal data boundary and outputs the outlier score. Isolation Forest: Outputs outlier scores based on a randomized splitting tree. LOF (Local Occurrence Factor): Calculates local density deviation and outputs anomaly scores. The outputs of multiple base learners are used as new feature vectors: Logistic regression was used. ,in, This is the Sigmoid function, which outputs the probability of an anomaly.

[0074] In the Isolation Forest example, remote sensing data from Tajikistan from 2016 to 2020 were processed.

[0075] ① Extract NDVI and precipitation sequences using a sliding window, with a time interval of 7 days; ② The spatial model used is Isolation Forest (150 trees, 0.3% of the sample size). ③ The time series model uses LSTM (input length 4, hidden layers 64, learning rate 0.001). ④ Stacking logistic regression is used in the fusion phase: ; ⑤ Accuracy evaluation: AUC=0.91, F1=0.92, indicating that the fusion model has the best performance in terms of spatial consistency and temporal trend prediction.

[0076] To address the challenges of sparsely distributed locust plague samples with severe class imbalance, the inapplicability of traditional accuracy metrics, and the lack of spatial consistency verification and dynamic habitat visualization in prediction results, which hinders practical early warning deployment, this invention proposes introducing classification evaluation metrics such as Precision-Recall, F1-score, and MCC, combined with spatial consistency measures such as IOU (Intersection over Union) and Kappa coefficient, to achieve comprehensive and objective model evaluation and optimization.

[0077] In actual testing, the Isolation Forest model achieved F1=0.92, AUC=0.90, Kappa=0.86, IOU=0.81, Precision=0.89, and Recall=0.94. These results demonstrate that the model maintains stable predictive ability even with imbalanced samples, and its spatial consistency is approximately 9% higher than traditional SVM methods.

[0078] To address the problem that existing models often output results in tabular or single-frame images, lacking continuous multi-temporal habitat visualization and thus failing to provide actionable regional-level response measures for agricultural disaster prevention, this invention proposes generating spatial raster dynamic maps from the model's prediction output. Combined with a GIS platform, this creates continuous prediction maps and heatmaps, facilitating early warning and regional prevention and control deployment.

[0079] In one embodiment, the present invention was experimentally simulated.

[0080] The experimental procedure is as follows: ① Data acquisition and registration: A unified spatial coordinate system was established using multi-source data, including CHIRPS precipitation data, MOD09GA NDVI, ERA5 temperature data, and ISRIC soil data. ② Data normalization and time alignment processing; ③ Feature extraction: Calculate and standardize GRt, CVI, AI, and SAI metrics; ④ Model training: The model is trained using Python scikit-learn and TensorFlow, with a 70% training set, 15% validation set, and 15% test set. ⑤ Model fusion and evaluation: Through comprehensive evaluation using F1, AUC, Kappa, and IOU indicators, the final output is a risk map and a dynamic time series map.

[0081] Experimental Location: This invention underwent field sampling and verification in Tajikistan. Tajikistan's geographical environment and climate provide a challenging practical application scenario for locust plague prediction. The method of this invention is not only applicable to locust plagues but can also be extended to the monitoring and prediction of other large-scale agricultural pests such as grass moths and armyworms, as well as forest pests (such as pine wilt nematode) and crop diseases (such as wheat powdery mildew). By adjusting the input features (such as EVI, LAI, LST) and training samples, model transfer and reuse can be achieved, demonstrating broad applicability.

[0082] Validation Method: In Tajikistan, the research team continuously tracked and monitored the occurrence of locust plagues by collecting multi-source remote sensing data (LULC, NDVI, meteorological data, soil moisture, etc.) and ground monitoring data. The machine learning model proposed in this invention was used to analyze the distribution and prediction of locust plagues, and the results were compared and validated with actual locust plague events.

[0083] Accuracy Evaluation: Experimental results show that the accuracy of the locust plague prediction method of this invention is between 0.7 and 0.9, demonstrating high accuracy. The high accuracy of the prediction results indicates that the method of this invention can effectively identify the occurrence area and intensity of locust plagues, and can predict the spatiotemporal distribution of locust plagues relatively well.

[0084] In the description of this specification, the use of terms such as "Embodiment 1," "this embodiment," or "in one embodiment" indicates that the specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example; moreover, the specific features, structures, materials, or characteristics described may be combined in any appropriate manner in one or more embodiments or examples.

[0085] In the description of this specification, the terms "connection," "installation," "fixing," "setting," and "having" are interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0086] In the description of this specification, relational terms such as “first” and “second” are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase “comprising one…” does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0087] The above description of the embodiments is intended to enable those skilled in the art to understand and apply the technology of this invention. Those skilled in the art can easily make various modifications to these examples and apply the general principles described herein to other embodiments without creative effort. Therefore, this invention is not limited to the above embodiments. Modifications in the following situations should be within the scope of protection of this invention: ① New technical solutions implemented based on the technical solution of this invention and combined with existing common knowledge, where the technical effects of the new technical solution do not exceed the technical effects of this invention; ② Equivalent substitutions of some features of the technical solution of this invention using known technology, resulting in the same technical effects as those of this invention; ③ Extendable technical solutions based on the technical solution of this invention, where the substantive content of the extended technical solution does not exceed the technical solution of this invention; ④ Equivalent transformations made using the content of this specification and drawings, directly or indirectly applied to other related technical fields.

Claims

1. A method of biological disaster prediction, characterized by, The method comprises the following steps: S100, integrating multi-source remote sensing and environmental data and field investigation samples to construct an initial data set; S200, training a spatial prediction model set and a time series prediction model based on the initial data set, wherein the spatial prediction model set comprises at least two anomaly detection models; S300, selecting the anomaly detection model with the best performance in the spatial prediction model set; S400, outputting a prediction result based on the anomaly detection model with the best performance and the time series prediction model.

2. The method of claim 1, wherein, The multi-source remote sensing and environmental data specifically comprises at least one of soil properties, meteorological factors, vegetation indexes, terrain data or land use types.

3. The method of claim 1, wherein, The training of the spatial prediction model set and the time series prediction model based on the initial data set specifically comprises: S210, obtaining a static data set and a time series data set based on the initial data set; S220, training the spatial prediction model set based on the static data set; training the time series prediction model based on the time series data set.

4. The method of claim 3, wherein, The spatial prediction model set specifically comprises at least two of support vector machines, isolation forests, one-class support vector machines or local outlier factors; and / or The time series prediction model specifically comprises one of long short-term memory networks, bidirectional long short-term memory networks, gated recurrent units or time-domain convolutional networks.

5. The method of claim 3, wherein, The obtaining of the time series data set based on the initial data set specifically comprises: for dynamic variables in the initial data set, extracting time series by using a sliding time window to construct a time series data set.

6. The method of claim 1, wherein, The S300 specifically comprises: S310, calculating a probability value of each model in the spatial prediction model set in which the prediction positive sample score is higher than the negative sample score; S320, selecting the anomaly detection model with the highest probability value as the anomaly detection model with the best performance.

7. The method of claim 1, wherein, The S400 further comprises: S500, generating a risk spatial distribution map and a dynamic time series atlas based on the prediction result.

8. The method of claim 7, wherein, The S500 specifically comprises: S510, calculating a percentage of prediction correct results in total samples and / or a harmonic mean of precision and recall of the prediction result; S520, when the percentage and / or the harmonic mean is less than a threshold value, adjusting parameters of the model in the S400 and repeating the S500; when the percentage and / or the harmonic mean is greater than or equal to the threshold value, generating the risk spatial distribution map and the dynamic time series atlas.

9. A biological disaster prediction system characterized by comprising: The method comprises: a data preprocessing module for collecting and preprocessing multi-source remote sensing and environmental data and field investigation samples to generate a static data set and a time series data set; a static multi-model training module for training at least two anomaly detection models by using the static data set; a dynamic time series modeling module for training a time series prediction model by using the time series data set; a model selection module for evaluating a plurality of anomaly detection models output by the static multi-model training module and selecting the anomaly detection model with the best performance according to the evaluation result; and a prediction result output module for outputting a prediction result based on the anomaly detection model with the best performance and the time series prediction model. The model verification optimization module receives the prediction results of the performance-optimized abnormality detection model and the time series prediction model, verifies the prediction results by using real-time data, adjusts model parameters, and optimizes prediction performance.

10. The system of claim 9, wherein, Also included are: The visualization module generates a risk space distribution map and a dynamic time series atlas based on the prediction results of the optimized model.

Images