A method for spatial mapping of rainfall erosivity based on cascading machine learning

Abstract

The application discloses a rainfall erosion force spatial mapping method based on cascaded machine learning, constructs a prior knowledge guided cascaded RE estimation model PKiCM, and the PKiCM is a cascaded estimation architecture, the cascaded estimation architecture comprises a classification sub-model and a regression sub-model, and the classification sub-model and the regression sub-model are coupled through decision logic; the classification sub-model and the regression sub-model integrate spatiotemporal prior knowledge features and perform model training through a multi-source heterogeneous data set and a rainfall erosion force value; the regression sub-model outputs a rainfall erosion force estimated value, splices all rainfall erosion force estimated values, generates a regional rainfall erosion force spatial distribution map, deeply fuses a threshold triggering mechanism and spatiotemporal characteristics of RE and a data-driven paradigm through the cascaded architecture, and thus the accuracy and physical consistency of the RE estimated value are improved.

Description

Technical Field

[0001] This invention belongs to the field of environmental science and geographic information technology, specifically relating to a spatial mapping method for rainfall erosivity based on cascaded machine learning. Background Technology

[0002] Soil erosion seriously threatens the stability of ecosystems and the sustainability of agricultural development, and has become a focal environmental problem of global concern. Studies show that the global annual loss of topsoil due to soil erosion reaches as high as 25-40 billion tons, which not only directly causes crop yield reduction, but also weakens the carbon storage capacity of terrestrial ecosystems and induces the degradation of soil nutrient loss and water retention functions.

[0003] The Universal Soil Loss Equation (USLE) and its revised version (RUSLE) are widely used for the quantitative assessment of soil erosion. Rainfall erosivity (RE) is used to measure the potential for soil stripping and transport induced by rainfall. Unlike relatively stable underlying surface factors such as topography, soil, and vegetation, rainfall erosivity exhibits extremely high sensitivity to climate change; an increase in the frequency and intensity of extreme rainfall events will directly and significantly enhance regional erosion risk. Achieving high-precision, large-scale mapping of rainfall erosivity is of great significance for research on soil erosion evolution and environmental protection.

[0004] The core of rainfall erosivity estimation lies in accurately characterizing the complex nonlinear relationship between rainfall kinetic energy and erosion potential. The erosivity of a single rainfall event is typically represented by the EI30 index, which is the product of the total rainfall kinetic energy E and the maximum 30-minute rainfall intensity I30. Although the derivation of EI30 is based on a well-defined physical mechanism, its dependence on minute-level high temporal resolution data makes it difficult to apply in areas with sparse meteorological stations or scarce data records. To address this, researchers have focused on developing simplified estimation models based on daily, monthly, and annual rainfall. Based on this, regional simplified models developed for China, such as the Xie model and the Zhang model, have shown high local adaptability and have been widely used. However, these models only solve the problem of quantifying rainfall erosivity at discrete stations. In large-scale applications, how to balance spatial continuity and local accuracy to estimate rainfall erosivity at the regional scale remains a key challenge for current research.

[0005] Directly calculating rainfall erosivity based on rain gauge observations and then using spatial interpolation to fill data gaps in areas without rain gauges is a common method for obtaining the spatial distribution of rainfall erosivity. While rain gauges provide the most accurate rainfall observations, the introduction of spatial interpolation introduces significant uncertainty into the final spatial distribution of rainfall erosivity. Although rainfall events themselves exhibit clear spatial trends, strong spatial autocorrelation, and relatively stable spatial variability, the strength of spatial autocorrelation varies depending on the time period or sampling scheme. Therefore, when the density of the rain gauge network is insufficient to cover the influence range of rainfall autocorrelation, the interpolation algorithm cannot accurately construct the true spatial distribution of rainfall erosivity. Studies have clearly shown that there are significant differences in the spatial distribution of rainfall erosivity generated using different interpolation methods, and this difference reduces the reliability of the interpolation results in areas without rain gauges.

[0006] The emergence of high-resolution satellite rainfall data has made it possible to estimate the spatial distribution of large-scale rainfall erosivity. Studies have shown that the application of satellite data has greatly improved the ability to quantify and monitor soil erosion at regional scales. However, there is an inherent systematic bias between satellite and rain gauge observations of rainfall, which is particularly pronounced in capturing heavy rainfall and rainfall in areas with complex topography. Research has clearly indicated that the accuracy of rainfall products decreases with increasing rainfall amount and their applicability in complex regions is severely limited, due to the inherent limitations of their inversion algorithms. These issues directly affect the accuracy of calculating rainfall erosivity using satellite rainfall products. For example, Kim et al. demonstrated the potential of the CMORPH product for mapping rainfall erosivity at intercontinental scales, but still pointed out that the CMORPH product needs further refinement to improve the accuracy of rainfall erosivity estimation; Liu et al. compared the results of various rainfall products in estimating rainfall erosivity in China, pointing out that satellite rainfall products have significant uncertainties in characterizing rainfall erosivity, and different products exhibit significant differences in applicability. Therefore, even if satellite rainfall data provides ideal spatial coverage, it is still impossible to achieve high-precision estimation of regional rainfall erosivity. The satellite-ground fusion method has proven to be an effective approach for estimating large-scale rainfall erosivity, while machine learning models, with their superior nonlinear mapping capabilities and advantages in integrating multi-source heterogeneous data, have been widely applied in rainfall erosivity estimation research. However, as Karpatne et al. pointed out, purely data-driven machine learning methods often face challenges such as inconsistencies in physical logic, poor interpretability, and weak out-of-distribution (OOD) generalization ability due to a lack of adherence to established scientific principles. In light of this, the knowledge-guided machine learning (KGML) paradigm has received widespread attention. It advocates explicitly embedding prior knowledge from specific domains, such as physical laws and logical constraints, into the model architecture or learning algorithm to enhance the interpretability of data-driven models for complex physical processes. In the field of rainfall erosivity estimation, the necessity of this knowledge-guided approach is particularly prominent. First, rainfall erosion processes exhibit a significant threshold-triggered effect, meaning that rainfall only possesses erosive power when it reaches a specific magnitude, which makes rainfall erosivity exhibit a significant discontinuous spatial distribution.

[0007] However, most mainstream machine learning models currently employ a single-stage global regression architecture, neglecting the discontinuity of estimation results caused by trigger thresholds, which weakens the model's ability to estimate rainfall erosivity. Secondly, existing machine learning research largely focuses on feature mapping of independent samples, failing to explicitly capture the inherent spatiotemporal characteristics of rainfall erosivity. These limitations collectively restrict the reliability of existing machine learning models in achieving high-precision rainfall erosivity estimation on a large scale and in complex terrain. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a spatial mapping method for rainfall erosivity based on cascaded machine learning, which overcomes the limitations of traditional rainfall erosivity estimation methods and improves the accuracy and physical consistency of regional rainfall erosivity estimation.

[0009] The specific plan is as follows: A prior-knowledge-informed cascaded RE estimation model PKICM is constructed. The PKICM is a cascaded estimation architecture, which includes a classification sub-model and a regression sub-model. The classification sub-model and the regression sub-model are coupled through decision logic. The classification and regression sub-models integrate spatiotemporal prior knowledge features and are trained using multi-source heterogeneous datasets and rainfall erosivity values, including ground-observed rainfall erosivity (RE). GD and satellite observation of rainfall erosivity RE SDThe regression sub-model outputs a rainfall erosivity estimate, and all rainfall erosivity estimates are stitched together to generate a spatial distribution map of regional rainfall erosivity.

[0010] The decision logic of the classification sub-model and the regression sub-model is that the classification sub-model outputs an erosivity discrimination result. If the erosivity discrimination result is a non-erosive event, the final output is zero. If the erosivity discrimination result is an erosive event, the regression sub-model outputs a rainfall erosivity prediction value.

[0011] The classification sub-model is a classification model built based on the Extreme Gradient Boosting (XGBoost) algorithm. This sub-model uses a dataset and satellite-observed rainfall erosivity (RE) data. SD Using spatiotemporal prior knowledge features as input and the types of rainfall events measured on the ground as target variables, a classification sub-model is trained.

[0012] The regression sub-model is a regression model built based on the extreme gradient boosting (XGBoost) algorithm. This regression sub-model uses a dataset and satellite-observed rainfall erosivity R... ESD Using spatiotemporal prior knowledge features as input, and ground-observed rainfall erosivity R... EGD As the target variable, a regression sub-model is trained.

[0013] The multi-source heterogeneous dataset includes rain gauge observation data, CHIRPS data, NDVI data, spatial information data, and temporal information data. Preprocessing is also included before training with the multi-source heterogeneous dataset, which includes integrity screening, temporal matching, and spatial matching.

[0014] The ground-based observed rainfall erosivity RE GD and satellite observation of rainfall erosivity RE SD The calculation is performed using a preset formula for calculating rainfall erosivity, which is as follows: Among them, RE i The erosivity of rainfall in the i-th half-month of the year is expressed in MJ·mm / (hm²). 2 ·h); n is the number of days with erosive rainfall during that period; R x For daily rainfall, when R x When the rainfall erosivity is less than 10 mm, the rainfall erosivity is 0, and z is a constant. From May to September, z = 0.3937, and from October to April of the following year, z = 0.3101. The monthly rainfall erosivity is obtained by summing the calculation results of the two and a half months within that month. The ground-based observed rainfall erosivity RE GD It is the rainfall erosivity calculated based on rain gauge data; The satellite-observed rainfall erosivity RESD It is the rainfall erosion force calculated based on satellite data.

[0015] The spatiotemporal prior knowledge features include features for constructing spatial proximity, features for constructing spatial heterogeneity, and features for constructing temporal heterogeneity.

[0016] The method for constructing spatial proximity features is to use... k-nn Algorithms are used to identify the location with the shortest Euclidean distance to the target position. k One observation station; k Using the observations from each observation station as spatial guides, spatial neighborhood features are constructed for rainfall event discrimination and rainfall erosivity estimation tasks, respectively, by utilizing the corresponding rainfall event types and rainfall erosivity values. Rainfall events are categorized into erosive and non-erosive events, with 1 representing an erosive event and 0 representing a non-erosive event. The formulaic expression of spatial proximity features is as follows: in, Represents the distance from the target location k Rainfall event types at each observation station; x The latitude of the target location. y The longitude of the target location. z The elevation of the target location; These represent the types of rainfall events that are closest to the target location at the same time, in descending order of distance. Represents the distance from the target location k Observations of surface rainfall erosivity at each observation station; This represents the observed ground rainfall erosivity at the same time, from closest to furthest from the target location.

[0018] The method for constructing spatial heterogeneity features involves converting the latitude, longitude, and elevation information of the target station into a three-dimensional spatial coordinate vector in a Cartesian coordinate system. Among them, LOC x,y,z LOC represents the coordinate vector of the target site. x LOC represents the latitude and longitude coordinates of the target site. y LOC represents the longitude coordinate vector of the target station. z This represents the elevation coordinate vector of the target station, where R represents the Earth's radius. Latitude of target location xThe radian value, θ is the longitude of the target position. y The radian value.

[0019] The method for constructing temporal heterogeneity features is as follows: treating the year and month as independent category identifiers, and converting them into binary feature vectors of a preset dimension; the temporal feature of the m-th year and j-th month can be represented as: Among them, TRF m,j Represents the vector of year m and month j, TRF year Represents a year vector, TRF month Represents the month vector, with 1 located at the m-th position of the year vector and the j-th position of the month vector, respectively.

[0020] This invention discloses a spatial mapping method for rainfall erosivity based on cascaded machine learning, establishing a cascaded RE estimation model PKICM guided by prior knowledge. The workflow of PKICM mainly consists of two stages: (1) Data preprocessing and RE calculation: This includes data integrity screening and spatiotemporal matching of multi-source heterogeneous datasets, followed by calculation of the ground-observed rainfall erosivity RE. GD and satellite observation of rainfall erosivity RE SD ; (2) Model building: The focus is on developing a cascaded estimation model based on the knowledge-guided machine learning (KGML) paradigm, including a classification sub-model and a regression sub-model. The classification sub-model first identifies the occurrence of erosive events, and the regression sub-model then performs specific RE value estimation for the identified events. The rainfall erosivity estimates of all grid cells in the study area are spliced ​​together to generate a spatial distribution map of regional rainfall erosivity.

[0021] The core design of this invention uses a cascaded architecture to deeply integrate the threshold triggering mechanism and spatiotemporal characteristics of RE with the data-driven paradigm, thereby improving the accuracy and physical consistency of RE estimates. Attached Figure Description

[0022] Figure 1 This is the overall architecture diagram of the present invention.

[0023] Figure 2 It is the spatial consistency and temporal heterogeneity of rainfall events.

[0024] Figure 3 It is the spatial correlation and temporal heterogeneity of rainfall erosivity values.

[0025] Figure 4 This is a table of evaluation indicators and calculation methods.

[0026] Figure 5This is the overall performance graph of the cascaded RE estimation model. Detailed Implementation

[0027] The technical solutions in the embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the implementation of the present invention, and not all of it. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0028] like Figure 1 As shown, a spatial mapping method for rainfall erosivity based on cascaded machine learning is proposed, which constructs a cascaded RE estimation model PKICM guided by prior knowledge. The PKICM is a cascaded estimation architecture, which includes a classification sub-model and a regression sub-model. The classification sub-model and the regression sub-model are coupled through decision logic. The classification and regression sub-models integrate spatiotemporal prior knowledge features and are trained using multi-source heterogeneous datasets and rainfall erosivity values, including ground-observed rainfall erosivity (RE). GD and satellite observation of rainfall erosivity RE SD The regression sub-model outputs a rainfall erosivity estimate, and all rainfall erosivity estimates are stitched together to generate a spatial distribution map of regional rainfall erosivity.

[0029] The decision logic of the classification sub-model and the regression sub-model is that the classification sub-model outputs an erosivity discrimination result. If the erosivity discrimination result is a non-erosive event, the final output is zero. If the erosivity discrimination result is an erosive event, the regression sub-model outputs a rainfall erosivity prediction value.

[0030] The classification sub-model is a classification model built based on the Extreme Gradient Boosting (XGBoost) algorithm. This sub-model uses a dataset and satellite-observed rainfall erosivity (RE) data. SD Using spatiotemporal prior knowledge features as input and the types of rainfall events measured on the ground as target variables, a classification sub-model is trained.

[0031] The regression sub-model is a regression model built based on the extreme gradient boosting (XGBoost) algorithm. This regression sub-model uses a dataset and satellite-observed rainfall erosivity R... ESD Using spatiotemporal prior knowledge features as input, and ground-observed rainfall erosivity R... EGD As the target variable, a regression sub-model is trained.

[0032] The multi-source heterogeneous dataset includes rain gauge observation data, CHIRPS data, NDVI data, spatial information data, and temporal information data. Preprocessing is also included before training with the multi-source heterogeneous dataset, which includes integrity screening, temporal matching, and spatial matching.

[0033] To achieve accurate estimation of rainfall erosivity, this embodiment utilizes a multi-source heterogeneous dataset, including rain gauge observations, satellite observations, and surface environmental factors. Rain gauge observation data is sourced from the "Daily Dataset of China's Surface Climate Data." Satellite observation rainfall data employs the Climate Hazards Group Infrared Rainfall Product version 2.0 (CHIRPSv2.0). This product integrates satellite remote sensing signals and ground station measurements, boasting a high spatial resolution of 0.05° and long-term series data, and has been widely used in global and regional-scale hydrological process simulations. Surface environmental factors include the daily seamless NDVI raster dataset for China developed by Li et al., and the 30m resolution ASTERGDEMv3 dataset for China.

[0034] To ensure the continuity of the spatiotemporal sequence, this embodiment conducted strict quality control on the original sequence, eliminating stations with a large number of missing data, and finally selected 2238 valid observation stations, covering the time span from 1982 to 2019.

[0035] The rainfall erosivity value includes the ground-observed rainfall erosivity RE. GD and satellite observation of rainfall erosivity RE SD The ground-based observed rainfall erosivity RE GD and satellite observation of rainfall erosivity RE SD The calculation is performed using a preset formula for calculating rainfall erosivity, which is as follows: Among them, RE i The erosivity of rainfall in the i-th half-month of the year is expressed in MJ·mm / (hm²). 2 ·h); n is the number of days with erosive rainfall during that period; R x For daily rainfall, when R x When the rainfall erosivity is less than 10 mm, the rainfall erosivity is 0, and z is a constant. From May to September, z = 0.3937, and from October to April of the following year, z = 0.3101. The monthly rainfall erosivity is obtained by summing the calculation results of the two and a half months within that month. The ground-based observed rainfall erosivity RE GD It is the rainfall erosivity calculated based on rain gauge data; The satellite-observed rainfall erosivity RE SD It is the rainfall erosion force calculated based on satellite data.

[0036] The spatiotemporal prior knowledge features include features for constructing spatial proximity, features for constructing spatial heterogeneity, and features for constructing temporal heterogeneity.

[0037] like Figure 2 and Figure 3 As shown, rainfall erosivity exhibits significant spatial correlation and temporal heterogeneity. Figure 2 (a) shows the consistency between rainfall events at a specific location and their neighboring observations, from Figure 2 As shown in (a), erosive rainfall events exhibit a high degree of spatial consistency; Figure 3 (a) is the ground-observed rainfall erosivity RE at a specific location. GD The correlation coefficient between the value and its neighboring observations; and the ground-observed rainfall erosivity RE at the target location. GD The values ​​also showed strong spatial autocorrelation characteristics with neighboring stations; erosive rainfall events and ground-observed rainfall erosivity RE at the target location. GD The spatial dependence of both decreases significantly with increasing spatial distance.

[0038] Meanwhile, ground-based observations of rainfall erosivity RE GD and satellite observation of rainfall erosivity RE SD The correlation between them is moderated by climate seasonality, exhibiting unique temporal heterogeneity. Figure 2 (b) shows the consistency between rainfall events observed by rain gauges and satellites at different times; Figure 3 (b) Surface-observed rainfall erosivity RE at different times GD and satellite observation of rainfall erosivity RE SD The correlation coefficient between them.

[0039] Therefore, this embodiment transforms the aforementioned spatiotemporal prior knowledge into spatial neighborhood features, spatial heterogeneity features, and temporal discretization features. These features are embedded into the classification and regression sub-models of the cascaded model, respectively, to significantly enhance the model's ability to autonomously learn spatiotemporal variation features.

[0040] The method for constructing spatial proximity features is to use... k-nn Algorithms are used to identify the location with the shortest Euclidean distance to the target position. k One observation station; k Using the observations from each observation station as spatial guides, spatial neighborhood features are constructed for rainfall event discrimination and rainfall erosivity estimation tasks, respectively, by utilizing the corresponding rainfall event types and rainfall erosivity values. Rainfall events are categorized into erosive and non-erosive events, with 1 representing an erosive event and 0 representing a non-erosive event. The formulaic expression of spatial proximity features is as follows: in, Represents the distance from the target location k Rainfall event types at each observation station; x The latitude of the target location. y The longitude of the target location. z The elevation of the target location; These represent the types of rainfall events that are closest to the target location at the same time, in descending order of distance. Represents the distance from the target location k Observations of surface rainfall erosivity at each observation station; This represents the observed ground rainfall erosivity at the same time, from closest to furthest from the target location.

[0041] Due to the local heterogeneity of rainfall events and the spherical nature of the Earth, this embodiment further constructs spatial heterogeneity features to enhance the model's ability to learn spatial heterogeneity between distant samples.

[0042] The method for constructing spatial heterogeneity features involves converting the latitude, longitude, and elevation information of the target station into a three-dimensional spatial coordinate vector in a Cartesian coordinate system. Among them, LOC x,y,z LOC represents the coordinate vector of the target site. x LOC represents the latitude and longitude coordinates of the target site. y LOC represents the longitude coordinate vector of the target station. z This represents the elevation coordinate vector of the target station, where R represents the Earth's radius. Latitude of target location x The radian value, θ is the longitude of the target position. y The radian value.

[0043] To ensure that the model can accurately represent the evolution relationship between ground and satellite observations at different time points, the time information was expanded and discretized, i.e., a time heterogeneity feature was constructed. The method for constructing temporal heterogeneity features is as follows: treating the year and month as independent category identifiers, and converting them into binary feature vectors of a preset dimension; the temporal feature of the m-th year and j-th month can be represented as: Among them, TRF m,j Represents the vector of year m and month j, TRF year Represents a year vector, TRFmonth Represents the month vector, with 1 located at the m-th position of the year vector and the j-th position of the month vector, respectively.

[0044] Compared to traditional continuous numerical encoding, this embodiment can treat the year and month as independent category identifiers. For the year range from 1982 to 2019, the year can be converted into a 38-dimensional binary feature vector and the month into a 12-dimensional binary feature vector to avoid the linear assumption bias introduced by continuous numerical values. This discretization labeling method gives the model great flexibility, enabling it to learn autonomously and capture complex heterogeneous correlation features at specific time points, such as the year of extreme climate events.

[0045] In this embodiment, the classification sub-model focuses on the identification of erosive rainfall events. The classification sub-model is built based on the Extreme Gradient Boosting (XGBoost) algorithm. The classification sub-model uses the Normalized Difference Vegetation Index (NDVI) of the dataset and the RE of satellite-observed rainfall. SD Spatial proximity features for rainfall event classification (Type) k Using spatial heterogeneity features (LOC) and temporal heterogeneity features (TRF) as inputs, and ground-measured rainfall event types as target variables, a classification sub-model is trained. The classification sub-model is represented as: .

[0046] The regression sub-model focuses on rainfall erosivity estimation. It is a regression model built using the Extreme Gradient Boosting (XGBoost) algorithm. This sub-model uses the Normalized Difference Vegetation Index (NDVI) of the dataset and satellite-observed rainfall erosivity (RE) as its parameters. SD Spatial proximity features RE k Using spatial heterogeneity characteristics (LOC) and temporal heterogeneity characteristics (TRF) as inputs, and ground-observed rainfall erosivity (RE) as input, the results were analyzed. GD Use it as the target variable to train a regression sub-model; The regression sub-model is represented as follows: .

[0047] The final output of the cascaded RE estimation model PPiCM is the rainfall erosivity prediction value (RE). pre It is determined by both the classification results and the regression estimates. Specifically, if the classification sub-model identifies non-erosive events, i.e., RE... cls = 0, then the final output is directly assigned zero; if the event is identified as an erosive event, i.e., RE cls = 1, then the model will use the rainfall erosivity estimate RE calculated by the regression sub-model. pre .

[0048] For the prior knowledge-guided cascaded RE estimation model PPiCM in this embodiment, a spatial ten-fold cross-validation strategy combining stratified sampling is used for performance evaluation.

[0049] Stratified sampling based on geographical partitions is used to ensure the representativeness of the training and validation set sample distributions, and to guarantee that each sample participates in the validation set and only once. This strategy effectively reduces evaluation bias caused by spatial clustering of samples in specific regions.

[0050] For a given validation sample, its nearest neighbor features are retrieved only from the training set sites, thus simulating a scenario for estimating rainfall erosivity in areas without observation stations. This setup ensures robustness and reliability in accuracy evaluation. The hyperparameters used in the regression and classification model branches of PPiCM were optimized using raytune, with k = 6 set for spatial nearest neighbor feature selection.

[0051] The selected evaluation metrics include root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 The Critical Success Index (CSI), Probability of Detection (POD), and False Alarm Rate (FAR) are among the metrics used. These metrics are calculated as follows: Figure 4 As shown.

[0052] Figure 4 middle, and These represent the predicted value and the reference value, respectively. and They are and The mean, and They are and The standard deviation and mean of γ, where γ is the histogram intersection of a given histogram K for the observed model and a histogram L for the simulated model; X represents the number of samples; Y represents the number of times erosive rainfall occurred and was characterized as erosive rainfall; Z represents the number of times erosive rainfall occurred but was not characterized as erosive rainfall; and Z represents the number of times erosive rainfall did not occur but was characterized as erosive rainfall.

[0053] Furthermore, PPiCM and the XGBoost mapping model were compared. XGBoost uses satellite-observed rainfall erosivity (RE). SD Longitude, latitude, elevation, NDVI, year, and month were used as input features for regression and classification training, and the results of the two branches were fused to provide the rainfall erosivity estimation result.

[0054] The estimation results from both PKiCM and XGBoost were evaluated using aggregated outputs obtained through 10-fold validation; the rainfall erosivity results obtained from PKiCM cross-validation were named RE. U The rainfall erosivity results obtained from XGBoost cross-validation are named RE. T .

[0055] Figure 5 (ac) respectively show the observed rainfall erosivity RE SD The root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) of XGBoost and PKiCM estimation results. 2 Comparison chart; from Figure 5 (ac) indicates that the observed rainfall erosivity RE SD RE T With RE U RMSE, MAE and R 2 It can be observed that RE U Its performance was significantly better than RE SD With RE T .

[0056] RE U With RE SD In comparison, RE U RMSE and MAE decreased by 33.7% and 42% respectively, while R² increased by 44%; and RE T In comparison, RE U The RMSE and MAE decreased by 14.4% and 18.3% respectively, and the R² increased by 9.5%.

[0057] at the same time, Figure 5 (df) represents the observed rainfall erosivity RE SD A comparison chart of the Probability of Detection (POD), False Alarm Rate (FAR), and Success Index (CSI) of XGBoost and PKICM estimation results; from Figure 5 As shown in (df), PKICM exhibits the best overall performance, demonstrating the highest POD, the lowest FAR, and the highest CSI; RE U POD relative to RE SD With RE T They increased by 3.3% and 1.8% respectively, while FAR decreased by 41.67% and 12.05%, and CSI increased by 8.64% and 2.33%.

[0058] also, Figure 5 (gi) is used to observe rainfall erosivity RE SDThe root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) of XGBoost and PKICM estimation results in different years. 2 Comparison chart; Figure 5 (gi) The results show that PKICM significantly outperformed XGBoost and CHIRPS throughout the entire study period; Regarding RMSE, except for 1998, RE U The RMSE values ​​were all below 253, and the RMSE in 1998 was 277, still significantly better than the RE in the same year. SD and RE T The performance; in contrast, except for 1986, RE SD The RMSE was above 320 throughout the study period, RE T In most years, the value is above 260; Regarding MAE, except for 1998 and 2016, RE U The MAE values ​​were all below 118, and the MAE values ​​for the two years mentioned above were 123 and 121, respectively; in comparison, RE SD MAE is greater than 145 in all years, RE T In most years, the value is higher than 118; For R², RE U The RE value remained above 0.75 throughout the study period. SD The R² value consistently remained below 0.66, and in some years even below 0.5, while the RE value... T Although R² exceeds 0.75 in some years, it still performs worse than RE. U .

[0059] Furthermore, compared to RE SD With RE T RE U The performance fluctuates less across different years. Specifically, RE U RE SD With RE T The standard deviations of RMSE during the study period were 12.5, 17, and 30.2, respectively, and the standard deviations of MAE were 6.2, 8.8, and 18, respectively. 2 The standard deviations were 0.016, 0.019 and 0.057, respectively.

[0060] This invention discloses a spatial mapping method for rainfall erosivity based on cascaded machine learning, establishing a cascaded RE estimation model PKICM guided by prior knowledge. The workflow of PKICM mainly consists of two stages: (1) Data preprocessing and RE calculation: This includes data integrity screening and spatiotemporal matching of multi-source heterogeneous datasets, followed by calculation of the ground-observed rainfall erosivity RE. GD and satellite observation of rainfall erosivity RE SD ; (2) Model building: The focus is on developing a cascaded estimation model based on the knowledge-guided machine learning (KGML) paradigm, including a classification sub-model and a regression sub-model. The classification sub-model first identifies the occurrence of erosive events, and the regression sub-model then performs specific RE value estimation for the identified events. The rainfall erosivity estimates of all grid cells in the study area are spliced ​​together to generate a spatial distribution map of regional rainfall erosivity.

[0061] The core design of PKPiCM lies in the deep coupling of the discrimination logic for erosive rainfall, the spatiotemporal characteristics of precipitation evolution (RE), and the machine learning architecture. According to the calculation mechanism of RE, erosion energy is only generated when rainfall reaches a specific threshold; otherwise, the RE value should be strictly zero. However, traditional unified global regression models typically use a single continuous function to fit the entire interval of the rainfall signal. This architectural flaw leads to statistical interference between a large number of non-erosive samples (zero-value data) and non-zero erosion signals, resulting in significant smoothing errors.

[0062] To address this limitation, we construct a cascaded estimation architecture. First, the classification stage is responsible for determining the erosive properties of rainfall events. If a rainfall event is classified as non-erosive, its output value is directly assigned zero; only when the event is classified as erosive is the sample passed to the regression stage for specific RE estimation. This cascaded architecture is implemented based on the Extreme Gradient Boosting (XGBoost) algorithm, fully leveraging its superior ability to characterize complex nonlinear relationships, thereby achieving accurate estimation of the spatiotemporal patterns of RE under multi-source geographical covariate constraints.

[0063] The core design of this invention uses a cascaded architecture to deeply integrate the threshold triggering mechanism and spatiotemporal characteristics of RE with the data-driven paradigm, thereby improving the accuracy and physical consistency of RE estimates.

[0064] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.

Claims

1. A spatial mapping method for rainfall erosivity based on cascaded machine learning, characterized in that: A prior-knowledge-informed cascaded RE estimation model (PKiCM) is constructed. The PKICM is a cascaded estimation architecture, which includes a classification sub-model and a regression sub-model. The classification sub-model and the regression sub-model are coupled through decision logic. The classification and regression sub-models integrate spatiotemporal prior knowledge features and are trained using multi-source heterogeneous datasets and rainfall erosivity values, including ground-observed rainfall erosivity (RE). GD and satellite observation of rainfall erosivity RE SD The regression sub-model outputs a rainfall erosivity estimate, and all rainfall erosivity estimates are stitched together to generate a spatial distribution map of regional rainfall erosivity.

2. The spatial mapping method for rainfall erosivity based on cascaded machine learning according to claim 1, characterized in that: The decision logic of the classification sub-model and the regression sub-model is that the classification sub-model outputs an erosivity discrimination result. If the erosivity discrimination result is a non-erosive event, the final output is zero. If the erosivity discrimination result is an erosive event, the regression sub-model outputs a rainfall erosivity prediction value.

3. The spatial mapping method for rainfall erosivity based on cascaded machine learning according to claim 1, characterized in that: The classification sub-model is a classification model built based on the Extreme Gradient Boosting (XGBoost) algorithm. This sub-model uses a dataset and satellite-observed rainfall erosivity (RE) data. SD Using spatiotemporal prior knowledge features as input and the types of rainfall events measured on the ground as target variables, a classification sub-model is trained.

4. The spatial mapping method for rainfall erosivity based on cascaded machine learning according to claim 1, characterized in that: The regression sub-model is a regression model built based on the extreme gradient boosting (XGBoost) algorithm. This regression sub-model uses a dataset and satellite-observed rainfall erosivity R... ESD Using spatiotemporal prior knowledge features as input, and ground-observed rainfall erosivity R... EGD As the target variable, a regression sub-model is trained.

5. The spatial mapping method for rainfall erosivity based on cascaded machine learning according to claim 1, characterized in that: The multi-source heterogeneous dataset includes rain gauge observation data, CHIRPS data, NDVI data, spatial information data, and temporal information data. Preprocessing is also included before training with the multi-source heterogeneous dataset, which includes integrity screening, temporal matching, and spatial matching.

6. The spatial mapping method for rainfall erosivity based on cascaded machine learning according to claim 1, characterized in that: The ground-based observed rainfall erosivity RE GD and satellite observation of rainfall erosivity RE SD The calculation is performed using a preset formula for calculating rainfall erosivity, which is as follows: Among them, RE i The erosivity of rainfall in the i-th half-month of the year is expressed in MJ·mm / (hm²). 2 ·h); n is the number of days with erosive rainfall during that period; R x For daily rainfall, when R x When the rainfall erosivity is less than 10 mm, the rainfall erosivity is 0, and z is a constant. From May to September, z = 0.3937, and from October to April of the following year, z = 0.3101. The monthly rainfall erosivity is obtained by summing the calculation results of the two and a half months within that month. The ground-based observed rainfall erosivity RE GD It is the rainfall erosivity calculated based on rain gauge data; The satellite-observed rainfall erosivity RE SD It is the rainfall erosion force calculated based on satellite data.

7. The spatial mapping method for rainfall erosivity based on cascaded machine learning according to claim 1, characterized in that: The spatiotemporal prior knowledge features include features for constructing spatial proximity, features for constructing spatial heterogeneity, and features for constructing temporal heterogeneity.

8. The spatial mapping method for rainfall erosivity based on cascaded machine learning according to claim 7, characterized in that: The method for constructing spatial proximity features is to use... k-nn Algorithms are used to identify the location with the shortest Euclidean distance to the target position. k One observation station; k Using the observations from each observation station as spatial guides, spatial neighborhood features are constructed for rainfall event discrimination and rainfall erosivity estimation tasks, respectively, by utilizing the corresponding rainfall event types and rainfall erosivity values. Rainfall events are categorized into erosive and non-erosive events, with 1 representing an erosive event and 0 representing a non-erosive event. The formulaic expression of spatial proximity features is as follows: in, Represents the distance from the target location k Rainfall event types at each observation station; x The latitude of the target location. y The longitude of the target location. z The elevation of the target location; These represent the types of rainfall events that are closest to the target location at the same time, in descending order of distance. Represents the distance from the target location k Observations of surface rainfall erosivity at each observation station; This represents the observed ground rainfall erosivity at the same time, from closest to furthest from the target location.

9. The spatial mapping method for rainfall erosivity based on cascaded machine learning according to claim 7, characterized in that: The method for constructing spatial heterogeneity features involves converting the latitude, longitude, and elevation information of the target station into a three-dimensional spatial coordinate vector in a Cartesian coordinate system. Among them, LOC x,y,z LOC represents the coordinate vector of the target site. x LOC represents the latitude and longitude coordinates of the target site. y LOC represents the longitude coordinate vector of the target station. z This represents the elevation coordinate vector of the target station, where R represents the Earth's radius. Latitude of target location x The radian value, θ is the longitude of the target position. y The radian value.

10. The spatial mapping method for rainfall erosivity based on cascaded machine learning according to claim 7, characterized in that: The method for constructing temporal heterogeneity features is as follows: treating the year and month as independent category identifiers, and converting them into binary feature vectors of a preset dimension; the temporal feature of the m-th year and j-th month can be represented as: Among them, TRF m,j Represents the vector of year m and month j, TRF year Represents a year vector, TRF month Represents the month vector, with 1 located at the m-th position of the year vector and the j-th position of the month vector, respectively.

Images