Agricultural drought monitoring method based on spatio-temporal adaptive filling and XGBoost downscaling

Abstract

The application provides an agricultural drought monitoring method based on space-time adaptive filling and XGBoost downscaling, which comprises the following steps: taking each missing pixel in remote sensing data as a processing unit, implementing preliminary reconstruction of time dimension missing values through an improved Savitzky-Golay filter, implementing preliminary reconstruction of space dimension missing values through a geographical weighted regression model considering environmental similarity, and adaptively coupling the two reconstruction results; generating space-time continuous low-resolution soil moisture data through consistency correction of multi-source products; constructing a feature set fusing a temperature vegetation drought index, performing two-stage feature screening, and generating a high spatial resolution space-time seamless soil moisture data set through an XGBoost model optimized by Optuna; constructing a per-pixel long-time series soil moisture time series, calculating a standardized soil humidity index, and realizing drought grade division. The application significantly improves the space-time continuity, downscaling accuracy, large-scale generalization ability of passive microwave soil moisture products, and the accuracy and cross-region comparability of agricultural drought identification.

Description

Technical Field

[0001] This invention relates to the technical field of remote sensing data processing and agricultural drought monitoring, and in particular to an agricultural drought monitoring method based on spatiotemporal adaptive filling and XGBoost downscaling. Background Technology

[0002] With the intensification of global climate change and the deepening of refined water resource management, high-precision, long-term, and large-scale agricultural drought monitoring has become a core requirement for ensuring food security and sustainable agricultural development. Soil moisture, as the most direct physical parameter characterizing agricultural drought, has its spatiotemporal continuity and spatial resolution directly determining the reliability and precision of drought monitoring. Passive microwave remote sensing, with its sensitivity to soil dielectric properties, all-weather operation, and ability to penetrate vegetation, has become the mainstream method for large-scale soil moisture monitoring. However, due to limitations in the accuracy of inversion algorithms, interference from cloud and rain, radio frequency interference, and complex terrain, passive microwave soil moisture products generally suffer from severe spatiotemporal data gaps. Furthermore, constrained by antenna size and relatively weak natural radiation energy, its spatial resolution is typically only 25km to 50km, which is insufficient to meet the needs of refined regional-scale agricultural drought monitoring.

[0003] Traditional methods for reconstructing soil moisture deficits, such as spatial interpolation methods like spline interpolation, inverse distance weighting, and ordinary kriging, and time-series filtering methods like Savitzky-Golay filtering and HANTS filtering, primarily rely on single-dimensional statistical information and struggle to utilize both spatiotemporal information. Especially in areas with severe continuous data gaps or high spatial heterogeneity, reconstruction accuracy significantly decreases, leading to unreliable subsequent drought monitoring results. Traditional downscaling methods, such as empirically based multinomial regression, geographically weighted regression, and semi-empirical Dispatch (disaggregation based on a physical and theoretical scale change) methods, mainly rely on pre-defined linear or simplified physical relationships. They struggle to effectively characterize the complex nonlinear interactions between the Temperature-Vegetation-Drought Index (TVDI), multi-source environmental parameters (topography, meteorology, etc.), and soil moisture. Downscaling errors increase significantly, particularly in areas with drastic vegetation cover changes and complex topography.

[0004] Although some studies in recent years have attempted to introduce machine learning methods such as random forests and artificial neural networks into soil moisture downscaling tasks, improving prediction accuracy in specific scenarios, their inherent static feature fusion strategies still suffer from shortcomings such as insufficient adaptive sensing capabilities and inadequate integration of multi-source information, failing to meet the requirements of large-scale monitoring for high-precision and robust soil moisture data. While existing improved models based on ensemble learning have made progress in local areas, they still have the following limitations in long-term, large-scale agricultural drought monitoring applications: first, they are unable to effectively alleviate the problem of discontinuous monitoring data caused by spatiotemporal gaps in original passive microwave products; second, downscaling results often suffer from spatial heterogeneity loss and blurred details of local drought conditions; and third, they rely heavily on preset or static fusion weights and model structures, making it difficult to dynamically and adaptively integrate features based on the actual land cover, climate conditions, and spatial distribution characteristics of soil moisture in the input images. These problems severely restrict the reliability of soil moisture data in practical applications for agricultural drought assessment, drought early warning, and water resource management.

[0005] Therefore, developing an agricultural drought monitoring method that can achieve seamless spatiotemporal reconstruction and nonlinear adaptive downscaling of passive microwave soil moisture products is a technical challenge that urgently needs to be solved in the fields of remote sensing data processing and agricultural drought monitoring.

[0006] Problems with existing technologies include:

[0007] ① Passive microwave soil moisture products are spatiotemporally discontinuous, making it difficult for missing value reconstruction methods to effectively utilize both spatiotemporal information. Existing missing value reconstruction methods are mainly divided into spatial interpolation methods (such as inverse distance weighting and ordinary kriging) and time series filtering methods (such as Savitzky-Golay filtering and HANTS filtering). Spatial interpolation methods rely solely on the spatial correlation of neighboring pixels, resulting in a significant decrease in reconstruction accuracy in areas with continuous data loss or strong spatial heterogeneity. Time series filtering methods utilize only the evolutionary patterns of the time dimension, limiting their reconstruction capabilities for long-term continuous data loss (such as during the winter snow cover period), and are prone to artifacts at sequence boundaries. Both types of methods fail to effectively utilize both spatiotemporal information, particularly lacking an adaptive judgment mechanism for the similarity of the environment surrounding missing pixels. This leads to significant uncertainty in the soil moisture data after missing value reconstruction, making it difficult to support the reliability of subsequent agricultural drought monitoring.

[0008] ② Multi-source passive microwave products exhibit significant systemic biases, making long-term time-series stitching difficult. Sensor parameters, inversion algorithms, and transit times differ among products such as AMSR-E (The Advanced Microwave Scanning Radiometer for EOS), SMOS (Soil Moisture and Ocean Salinity), and AMSR2 (The Advanced Microwave Scanning Radiometer 2), resulting in substantial systemic biases. Existing correction methods often employ global linear correction, failing to achieve pixel-by-pixel temporal matching for refined correction. Even after correction, noticeable temporal stitching breaks remain, making it impossible to generate long-term, spatiotemporally continuous soil moisture datasets.

[0009] ③ Downscaling methods are insufficient in characterizing the nonlinear relationships between soil moisture and multi-source environmental parameters, and often employ static fusion strategies. Existing downscaling methods based on empirical statistics (such as multinomial regression and geographic weighted regression) rely on pre-defined linear relationships or simple spatial weights, making it difficult to effectively capture the complex nonlinear interactions between soil moisture and multi-source environmental parameters such as TVDI (Normalized Difference Vegetation Index), NDVI (Land Surface Temperature), topography, and meteorology. Semi-empirical methods, while introducing physical constraints, are still limited by the assumptions of simplified physical models, resulting in limited applicability under complex land surface conditions. More importantly, existing methods often use fixed feature fusion weights or static model structures, failing to dynamically and adaptively adjust based on the actual land cover, climate conditions, and spatial distribution characteristics of soil moisture in the input image. This rigid fusion mechanism leads to a lack of targeted perception and modeling capabilities when facing soil moisture inversion in different climate zones, seasons, and land surface types, resulting in problems such as loss of spatial heterogeneity in downscaling results and blurred details of local drought conditions.

[0010] In summary, existing technologies lack a method that can simultaneously address the dual bottlenecks of spatiotemporal discontinuity and low spatial resolution in passive microwave soil moisture products, and achieve end-to-end mapping from raw remote sensing data to agricultural drought levels. Summary of the Invention

[0011] To address the technical problems of existing passive microwave soil moisture products, such as spatiotemporal discontinuity, low spatial resolution, and insufficient accuracy in agricultural drought monitoring, this invention proposes an agricultural drought monitoring method based on spatiotemporal adaptive filling and XGBoost downscaling. By synergistically utilizing multi-source passive microwave products and coupling temporal and spatial information for missing value reconstruction, an XGBoost ensemble learning model is introduced to achieve nonlinear adaptive downscaling, and a precise drought mapping system is established based on a standardized soil moisture index, thereby achieving accurate and continuous monitoring of long-term, large-scale agricultural drought.

[0012] To achieve the above objectives, the technical solution of this invention is as follows: an agricultural drought monitoring method based on spatiotemporal adaptive filling and XGBoost downscaling, comprising the following steps:

[0013] Step 1: Using each missing pixel in the remote sensing data as a processing unit, preliminary reconstruction of the temporal dimension missing values ​​is achieved through an improved Savitzky-Golay filter, and preliminary reconstruction of the spatial dimension missing values ​​is achieved through a geographic weighted regression model that takes into account environmental similarity. The least squares method is used to adaptively couple the two reconstruction results. Spatiotemporally continuous low-altitude resolution soil moisture data is generated through multi-source product consistency correction.

[0014] Step 2: Construct a feature set that integrates temperature, vegetation drought index and multiple environmental factors. After two-stage feature screening, use the Optuna-optimized XGBoost model to downscale the generated low-resolution soil moisture data to generate a high spatial resolution spatiotemporal seamless soil moisture dataset.

[0015] Step 3: Construct a pixel-by-pixel long-time soil moisture time series using a high spatial resolution spatiotemporal seamless soil moisture dataset, calculate the standardized soil moisture index, and classify drought levels based on the standardized soil moisture index.

[0016] Preferably, the improved method for preliminary reconstruction of missing time-series dimensions using Savitzky-Golay filtering is as follows:

[0017] For a missing pixel s in remote sensing data, in a continuous time series... In the middle, for the target time step j, by analyzing the missing pixels s within the time window [ Weighted regression was performed on the effective observations within the target time step to obtain the missing pixel s. Reconstructed soil moisture time series observations:

[0018] ;

[0019] Where T is the total number of time steps. Let be the valid observation of the missing pixel s at time step t, and let i be the time offset index relative to the target time step j. It is half the size of the sliding window, and N is the number of convolution kernels, the size of which is equal to the size of the sliding window; For the missing pixel s at time step Valid observations, The weights are Gaussian window function weights. These are the corrected filter coefficients;

[0020] The method for preliminary reconstruction of spatial dimension missing values ​​using the geographic weighted regression model that takes into account environmental similarity is as follows: For missing pixels s under the same time index, spatial dimension reconstruction is performed using information from their spatially neighboring pixels to obtain the spatial dimension of the missing pixel s at the target time step. Reconstructed spatial observations of soil moisture:

[0021] ;

[0022] In the formula, These are the geographic coordinates of the missing pixel s; This represents the number of effective neighboring pixels within the pixel sliding window. For the first pixel in the sliding window Effective observations of neighboring pixels; It is the local regression intercept term. It is the first The local regression slope of the neighboring pixels; For the random error term of the missing pixel s;

[0023] The method of adaptively coupling the two reconstruction results using the least squares method is as follows: the missing pixel s at the target time step Reconstructed soil moisture data ;

[0024] Soil moisture observations were fitted using the least squares method with Savitzky-Golay filtering and the GWR model to reconstruct the soil moisture value of pixel s with high accuracy by integrating spatiotemporal information. and All are fitting coefficients.

[0025] Preferably, the method for multi-source product consistency correction is as follows: using a linear regression model, taking the soil moisture data reconstructed from missing values ​​of AMSR-E products as the benchmark dataset, performing pixel-by-pixel consistency correction on the soil moisture data reconstructed from missing values ​​of SMOS and AMSR2 products, and then splicing the corrected multi-source data in time series to finally generate a long-term spatiotemporally continuous soil moisture dataset.

[0026] Preferably, the Gaussian window function weights ,in, The standard deviation of the Gaussian window; the corrected filter coefficients. Obtained through polynomial least squares fitting;

[0027] The local regression intercept and local regression slope are solved using the weighted least squares method, and their matrix representation is as follows:

[0028] ;

[0029] In the formula, vector It is ( A +1)*1 dimensional column vector containing all regression coefficients at the locations of missing pixels, and the vector... ; For the valid observations of the missing pixel s, The effective observation value of the neighboring pixels, This is the spatial weight matrix;

[0030] The method for solving the spatial weight matrix is ​​as follows: classify pixels based on distance: let the pixel sliding window be within Among the valid pixels, there are similar pixels. One, using subscripts Indicates that there are f dissimilar pixels, indicated by subscripts. Indicates; the first in the sliding window Weights of similar pixels and the Weights of dissimilar pixels They are respectively:

[0031] , ;

[0032] In the formula, and These are the distances from similar pixel q and dissimilar pixel g to the missing pixel s, respectively. and These are the weighting coefficients for similar pixels and the weighting coefficients for dissimilar pixels, respectively.

[0033] The least squares method was used to fit the spatial and temporal observations of soil moisture to obtain the fitting coefficients. and The implementation method is as follows: construct the sum of squares. ;in, It is the number of effective pixels in a remote sensing image. It is an index of valid samples. It is the first The observation value of each effective pixel, and They are the first Spatial and temporal soil moisture observations of 1 effective pixel; when the sum of squares When the minimum value is reached, the fitting coefficient can be obtained. and ;

[0034] The method for per-pixel consistency correction is as follows: For each pixel r, the corrected soil water value is: In the formula, The effective observations of the SMOS or AMSR2 product at pixel r before correction. and These are the time means of the baseline dataset and the uncorrected dataset at cell r, respectively. Let r be the scalar scaling factor for pixel r, and the calculation formula is: In the formula, and These represent the standard errors of the baseline dataset and the uncorrected dataset, respectively. It is the correlation coefficient between the baseline dataset and the uncorrected dataset.

[0035] Preferably, the features used to construct the feature set integrating the Temperature-Vegetation-Drought Index (TVDI) and multiple environmental factors include: TVDI, surface temperature and its derived features, Normalized Difference Vegetation Index (NDVI), Digital Elevation Model (DEM) and its derived topographic factors, meteorological parameters, and soil property parameters; the data of all features are resampled to a uniform low resolution to obtain the input feature set.

[0036] The derived features include diurnal temperature range or temperature spatiotemporal gradient; the derived topographic factors include at least one of slope, aspect, and topographic humidity index; the meteorological parameters include precipitation, air temperature, solar radiation, and wind speed; and the soil property parameters include soil texture or field water holding capacity.

[0037] The two-stage feature selection method is as follows: a recursive feature elimination method is used to perform initial feature screening on the features in the input feature set; the average SHAP value of the candidate features selected in the initial screening is calculated using the SHAP global interpretation method, and candidate features with an average SHAP value lower than a preset threshold are removed, while the core features are retained to form the final feature subset.

[0038] Preferably, the training method of the XGBoost model is as follows: the long-term spatiotemporal continuous soil moisture dataset generated in step one is used as the target variable, and the final feature subset is used as the input features to form a training dataset; the training dataset is randomly divided into a training set and a validation set according to a certain ratio; the XGBoost model is trained on the training set based on the optimal hyperparameters obtained by Optuna search; the early stopping mechanism is monitored on the validation set to determine the optimal number of trees; after training, a trained XGBoost model that characterizes the mapping relationship between the input features and the target variable is obtained.

[0039] The input feature set is upscaled to a high resolution and input to a pre-trained XGBoost model. Soil moisture values ​​are predicted pixel by pixel, and a high spatial resolution, spatiotemporally seamless soil moisture dataset is output.

[0040] Preferably, the recursive feature elimination method recursively constructs an XGBoost base model and removes the least important features. The implementation method is as follows: using XGBoost as the base model, 5-fold cross-validation is used to evaluate the average accuracy under different numbers of features, and the number of features to be retained when the model accuracy is optimal is determined to obtain a subset of candidate features.

[0041] The XGBoost model approximates the true value by integrating K classification and regression trees, for the th The predicted soil moisture value for each sample is: In the formula, It is an XGBoost model; H is the total number of regression trees; It is the first A regression tree; For the first The input feature vector of each sample; F is the function space of all possible classification and regression trees;

[0042] The loss function of the XGBoost model is the h-th tree regression. regularization term for:

[0043] ;

[0044] In the formula, Let be the total number of leaf nodes in the h-th regression tree. It is the weight of the o-th leaf node; , and These are the hyperparameters used to adjust the penalty strength, L2 regularization strength, and L1 regularization strength coefficient, respectively;

[0045] The hyperparameters of the XGBoost model include tree structure parameters, random sampling parameters, regularization parameters, and learning rate parameters; the Optuna framework combined with the Tree-structured Parzen Estimator is used for automated Bayesian hyperparameter optimization.

[0046] Preferably, the method for calculating the standardized soil moisture index is: extracting the soil moisture time series SM of each pixel in the high-spatial-resolution spatiotemporal seamless soil moisture dataset, and performing probability distribution fitting on the soil moisture time series SM through five theoretical probability distribution functions; using the K-S test to perform hypothesis testing on the goodness of fit of each probability distribution, and determining the ability of each probability distribution to describe the temporal characteristics of pixel soil moisture by calculating the maximum absolute difference between the empirical cumulative frequency and the theoretical cumulative frequency of soil moisture;

[0047] For each pixel, the probability distribution with the maximum goodness of fit of the K-S test probability distribution and passing the significance level is selected as the optimal probability distribution model of the pixel, and the cumulative probability cumulative distribution of the soil moisture value corresponding to the optimal probability distribution model of pixel S is recorded as ; The standardized soil moisture index is calculated as: ; In the formula, is the inverse function of the cumulative probability function of the standard normal distribution;

[0048] The drought grade division method is: standardized soil moisture index SSI > 0 indicates no drought; -1.0 < SSI ≤ 0 indicates light drought; -1.5 < SSI ≤ -1.0 indicates moderate drought; -2.0 < SSI ≤ -1.5 indicates severe drought; standardized soil moisture index SSI ≤ -2.0 indicates extreme drought.

[0049] Preferably, the ranges of the hyperparameters of the XGBoost model include: the maximum depth of the tree ranges from 3 to 7; the minimum sum of sample weights of child nodes ranges from 20 to 120; the minimum loss reduction threshold required for leaf node splitting ranges from 0.0 to 10.0; the sample sampling ratio during training of each tree ranges from 0.6 to 1.0; the feature sampling ratio during training of each tree ranges from 0.6 to 1.0; the L1 regularization coefficient ranges from 0.0 to 5.0; the L2 regularization coefficient ranges from 3.0 to 50.0 on a logarithmic scale, and the learning rate ranges from 0.02 to 0.10 on a logarithmic scale; the early stopping mechanism is to automatically terminate training when the loss function on the validation set does not decrease for 50 consecutive rounds;

[0050] The five theoretical probability distribution functions include Gamma distribution, normal distribution, log-normal distribution, generalized extreme value distribution, and Pearson type III distribution;

[0051] Pixel-by-pixel statistical verification showed that the best-fit distribution of pixels in the study area is a normal distribution, and the standardized soil moisture index is: In the formula, μ and σ are the mean and standard deviation of the soil moisture time series SM, respectively.

[0052] An agricultural drought monitoring system based on spatiotemporal adaptive imputation and XGBoost downscaling includes:

[0053] The data preprocessing module is used to acquire multi-source passive microwave soil moisture data, remote sensing data, meteorological data, and topographic and soil data, and to complete the spatiotemporal matching and preprocessing of the data.

[0054] The missing value reconstruction module, connected to the data preprocessing module, is used to perform temporal reconstruction of improved Savitzky-Golay filtering, spatial reconstruction of the geographic weighted regression model that takes into account environmental similarity, and least squares spatiotemporal coupling to complete the missing value reconstruction of soil moisture data; and to perform pixel-by-pixel consistency correction and temporal stitching of multi-source passive microwave data to generate a long-term spatiotemporal continuous soil moisture dataset.

[0055] The downscaling module, connected to the missing value reconstruction module, is used to construct the input feature set, perform two-stage feature selection, and train the Optuna-optimized XGBoost model to complete the downscaling of long-term spatiotemporal continuous soil moisture datasets and generate high spatial resolution spatiotemporal seamless soil moisture products.

[0056] The drought identification module, connected to the downscaling module, is used to construct long-term soil moisture time series, calculate standardized soil moisture index, and complete drought level classification and drought event identification.

[0057] The beneficial effects of this invention are as follows: Based on the reconstruction of missing values ​​in a temporal-spatial dual-dimensional coupling, it improves the Savitzky-Golay filtering and uses a geographically weighted regression (GWR) model that takes into account environmental similarity to complete the initial reconstruction of missing values ​​in both the temporal and spatial dimensions. High-precision reconstruction of missing pixels is achieved through least-squares coupling, enhancing the temporal continuity and spatial distribution consistency of soil moisture data. To address the systematic bias of multi-source passive microwave products, a pixel-by-pixel refined consistency correction framework is constructed to generate a spatiotemporally continuous soil moisture dataset. Furthermore, an XGBoost downscaling scheme integrating multi-source environmental features is proposed. This involves constructing a multi-dimensional feature set by coupling temperature, vegetation, and drought indices with meteorological, topographic, and soil attributes, followed by two-stage feature selection and Optuna (Optimize-Tuning)-TPE (Tree-structured Parzen) analysis. This invention employs Bayesian-optimized model training (using an estimator) to achieve high-precision spatial downscaling of passive microwave soil moisture products, generating high-resolution, spatiotemporally seamless soil moisture products. Ultimately, it establishes a precise agricultural drought identification system adapted to long-term seamless products. By fitting a standardized soil moisture index with the optimal probability distribution, it achieves accurate identification and spatiotemporal feature extraction of agricultural drought events at different levels, improving the cross-regional comparability and precision of drought monitoring. This invention effectively solves the core problems of existing passive microwave soil moisture products, such as spatiotemporal discontinuity, poor consistency of multi-source data, low spatial resolution, insufficient accuracy in agricultural drought identification, and weak cross-regional comparability. It significantly improves the automation, intelligence, and reliability of refined agricultural drought monitoring, early warning, and drought mitigation decision support.

[0058] This invention constructs a complete technical system for temporal-spatial dual-dimensional coupling, including missing value reconstruction, multi-source data fine correction, multi-feature fusion machine learning downscaling, and adaptive agricultural drought identification. It integrates an improved Savitzky-Golay filtering and GWR coupling reconstruction module that takes into account environmental similarity, a pixel-by-pixel multi-source data consistency correction module, an Optuna-TPE optimized XGBoost downscaling module, and an optimal distribution fitting drought identification module. This significantly improves the spatiotemporal continuity, downscaling accuracy, large-scale generalization ability of passive microwave soil moisture products, as well as the accuracy and cross-regional comparability of agricultural drought identification.

[0059] First, in the missing value reconstruction stage, the temporal-spatial dual-dimensional coupled reconstruction framework proposed in this invention effectively solves the problems of weak high-frequency noise suppression and severe temporal boundary fitting artifacts in the traditional Savitzky-Golay filtering algorithm by introducing a Gaussian window function and Sinc-corrected filter coefficients. By adding an environmental similarity weight to the GWR spatial interpolation model, the interpolation accuracy of complex land surface types with strong spatial heterogeneity is significantly improved. Furthermore, the deep coupling of temporal and spatial dimension reconstruction results is achieved through least squares fitting, fully exploring the complementary information of the two dimensions. While reducing computational complexity, this enhances the model's robust reconstruction capability for temporally continuous missing values ​​and large-area spatial missing values, laying a foundation for the construction of long-term soil moisture datasets that are temporally continuous, spatially complete, and accurate.

[0060] Secondly, in the multi-source data correction and downscaling stage, the pixel-by-pixel refined consistency correction strategy introduced in this invention can adaptively complete the systematic deviation correction of SMOS, AMSR2 products and the benchmark AMSR-E product based on the long-term temporal variation characteristics of different pixels, completely eliminating the temporal splicing breakpoints between multi-source data, and realizing the seamless construction of a long-term, spatiotemporally continuous soil moisture dataset from 2003 to 2023. The RFE (Recursive Feature Elimination)-SHAP (Shapley Additive Explanations) two-stage feature screening strategy designed in the downscaling process, combined with Optuna-TPE Bayesian hyperparameter optimization and a dual-stage early stopping mechanism, effectively solves the core limitations of traditional static downscaling methods, such as strong regional dependence, insufficient generalization ability, and high risk of overfitting. This enables the model to adapt to the complex and diverse regional characteristics of climate and land surface types, maintain stable downscaling accuracy over a large scale, and realize the scale extrapolation of passive microwave soil moisture products from a coarse resolution of 25km to a fine resolution of 1km or even higher, meeting the core needs of refined monitoring of agricultural drought at the county and field scales.

[0061] Furthermore, in the agricultural drought identification stage, this invention integrates a dual-module collaborative system of optimal probability distribution adaptive matching and standardized drought level classification. Through goodness-of-fit testing of five mainstream probability distributions, it adaptively determines the optimal fitting distribution for the soil moisture time series of each pixel, significantly reducing the calculation error of the standardized soil moisture index (SSI). Combining regional agricultural drought occurrence characteristics with World Meteorological Organization standards, it completes a localized drought level classification system, achieving accurate identification of four levels of agricultural drought events—light, moderate, severe, and extreme—and standardized quantification that is comparable across regions. The synergistic effect of these two systems significantly suppresses the interference of regional soil moisture background differences on drought identification results, enhances the model's adaptability to agricultural drought characteristics in different regions, and effectively compensates for the shortcomings of traditional fixed-distribution drought identification methods, such as insufficient accuracy and poor cross-regional comparability.

[0062] Ultimately, the aforementioned modules collectively form a synergistically optimized whole: the temporal-spatial dual-dimensional coupled reconstruction module provides temporally continuous and spatially complete basic soil moisture data; the pixel-by-pixel multi-source correction and adaptive downscaling module achieves the construction of high-resolution, highly generalizable, spatiotemporally seamless soil moisture products; and the adaptive drought identification system completes the accurate mapping and feature extraction from soil moisture dynamics to agricultural drought events. This invention's end-to-end technical system not only achieves automated construction of long-term, high-precision soil moisture products and refined, standardized monitoring of agricultural drought in complex geographical scenarios, but also maintains excellent computational efficiency and operational adaptability, providing reliable and efficient remote sensing analysis technology support for core applications such as refined early warning of regional agricultural drought, drought relief and disaster reduction decision-making, and national food security. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a flowchart of the present invention.

[0065] Figure 2 This is a spatial distribution comparison map of the original passive microwave soil moisture image and the downscaled soil moisture image of Henan Province.

[0066] Figure 3 This is a scatter plot and fitting analysis diagram of the measured soil moisture and downscaled soil moisture datasets for Henan Province.

[0067] Figure 4A comparative chart showing the accuracy assessment of downscaled soil moisture products for different land cover types in Henan Province.

[0068] Figure 5 The results of the verification of drought-affected crop area in Henan Province from 2003 to 2022 are as follows: (a) comparison of drought-affected area extracted based on original soil moisture data with official area, and (b) comparison of drought-affected area extracted based on soil moisture data of this invention with official area.

[0069] Figure 6 This represents the number of years that different regions of Henan Province experienced drought from 2003 to 2023. Detailed Implementation

[0070] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] Example 1

[0072] like Figure 1 As shown, an agricultural drought monitoring method based on spatiotemporal adaptive data imputation and XGBoost downscaling includes: First, developing a spatiotemporal adaptive data imputation algorithm, synergistically utilizing AMSR-E, SMOS, and AMSR2 multi-source passive microwave products, coupled with improved Savitzky-Golay filtering, geographically weighted regression, and least squares methods, to reconstruct missing values ​​in regional passive microwave soil moisture data, generating spatiotemporally continuous low-resolution soil moisture data; Second, integrating multi-source environmental variables such as TVDI, NDVI, LST, topographic factors, and meteorological parameters, constructing an XGBoost-based nonlinear downscaling model to downscale the low-resolution soil moisture data to 1km resolution, generating a spatiotemporally seamless high-resolution soil moisture dataset; Finally, achieving accurate mapping from soil moisture to agricultural drought levels based on a standardized soil moisture index, and systematically analyzing the spatiotemporal evolution characteristics of agricultural drought in a region (such as Henan Province). Through these improvements and innovations, this invention effectively solves the dual bottlenecks of spatiotemporal discontinuity and low resolution in passive microwave soil moisture products, significantly improving the accuracy and spatiotemporal coverage of agricultural drought monitoring. Validation results from 2003 to 2023 on different land cover types show that the present invention is significantly superior to mainstream comparative methods in core indicators such as missing value reconstruction accuracy, downscaling error, and drought identification consistency, while maintaining high computational efficiency and spatiotemporal stability, providing reliable technical support for the dynamic monitoring and early warning of long-term, large-scale agricultural drought.

[0073] This invention proposes a three-step technical solution: First, spatiotemporal adaptive missing value reconstruction to solve the spatiotemporal discontinuity problem of passive microwave soil moisture products and generate spatiotemporally continuous low-resolution (25km) soil moisture data; Second, based on the XGBoost nonlinear downscaling model, to solve the problem of excessively low spatial resolution and generate a 1km resolution spatiotemporally seamless soil moisture dataset; Third, based on SSI-based precise mapping of agricultural drought, to achieve standardized conversion from soil moisture to agricultural drought level and to conduct spatiotemporal evolution characteristic analysis.

[0074] Step 1: Spatiotemporal Adaptive Missing Value Reconstruction. Using each pixel with missing values ​​in the remote sensing data as a processing unit, preliminary reconstruction of the temporal dimension of missing values ​​is achieved through an improved Savitzky-Golay filter. Preliminary reconstruction of the spatial dimension of missing values ​​is achieved through a geographic weighted regression (GWR) model that considers environmental similarity. The least squares method is then used to adaptively couple the two reconstruction results to complete the missing value reconstruction for that pixel. Furthermore, a spatiotemporally continuous low-altitude resolution soil moisture dataset is generated through multi-source product consistency correction.

[0075] (1) Improved Savitzky-Golay filter timing reconstruction

[0076] Savitzky-Golay filtering is a local polynomial regression algorithm based on a moving window. It uses each pixel as a basic unit and employs a high-order polynomial for local fitting in the time dimension. Traditional Savitzky-Golay filtering has poor noise suppression capabilities above the cutoff frequency and is prone to artifacts near data boundaries. This invention improves upon traditional Savitzky-Golay filtering by: (a) introducing a Gaussian window function as fitting weights, significantly improving high-frequency noise suppression capabilities; and (b) using a convolution kernel based on the Sinc function to correct filter coefficients, significantly reducing artifacts near data boundaries.

[0077] For the s-th missing pixel in remote sensing data, let its time series containing soil moisture observations in a continuous time series (such as daily or monthly) be . Where T is the total number of time steps, Let be the soil moisture observation value of the s-th missing pixel at time step t (if data is missing at this time step, then...). (If null), for the target time step j, the reconstructed soil moisture observation value This can be achieved by considering the cell within a time window. We obtain the following by weighted regression of the effective observations within the range:

[0078] ; (1)

[0079] In the formula, s is the index of the missing pixel, j is the index of the target time step, and i is the time offset relative to the target time step j. It is half the size of the sliding window (taken before and after the target time step). (N is the number of time steps, set by the user according to the data time resolution), where N is the number of convolution kernels, and its size is equal to the size of the sliding window (i.e., 2^N). ). For the s-th missing value at time step Soil moisture observation values, The weights of the Gaussian window function (the specific expression is...) ,in (Standard deviation of Gaussian window) These are the corrected filter coefficients (obtained through traditional polynomial least squares fitting). For the s-th missing pixel at time step The temporal reconstruction results of the Savitzky-Golay filter.

[0080] (2) Geographically Weighted Regression (GWR) Model that Considers Environmental Similarity

[0081] Since spatially neighboring pixels with similar meteorological conditions and land surface types typically exhibit similar soil moisture values, this invention, after completing temporal reconstruction, further reconstructs the spatial dimension of the s-th missing pixel under the same time index using information from its spatially neighboring pixels. The model expression is:

[0082] ; (2)

[0083] In the formula, The geographic coordinates of the missing pixel s (read from remote sensing products); This represents the number of effective neighboring pixels within the sliding window. For the first in the window The index of the nearest neighboring cells; For the first The effective soil moisture observations of neighboring pixels are extracted from a 15×15 pixel sliding window centered on pixel s. The result of GWR spatial reconstruction for pixel s; It is the local regression intercept term. It is the first The local regression slope of a neighboring cell reflects the degree of contribution of that neighboring cell to the missing cell. This is a random error term, automatically estimated from the model fitting residuals, requiring no manual assignment.

[0084] To solve for all the aforementioned local parameters, this invention employs the weighted least squares method, whose matrix expression is as follows:

[0085] ; (3)

[0086] In the formula, Not a single value, but a ( A 1-dimensional column vector (+1). This vector completely contains all regression coefficients at the location of the missing pixel:

[0087] ;

[0088] For soil moisture observations with missing pixel s, Soil moisture observations for neighboring pixels. This is the spatial weight matrix. In constructing the spatial weight matrix, this invention introduces environmental similarity judgment: neighboring pixels closer to the missing pixel 's' have higher environmental similarity and should be assigned greater weight. The construction steps are as follows:

[0089] Calculate the Euclidean distance from each neighboring cell within the window to the missing cell s:

[0090] ; (4)

[0091] In the formula, , For the first in the window The geographic coordinates of the neighboring pixels; For the first Spatial distance from neighboring pixels to the missing pixel s. Pixel categories are defined based on distance: Let the distance within the window be... Among the valid pixels, there are similar pixels. One, using subscripts Indicated. There are f dissimilar pixels, indicated by subscripts. Indicates. The first [item] in the window. Weights of similar pixels and the Weights of dissimilar pixels They are respectively:

[0092] , ; (5)

[0093] In the formula, and These are the distances from similar pixel q and dissimilar pixel g to the missing pixel s, respectively. and These are the weighting coefficients for similar and dissimilar pixels, respectively. Based on multiple experiments, the values ​​for these two values ​​are 3 and 1, respectively.

[0094] (3) Results of Least Squares Coupled Temporal Reconstruction and Spatial Reconstruction

[0095] Soil moisture values ​​reconstructed from Savitzky-Golay filtering and the GWR model were fitted using the least squares method to integrate spatiotemporal information and reconstruct the soil moisture value of pixel s with high accuracy.

[0096] ; (6)

[0097] In the formula, This is the reconstructed soil moisture value. and These are the soil moisture values ​​reconstructed based on the GWR model and Savitzky-Golay filtering, respectively.

[0098] To determine the best fit coefficients and The least squares method is used to fit the data obtained from the two methods to obtain the fitting coefficients.

[0099] ; (7)

[0100] in, It is the number of effective pixels in the remote sensing image (i.e., the total number of effective training samples). It is an index of valid samples. It is the first Soil moisture values ​​for each valid sample. and They are the first Soil moisture values ​​obtained from the GWR model and Savitzky-Golay filtering for one effective sample. This is the sum of squares of the differences between the weighted sum of the two reconstruction results and the true value on all training samples.

[0101] When the sum of squares When the minimum value is reached, the fitting coefficients can be obtained using formulas (8) and (9). and .

[0102] =0; (8)

[0103] =0; (9)

[0104] (4) Consistency Correction of Multi-Source Passive Microwave Soil Moisture Data: Due to systematic biases among different passive microwave remote sensing products such as AMSR-E, SMOS, and AMSR2, this invention uses soil moisture data reconstructed from missing values ​​in the AMSR-E product as the benchmark dataset. It performs pixel-by-pixel consistency correction on the soil moisture data reconstructed from missing values ​​in the SMOS and AMSR2 products. The corrected multi-source data are then time-series stitched together to generate a long-term (2003-2023) spatiotemporally continuous soil moisture dataset. This invention employs a linear regression model, using the AMSR-E product as the benchmark, to perform pixel-scale consistency correction on the SMOS and AMSR2 products. For each pixel r, the correction formula is as follows:

[0105] ; (10)

[0106] In the formula, r represents the cell index. This is the corrected soil moisture value. The soil moisture value at pixel r for the SMOS or AMSR2 product before correction. and These are the time means of the baseline dataset and the uncorrected dataset at cell r, respectively. Let r be the scalar scaling factor for pixel r, and its calculation formula is as follows:

[0107] ; (11)

[0108] In the formula, and These represent the standard errors of the baseline dataset and the uncorrected dataset, respectively. It is the correlation coefficient between the baseline dataset and the uncorrected dataset.

[0109] Step 2: Construct a nonlinear downscaling model based on XGBoost: Construct a feature set that integrates temperature, vegetation, drought index and multiple environmental factors. After two-stage feature selection, use the Optuna-optimized XGBoost model to downscale the soil moisture data in the low-resolution soil moisture data generated in Step 1, generating a high spatial resolution spatiotemporal seamless soil moisture dataset.

[0110] The Temperature-Vegetation-Drought Index (TVDI) effectively characterizes the spatiotemporal evolution of soil moisture and is widely used in soil moisture retrieval and downscaling. Considering that climatic and environmental factors such as rainfall, temperature, elevation, and solar radiation can potentially influence the relationship between TVDI and soil moisture, this invention employs the XGBoost algorithm. First, it learns the correlation between low-spatial-resolution TVDI and soil moisture under different climatic and environmental influences. Then, it inputs high-spatial-resolution TVDI and corresponding climatic and environmental factors into the XGBoost model to achieve scale extrapolation of soil moisture products, ultimately obtaining high-spatial-resolution soil moisture data.

[0111] (1) TVDI calculation

[0112] The Temperature Vegetation Dryness Index (TVDI) is used to characterize surface water stress, and its calculation formula is as follows:

[0113] ;(12)

[0114] ; (13)

[0115] ;(14)

[0116] In the formula, the Temperature-Vegetation Aridity Index (TVDI) ranges from 0 to 1, with values ​​closer to 1 indicating drier surfaces and values ​​closer to 0 indicating wetter surfaces; T s The actual surface temperature of the pixel is obtained from remote sensing LST data; T smin The normalized difference vegetation index (NDVI) information, representing the theoretical minimum land surface temperature under the same NDVI conditions, can be obtained from remote sensing NDVI products; T smax This represents the theoretical maximum surface temperature under the same NDVI conditions.

[0117] The regression coefficients a1, b1, a2, and b2 are calculated as follows: LST and NDVI data of all effective pixels in the study area are extracted. The range of normalized differential vegetation index (NDVI) values ​​is divided into several sub-intervals at equal intervals. The 98th percentile of the LST values ​​in each sub-interval is taken as the theoretical maximum surface temperature T for that interval. smax The 2nd quantile is taken as the theoretical minimum surface temperature T in this interval. sminThe least squares linear regression was performed to obtain a1 and b1 in formula (13) and a2 and b2 in formula (14). The regression coefficients obtained by fitting were substituted back into formulas (13) and (14), and then into formula (12) to calculate the Temperature Vegetation Drought Index (TVDI) pixel by pixel. The TVDI calculated by the Temperature Vegetation Drought Index is one of the core input features of the subsequent XGBoost downscaling model.

[0118] (2) XGBoost downscaling model

[0119] XGBoost is an ensemble learning algorithm based on the gradient boosting framework. It is an efficient and scalable tree augmentation system optimized from Gradient Boosting Decision Tree (GBDT). Compared with traditional GBDT, XGBoost introduces several key improvements: First, it uses a second-order Taylor expansion in the objective function (i.e., utilizing information from both the first and second gradients), resulting in faster convergence and more accurate optimization direction compared to GBDT which only uses the first derivative. Second, it adds L1 (Lasso) and L2 (Ridge) regularization terms to the loss function, effectively controlling model complexity and preventing overfitting. Third, it supports column sampling (feature sampling) and row sampling (sample sampling), enhancing the model's generalization ability. Fourth, it has a built-in mechanism for handling missing values, automatically learning the optimal splitting direction for missing values. Fifth, it employs parallelization, cache-aware access, and out-of-kernel computation techniques, significantly improving training efficiency. These characteristics give XGBoost a significant advantage in handling the complex mapping relationship between high-dimensional, nonlinear, multi-source heterogeneous features (such as TVDI, NDVI, LST, precipitation, air temperature, solar radiation, etc.) and the target variable (soil moisture).

[0120] The core logic of the downscaling method in this invention is as follows: First, at a low-resolution scale (25km), the XGBoost model is used to learn the nonlinear regression relationship between multi-source environmental parameters and soil moisture; then, the trained model is applied to the same set of environmental parameters at a high resolution (1km) to predict the spatial distribution of soil moisture at a high resolution. The core assumption of this method is that the statistical relationship between environmental parameters and soil moisture has scale consistency at different spatial resolutions, that is, the mapping relationship learned at low resolution can be transferred to high-resolution applications.

[0121] Specifically, the XGBoost downscaling model constructed in this invention includes the following key elements:

[0122] (1) Construction of input feature set. This invention comprehensively considers multiple environmental variables closely related to soil moisture and constructs a high-dimensional feature set. Specifically, this includes: Temperature-Vegetation-Drought Index (TVDI), which reflects the surface dryness and wetness through the feature space of LST and NDVI and is the core feature of downscaling, calculated by step two (1); Surface temperature (LST) and its derived features (such as diurnal temperature range and temperature spatiotemporal gradient), extracted from remote sensing LST data; Normalized Difference Vegetation Index (NDVI), which characterizes vegetation cover and growth status, obtained from remote sensing NDVI products; Digital Elevation Model (DEM) and its derived topographic factors (slope, aspect, topographic humidity index), used to characterize the influence of topography on soil moisture redistribution, extracted from SRTM (Shuttle RadarTopography Mission) DEM data; Meteorological parameters, including precipitation, air temperature, solar radiation, wind speed, etc., obtained from ERA5 reanalysis data or meteorological station interpolation data, reflecting atmospheric forcing conditions; Soil property parameters (such as soil texture and field water holding capacity), extracted from the World Soil Database or China Soil Database, and included in areas where data is available (from China Soil Database). All the above input features were resampled to a uniform low resolution (25 km) to obtain the input feature set for model training, and a high resolution (1 km) for prediction.

[0123] (2) Two-stage feature selection strategy. The multi-source environmental variables (also known as feature variables) constructed by the above operations have high dimensionality, and there may be multicollinearity problems among the variables. This will not only increase the computational cost of model training, but also significantly reduce the interpretability of the model, and even introduce noise leading to overfitting. To this end, this invention designs a two-stage feature selection strategy. The first stage uses Recursive Feature Elimination (RFE) to perform initial feature screening. RFE locates the minimum feature subset that maximizes the model accuracy by recursively constructing the XGBoost base model and removing the least important features. In specific operations, using XGBoost as the base model, 5-fold cross-validation is used to evaluate the average accuracy under different numbers of features, determine the number of features to retain when the model accuracy is optimal, and obtain the candidate feature subset. The second stage introduces the SHAP (Shapley Additive Explanations) global interpretation method to quantitatively evaluate the independent contribution of the candidate features selected in the first stage. SHAP is based on the Shapley value in game theory and can assign a contribution score to each feature, reflecting the marginal contribution of the feature to the model output. Through SHAP attribution analysis, this invention calculates the average SHAP value for each candidate feature, eliminates variables with weak contributions (average SHAP value below a preset threshold), and ultimately retains 8-12 core features to form the final feature subset. This core feature set is used for the final training of the XGBoost model; features not selected are not used in either the training or prediction phases. This selection process not only reduces model dimensionality but also enhances model interpretability, enabling researchers to clearly identify which environmental factors play a dominant role in soil moisture changes in different regions and seasons.

[0124] (3) Determination of target variable. The target variable in the model training phase is the spatiotemporally continuous low-resolution (25km) soil moisture data generated in step one. This data has undergone spatiotemporally adaptive imputation and multi-source consistency correction, and has high spatiotemporal continuity and accuracy, providing reliable training labels for the downscaling model.

[0125] (4) Mathematical principles of model training. XGBoost approximates the true value by integrating K classification and regression trees (CART), and its prediction function is:

[0126] ; (15)

[0127] In the formula, e is the index of the training sample. It is the XGBoost model for the first Predicted soil moisture values ​​for each sample; H is the total number of regression trees; It is the first A regression tree; For the first The input feature vector of each sample (including multi-source environmental variables such as TVDI, LST, NDVI, DEM, precipitation, temperature, and solar radiation); F is the function space of all possible CART trees.

[0128] To better control model complexity and avoid overfitting, XGBoost's objective function incorporates L1 and L2 regularization terms. L1 regularization adds a sparse constraint to the weights of leaf nodes, while L2 regularization makes the model's learning process smoother by penalizing the squared terms of the weights. The calculation formula is:

[0129] (16)

[0130] In the formula, The regularization term corresponding to the h-th tree; Let be the total number of leaf nodes in the h-th tree. It is the weight of the o-th leaf node; , and These are hyperparameters used to adjust the penalty strength for model complexity, the intensity of L2 regularization, and the coefficient of L1 regularization, respectively. The introduction of regularization terms makes XGBoost more resistant to overfitting than traditional GBDT.

[0131] (5) Optuna-TPE Hyperparameter Optimization Strategy. The XGBoost model contains a large number of hyperparameters, including tree structure parameters (max_depth, min_child_weight, gamma), random sampling parameters (subsample, colsample_bytree), regularization parameters (reg_alpha, reg_lambda), and learning rate parameters (learning_rate). The values ​​of these hyperparameters directly determine the model's prediction accuracy and generalization ability. Traditional manual parameter tuning or grid search methods are inefficient and prone to getting stuck in local optima in high-dimensional parameter spaces. To address this, this invention uses the Optuna framework combined with a tree-structured Parzen Estimator (TPE) for automated Bayesian hyperparameter optimization. Optuna can dynamically adjust the parameter sampling distribution based on historical evaluation results, efficiently exploring high-dimensional parameter spaces within limited computational resources. Specifically, the hyperparameter search ranges set in this invention are as follows: max_depth (maximum tree depth) is searched between 3 and 7 to control the complexity of a single tree and prevent overfitting due to excessive depth; min_child_weight (minimum sum of child node sample weights) is searched between 20 and 120, the larger this parameter value, the more conservative the model, and the better it can suppress overfitting; gamma (minimum loss descent threshold required for leaf node splitting) is searched between 0.0 and 10.0, a larger gamma value will limit the tree branches and simplify the model structure; subsample (sample sampling ratio during training of each tree) is between 0.6 and 1.0. The XGBoost algorithm employs a multi-parameter search: row sampling enhances the model's generalization ability; `colsample_bytree` (the feature sampling ratio per tree during training) is searched between 0.6 and 1.0 to reduce coupling between features through column sampling; `reg_alpha` (L1 regularization coefficient) is searched between 0.0 and 5.0 to promote sparsity of leaf node weights; `reg_lambda` (L2 regularization coefficient) is searched on a logarithmic scale between 3.0 and 50.0 to smooth extreme weights in leaf nodes; and `learning_rate` is searched on a logarithmic scale between 0.02 and 0.10 to control the contribution of each tree to the final prediction. Simultaneously, the maximum number of iterations (`num_boost_round`) is set to 10,000, and an early stopping mechanism (`early_stopping_rounds=50`) is enabled, automatically terminating training when the loss function on the validation set fails to decrease for 50 consecutive iterations, thus dynamically determining the optimal number of trees. The XGBoost model with the optimal hyperparameters is the XGBoost model used in this invention.

[0132] Table 1. Hyperparameter settings for the XGBoost model

[0133]

[0134] The XGBoost model training process, specifically implemented as follows: The long-term, spatiotemporally continuous soil moisture dataset (25 km) generated in step one is used as the target variable. The TVDI calculated in step two (1) and the core feature set retained after two-stage feature filtering (including environmental variables such as LST, NDVI, DEM, precipitation, temperature, and solar radiation, all resampled to 25 km resolution) are used as input features. , constitute the training dataset { , The training dataset was randomly divided into training and validation sets in a 7:3 ratio. On the training set, an XGBoost model was trained using the optimal hyperparameter configuration obtained through Optuna search. On the validation set, early stopping was monitored to determine the optimal number of trees. After training, a model capable of characterizing... → The XGBoost model with mapping relationships.

[0135] High-resolution prediction. The aforementioned high-resolution (1 km) TVDI and corresponding environmental parameters (TVDI, LST, NDVI, DEM, precipitation, temperature, solar radiation, and other feature variables were upscaled from 25 km resolution to 1 km resolution through bilinear interpolation or nearest-neighbor resampling) were input into the trained XGBoost model to predict soil moisture values ​​pixel by pixel, ultimately outputting a 1 km resolution, spatiotemporally seamless soil moisture dataset. This dataset maintains the physical consistency of passive microwave soil moisture products (ensuring this through spatiotemporal reconstruction in step one) while significantly improving the spatial detail representation (achieved through XGBoost downscaling), effectively capturing the spatial heterogeneity of soil moisture at the county and watershed scales, providing a high-quality data foundation for subsequent precise monitoring of agricultural drought.

[0136] Step 3: Agricultural drought identification based on spatiotemporal seamless soil moisture products: Using the generated high spatial resolution spatiotemporal seamless soil moisture dataset, a pixel-by-pixel long-time series of soil moisture is constructed, and the standardized soil moisture index (SSI) is calculated to complete the accurate identification and classification of agricultural drought events.

[0137] The Standardized Soil Moisture Index (SSI) is one of the most effective indicators for evaluating agricultural drought. The present invention first extracts the time-series soil moisture sequence of each pixel from the generated high-resolution soil moisture dataset, performs probability distribution fitting on the time series, and systematically evaluates the applicability of five theoretical probability distribution functions (Gamma distribution, Normal distribution, Lognormal distribution, Generalized Extreme Value distribution, and Pearson Type III distribution). The Kolmogorov-Smirnov (K-S) test is used to conduct hypothesis testing on the goodness of fit of each distribution, and the descriptive ability of each distribution for the temporal characteristics of pixel soil moisture is determined by calculating the maximum absolute difference between the empirical cumulative frequency and the theoretical cumulative frequency of soil moisture. For each pixel, the distribution with the maximum p-value of the K-S test that passes the significance level is selected as the optimal probability distribution model for that pixel, and its cumulative distribution function is denoted as . On this basis, SSI is defined and calculated as:

[0138] ;

[0139] wherein SM is the extracted soil moisture time series of each pixel, is the cumulative probability of the soil moisture value corresponding to the optimal probability distribution model of pixel S, with a value ranging from 0 to 1; is the inverse function of the cumulative probability function of the standard normal distribution, that is, the standard normal quantile function. Statistical verification pixel by pixel shows that the optimal fitting distribution for the vast majority of pixels in the study area is the normal distribution, which is therefore equivalent to the standardization transformation as:

[0140] ; (17)

[0141] wherein μ and σ are respectively the mean value and standard deviation of the soil moisture time series SM of the pixel. According to the drought classification standard recommended by the World Meteorological Organization and combined with the data characteristics of the present invention, the drought grade classification system as shown in Table 2 is established. The agricultural drought grades are divided into: SSI>0 indicates no drought, -1.0<SSI≤0 indicates light drought, -1.5<SSI≤-1.0 indicates moderate drought, -2.0<SSI≤-1.5 indicates severe drought, and SSI≤-2.0 indicates extreme drought. Through this grade division, accurate identification of agricultural drought events of four grades (light, moderate, severe and extreme) and standardized quantification comparable across regions can be realized, which provides a scientific basis for the analysis of spatiotemporal evolution characteristics and early warning decision-making of regional agricultural drought.

[0142] Table 2 SSI values and corresponding drought grades

[0143]

[0144] The improvements of this invention are as follows:

[0145] (1) A high-precision soil moisture missing value reconstruction framework with temporal-spatial dual-dimensional coupling is introduced. In the missing value reconstruction stage, a hierarchical structure of temporal filtering-spatial interpolation-dual-dimensional coupling is adopted. The improved Savitzky-Golay filtering algorithm with overlapping sliding windows is used, and the filter coefficients are modified by Gaussian window function and Sinc function to alleviate the problems of weak high-frequency noise suppression and serious temporal boundary fitting artifacts of traditional SG filtering. A geographic weighted regression model that takes into account environmental similarity is adopted, and pixel similarity weight is introduced on the basis of geographic distance weight (Formula 4-5). While reducing spatial interpolation error, the adaptability to complex land surface types with strong spatial heterogeneity is greatly enhanced. The deep coupling of temporal and spatial dimension reconstruction results is achieved through least squares fitting, which fully explores the complementary information of dual dimensions and solves the core defects of traditional single-dimensional reconstruction methods such as insufficient information utilization and poor robustness. Thus, a basic data system with temporal continuity, spatial integrity and reliable accuracy is constructed for long-term soil moisture datasets.

[0146] (2) A pixel-by-pixel refined multi-source data correction and adaptive downscaling scheme is added. In the multi-source data fusion and downscaling process, a consistency correction strategy based on pixel-level statistical features of the time-series overlap period is introduced. This strategy can adaptively complete the system deviation correction of SMOS, AMSR2 products and the benchmark AMSR-E product according to the long-term time-series change characteristics of different pixels, completely eliminate the time-series splicing breakpoints of multi-source data, and realize the seamless construction of long-term spatiotemporal continuous soil moisture dataset. In the downscaling process, a multi-dimensional feature set integrating temperature, vegetation, drought index, vegetation, meteorology, topography and soil properties is constructed. The RFE-SHAP two-stage feature screening strategy is designed to realize redundant feature removal and core feature adaptive mining. Combined with Optuna-TPE Bayesian hyperparameter optimization and dual-stage early stopping mechanism, the problems of strong regional dependence, insufficient generalization ability and high risk of overfitting of traditional static downscaling methods are effectively solved, and high-precision and robust spatial downscaling of passive microwave soil moisture products in large-scale complex scenarios is realized.

[0147] (3) An integrated and adaptable agricultural drought precision identification system for long-term seamless products. In the drought identification process, the optimal probability distribution adaptive matching and standardized drought level classification are introduced in parallel. The optimal probability distribution matching determines the optimal fitting distribution of the soil moisture time series pixel by pixel adaptively through KS test of five mainstream probability distributions: gamma distribution, normal distribution, log-normal distribution, generalized extreme value distribution, and Pearson-III distribution, which greatly reduces the calculation error of the standardized soil moisture index (SSI). The standardized drought level classification combines the characteristics of agricultural drought occurrence in my country with the standards of the World Meteorological Organization to establish a five-level SSI classification system (no drought, mild drought, moderate drought, severe drought, and extreme drought), completes the localization adaptation of the drought level system, and realizes the precise identification of agricultural drought events of four levels and standardized quantification that is comparable across regions. The synergistic effect of the two significantly suppresses the interference of regional background differences on drought identification results, enhances the adaptability to agricultural drought characteristics in different regions and cross-regional comparability, and realizes the precise extraction and reliable analysis of the spatiotemporal evolution characteristics of agricultural drought.

[0148] Example 2

[0149] An agricultural drought monitoring system based on spatiotemporal adaptive imputation and XGBoost downscaling includes:

[0150] The data preprocessing module is used to acquire multi-source passive microwave soil moisture data (AMSR-E, SMOS, AMSR2), remote sensing data (LST, NDVI), meteorological data (precipitation, temperature, solar radiation), and topographic and soil data (DEM, soil texture), and to complete the spatiotemporal matching and preprocessing of the data. Specifically, all data are unified to the same spatial reference system (latitude and longitude projection), and data from different sources are resampled to the target resolution using bilinear interpolation or nearest-neighbor resampling methods. Quality labels are then applied to each data point to remove invalid values, providing standardized input for subsequent modules.

[0151] The missing value reconstruction module, connected to the data preprocessing module, is used to perform temporal reconstruction of improved Savitzky-Golay filtering, GWR spatial reconstruction that takes into account environmental similarity, and least squares spatiotemporal coupling to complete the missing value reconstruction of soil moisture data. It also performs pixel-by-pixel consistency correction and temporal stitching on multi-source passive microwave data to generate a long-term spatiotemporal continuous soil moisture dataset.

[0152] The downscaling module, connected to the missing value reconstruction module, is used to construct a downscaling feature set, perform two-stage feature selection, train the Optuna-optimized XGBoost model, complete the downscaling of soil moisture data, and generate a high spatial resolution spatiotemporal seamless soil moisture product.

[0153] The drought identification module, connected to the downscaling module, is used to construct long-term soil moisture sequences, calculate the SSI index, and complete drought level classification and drought event identification.

[0154] To comprehensively verify the reliability and regional applicability of the spatiotemporal seamless soil moisture product construction method and agricultural drought identification system proposed in this invention, a full-process, multi-scale accuracy evaluation system was constructed from four dimensions: downscaled product spatial visualization characterization, quantitative accuracy verification of station measured data, cross-performance evaluation under different land cover types, and applicability verification of agricultural drought identification index. This system systematically verifies the advanced nature and practicality of the method proposed in this invention.

[0155] Spatial visualization of soil moisture downscaling products

[0156] Using the downscaling method developed in this invention, spatial downscaling was performed on long-term passive microwave soil moisture data from Henan Province from 2003 to 2023. The downscaled soil moisture products, while retaining the original macroscopic spatial distribution pattern, significantly improved the ability to characterize spatial details. For example... Figure 2 As shown, the downscaling results not only fully preserve the overall gradient characteristics of soil moisture increasing from northeast to southwest in the original passive microwave data of Henan Province, but also finely depict the details of local small-scale spatial changes in soil moisture, solving the core problem of insufficient local detail representation ability of the original coarse-resolution product.

[0157] Quantitative accuracy testing of downscaling products based on site measurement data

[0158] To quantitatively verify the accuracy of downscaled soil moisture products, this invention selects three core quantitative evaluation indicators: Pearson correlation coefficient (R), systematic bias (Bias), and root mean square error (RMSE). The downscaled 1km resolution soil moisture products are then compared pixel-by-pixel with in-situ ground-based measurement data from Henan Province. The verification results are as follows: Figure 3 As shown, the downscaling product exhibits a high degree of agreement with the measured soil moisture at the site in Henan Province. Specifically, R = 0.871, and RMSE remains consistently at 0.058 m. 3 / m 3The relatively low level of accuracy demonstrates that the downscaling algorithm employed in this invention, while upscaling the original 25km resolution passive microwave data to a 1km high resolution, fully preserves the physical accuracy and temporal variation characteristics of the original data, without exhibiting the accuracy degradation problem during the downscaling process. From the sample distribution of the scatter plot, the vast majority of validation samples closely surround the 1:1 standard line, and the linear fitting trend highly coincides with the 1:1 line. Overall, the downscaled soil moisture product generated by this invention demonstrates excellent inversion accuracy and regional applicability in Henan Province, possessing the ability to support long-term, large-scale spatiotemporal dynamic analysis of soil moisture.

[0159] Cross-performance evaluation of downscaled products under different land cover types

[0160] To further comprehensively evaluate the overall performance of downscaling products under different land cover types, cross-validation was conducted using mainstream high-resolution soil moisture products of the same type from both domestic and international sources. The accuracy differences of the downscaling product (PIII), the 1km high-resolution product (PII), the 5km downscaling product (PI), and the original passive microwave product before downscaling were systematically analyzed under different land cover types. The results are as follows: Figure 4 As shown, under all land cover types, the accuracy of the downscaled soil moisture product (PIII) using this invention is significantly better than the original passive microwave product before downscaling. Specifically, the unbiased root mean square error (ubRMSE) is reduced by 25%–59%, and the correlation coefficient (R) is improved by 63%–86%. This result fully demonstrates the effectiveness and advancement of the downscaling algorithm of this invention. Compared with two other similar products, the downscaled data of this invention significantly improves the soil moisture inversion accuracy in all seven land cover types. Specifically, the accuracy in urban, shrubland, green space, and wasteland areas is improved by approximately 11.9%, 14.3%, 14.3%, and 12.8%, respectively, significantly improving the soil moisture estimation effect in complex land surface types and low vegetation cover areas. In summary, the downscaled soil moisture product generated by this invention maintains good inversion accuracy and spatial representativeness under different land cover types.

[0161] Applicability and reliability verification of the agricultural drought identification index

[0162] Agricultural drought primarily affects crops, and the area of ​​drought-affected crops is the most direct indicator of regional drought intensity. Time-series coupling analysis of drought-affected crop area and drought index can effectively verify the applicability and reliability of drought monitoring indices. This invention uses Henan Province as the verification area, comparing and analyzing officially reported statistical data on drought-affected crop area with drought-affected areas extracted from soil moisture products before and after downscaling based on this invention. This aims to evaluate the applicability and reliability of the SSI index in agricultural drought monitoring. The results are as follows: Figure 5As shown in the figure. The results indicate that, between 2003 and 2023, the drought-affected area extracted based on the downscaled soil moisture product of this invention showed a higher degree of agreement with the officially reported drought-affected area, and the correlation coefficient (R) was higher. 2 The SSI index, calculated based on downscaled soil moisture data, improved by 32.5% compared to the original product. This result fully demonstrates that the SSI index calculated based on downscaled soil moisture data of this invention can accurately characterize the occurrence process of regional agricultural drought. It has excellent applicability and good temporal stability in dynamic monitoring of agricultural drought, and can provide reliable quantitative indicator support for refined assessment, early warning and drought relief decision-making of regional-scale agricultural drought.

[0163] Based on the above multi-dimensional and full-process accuracy verification results, the spatiotemporal seamless soil moisture product construction method proposed in this invention can generate long-term soil moisture data with high spatial resolution, high accuracy, and spatiotemporal continuity. The product accuracy and regional adaptability are significantly better than similar mainstream products. The SSI agricultural drought identification system built based on this product can accurately match actual agricultural disasters, has excellent drought monitoring performance, and can fully meet the operational application needs of refined monitoring, early warning, and drought relief and disaster reduction decision-making in agricultural drought.

[0164] (2) Spatial distribution and variation of the frequency of drought years in agricultural drought in Henan Province

[0165] This invention, based on the annual average SSI of each pixel in Henan Province from 2003 to 2023, statistically analyzed the frequency of drought occurrences in the province over 21 years and plotted its spatial distribution map, as shown below. Figure 6 As shown in the figure, the frequency of drought occurrence in most pixels of Henan Province ranges from 7 to 12 years, with significant regional differences in spatial distribution. Areas with frequent drought (more than 10 years of drought) are widely distributed, mainly concentrated in western, central, northern, and eastern Henan. Central and northern Henan have a higher number of drought years and more frequent drought events. Due to topographical influence, the foehn wind effect is significant in northern Henan, leading to hot and dry winds, further exacerbating the drought situation in this region. In contrast, southern Henan has a lower number of drought years and is generally more humid. Overall, the above four regions face more severe drought problems in the long term, while southern Henan is relatively humid and has a lower drought risk. Figure 6 This visually illustrates the spatial pattern, reveals the regional differences in agricultural drought in Henan Province, and provides a reference for formulating drought resistance and disaster reduction strategies by region.

[0166] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An agricultural drought monitoring method based on spatiotemporal adaptive imputation and XGBoost downscaling, characterized in that, Includes the following steps: Step 1: Using each missing pixel in the remote sensing data as a processing unit, preliminary reconstruction of the temporal dimension missing values ​​is achieved through an improved Savitzky-Golay filter, and preliminary reconstruction of the spatial dimension missing values ​​is achieved through a geographic weighted regression model that takes into account environmental similarity. The least squares method is used to adaptively couple the two reconstruction results. Spatiotemporally continuous low-altitude resolution soil moisture data is generated through multi-source product consistency correction. Step 2: Construct a feature set that integrates temperature, vegetation drought index and multiple environmental factors. After two-stage feature screening, use the Optuna-optimized XGBoost model to downscale the generated low-resolution soil moisture data to generate a high spatial resolution spatiotemporal seamless soil moisture dataset. Step 3: Construct a pixel-by-pixel long-time soil moisture time series using a high spatial resolution spatiotemporal seamless soil moisture dataset, calculate the standardized soil moisture index, and classify drought levels based on the standardized soil moisture index.

2. The agricultural drought monitoring method based on spatiotemporal adaptive filling and XGBoost downscaling according to claim 1, characterized in that, The improved Savitzky-Golay filter method for preliminary reconstruction of missing time-series dimensions is as follows: For a missing pixel s in remote sensing data, in a continuous time series... In the middle, for the target time step j, by analyzing the missing pixels s within the time window [ Weighted regression was performed on the effective observations within the target time step to obtain the missing pixel s. Reconstructed soil moisture time series observations: ； Where T is the total number of time steps. Let be the valid observation of the missing pixel s at time step t, and let i be the time offset index relative to the target time step j. It is half the size of the sliding window, and N is the number of convolution kernels, the size of which is equal to the size of the sliding window; For the missing pixel s at time step Valid observations, The weights are Gaussian window function weights. These are the corrected filter coefficients; The method for preliminary reconstruction of spatial dimension missing values ​​using the geographic weighted regression model that takes into account environmental similarity is as follows: For missing pixels s under the same time index, spatial dimension reconstruction is performed using information from their spatially neighboring pixels to obtain the spatial dimension of the missing pixel s at the target time step. Reconstructed spatial observations of soil moisture: ； In the formula, These are the geographic coordinates of the missing pixel s; This represents the number of effective neighboring pixels within the pixel sliding window. For the first pixel in the sliding window Effective observations of neighboring pixels; It is the local regression intercept term. It is the first The local regression slope of the neighboring pixels; For the random error term of the missing pixel s; The method of adaptively coupling the two reconstruction results using the least squares method is as follows: the missing pixel s at the target time step Reconstructed soil moisture data ; Soil moisture observations were fitted using the least squares method with Savitzky-Golay filtering and the GWR model to reconstruct the soil moisture value of pixel s with high accuracy by integrating spatiotemporal information. and All are fitting coefficients.

3. The agricultural drought monitoring method based on spatiotemporal adaptive filling and XGBoost downscaling according to claim 2, characterized in that, The method for multi-source product consistency correction is as follows: using a linear regression model, soil moisture data reconstructed from missing values ​​of AMSR-E products is used as the benchmark dataset. The soil moisture data reconstructed from missing values ​​of SMOS and AMSR2 products are then subjected to pixel-by-pixel consistency correction. Finally, the corrected multi-source data are spliced ​​together in time series to generate a long-term, spatiotemporally continuous soil moisture dataset.

4. The agricultural drought monitoring method based on spatiotemporal adaptive filling and XGBoost downscaling according to claim 3, characterized in that, The Gaussian window function weights ,in, The standard deviation of the Gaussian window; the corrected filter coefficients. Obtained through polynomial least squares fitting; The local regression intercept and local regression slope are solved using the weighted least squares method, and their matrix representation is as follows: ； In the formula, vector It is ( A +1)*1 dimensional column vector containing all regression coefficients at the locations of missing pixels, and the vector... ; For the valid observations of the missing pixel s, The effective observation value of the neighboring pixels, This is the spatial weight matrix; The method for solving the spatial weight matrix is ​​as follows: classify pixels based on distance: let the pixel sliding window be within Among the valid pixels, there are similar pixels. One, using subscripts Indicates that there are f dissimilar pixels, indicated by subscripts. Indicates; the first in the sliding window Weights of similar pixels and the Weights of dissimilar pixels They are respectively: ， ； In the formula, and These are the distances from similar pixel q and dissimilar pixel g to the missing pixel s, respectively. and These are the weighting coefficients for similar pixels and the weighting coefficients for dissimilar pixels, respectively. The least squares method was used to fit the spatial and temporal observations of soil moisture to obtain the fitting coefficients. and The implementation method is as follows: construct the sum of squares. ;in, It is the number of effective pixels in a remote sensing image. It is an index of valid samples. It is the first The observation value of each effective pixel, and They are the first Spatial and temporal soil moisture observations of 1 effective pixel; when the sum of squares When the minimum value is reached, the fitting coefficient can be obtained. and ; The method for per-pixel consistency correction is as follows: For each pixel r, the corrected soil water value is: In the formula, The effective observations of the SMOS or AMSR2 product at pixel r before correction. and These are the time means of the baseline dataset and the uncorrected dataset at cell r, respectively. Let r be the scalar scaling factor for pixel r, and the calculation formula is: In the formula, and These represent the standard errors of the baseline dataset and the uncorrected dataset, respectively. It is the correlation coefficient between the baseline dataset and the uncorrected dataset.

5. The agricultural drought monitoring method based on spatiotemporal adaptive filling and XGBoost downscaling according to any one of claims 1-4, characterized in that, The features used to construct the feature set integrating the Temperature-Vegetation-Drought Index (TVDI) and multiple environmental factors include: TVDI, surface temperature and its derived features, Normalized Difference Vegetation Index (NDVI), Digital Elevation Model (DEM) and its derived topographic factors, meteorological parameters, and soil property parameters. The data of all features are resampled to a uniform low resolution to obtain the input feature set. The derived features include diurnal temperature range or temperature spatiotemporal gradient; the derived topographic factors include at least one of slope, aspect, and topographic humidity index; the meteorological parameters include precipitation, air temperature, solar radiation, and wind speed; and the soil property parameters include soil texture or field water holding capacity. The two-stage feature selection method is as follows: a recursive feature elimination method is used to perform initial feature screening on the features in the input feature set; the average SHAP value of the candidate features selected in the initial screening is calculated using the SHAP global interpretation method, and candidate features with an average SHAP value lower than a preset threshold are removed, while the core features are retained to form the final feature subset.

6. The agricultural drought monitoring method based on spatiotemporal adaptive filling and XGBoost downscaling according to claim 5, characterized in that, The training method of the XGBoost model is as follows: the long-term spatiotemporal continuous soil moisture dataset generated in step one is used as the target variable, and the final feature subset is used as the input feature to form the training dataset; the training dataset is randomly divided into training set and validation set according to the proportion; the XGBoost model is trained on the training set based on the optimal hyperparameters obtained by Optuna search; and the early stopping mechanism is monitored on the validation set to determine the optimal number of trees. After the training is completed, a trained XGBoost model that characterizes the mapping relationship between input features and target variables is obtained; The input feature set is upscaled to high resolution and input into the trained XGBoost model, the soil moisture value is predicted pixel by pixel, and a soil moisture dataset with high spatial resolution, spatiotemporal seamless coverage is output.

7. The agricultural drought monitoring method based on spatiotemporal adaptive filling and XGBoost downscaling according to claim 6, characterized in that, The recursive feature elimination method recursively constructs an XGBoost base model and removes the least important features, and the implementation method is as follows: taking XGBoost as the base model, using 5-fold cross-validation to evaluate the average accuracy under different feature quantities, determining the corresponding retained feature quantity when the model accuracy is optimal, and obtaining a candidate feature subset; The XGBoost model approximates the true value by integrating K classification and regression trees, for the th The predicted soil moisture value for each sample is: In the formula, It is an XGBoost model; H is the total number of regression trees; It is the first A regression tree; For the first The input feature vector of each sample; F is the function space of all possible classification and regression trees; The loss function of the XGBoost model is the h-th tree regression. regularization term for: ； In the formula, Let be the total number of leaf nodes in the h-th regression tree. It is the weight of the o-th leaf node; , and These are the hyperparameters used to adjust the penalty strength, L2 regularization strength, and L1 regularization strength coefficient, respectively; The hyperparameters of the XGBoost model include tree structure parameters, random sampling parameters, regularization parameters and learning rate parameters; the Optuna framework combined with the Tree-structured Parzen Estimator is used for automated Bayesian hyperparameter optimization.

8. The agricultural drought monitoring method based on spatiotemporal adaptive filling and XGBoost downscaling according to claim 6 or 7, characterized in that, The method for calculating the standardized soil moisture index is as follows: extract the soil moisture time series SM of each pixel in the high-spatial-resolution spatiotemporal seamless soil moisture dataset, and perform probability distribution fitting on the soil moisture time series SM through five theoretical probability distribution functions; use the K-S test to perform hypothesis testing on the goodness of fit of each probability distribution, and determine the description ability of each probability distribution for the temporal characteristics of pixel soil moisture by calculating the maximum absolute difference between the empirical cumulative frequency and the theoretical cumulative frequency of soil moisture; For each pixel, the probability distribution with the highest goodness of fit in the KS test and passing the significance level is selected as the optimal probability distribution model for that pixel. Let the cumulative probability distribution of soil moisture values ​​corresponding to the optimal probability distribution model for pixel S be denoted as... ; The standardized soil moisture index is calculated as follows: In the formula, It is the inverse function of the cumulative probability function of the standard normal distribution; The drought grade division method is as follows: standardized soil moisture index SSI>0 indicates no drought; -1.0<SSI≤0 indicates light drought; -1.5<SSI≤-1.0 indicates moderate drought; -2.0<SSI≤-1.5 indicates severe drought; standardized soil moisture index SSI≤-2.0 indicates extreme drought.

9. The agricultural drought monitoring method based on spatiotemporal adaptive filling and XGBoost downscaling according to claim 8, characterized in that, The ranges of the hyperparameters of the XGBoost model include: the maximum depth of the tree ranges from 3 to 7; the minimum sum of sample weights of child nodes ranges from 20 to 120; the minimum loss drop threshold required for leaf node splitting ranges from 0.0 to 10.0; the sampling ratio of samples during training of each tree ranges from 0.6 to 1.0; the sampling ratio of features during training of each tree ranges from 0.6 to 1.0; the L1 regularization coefficient ranges from 0.0 to 5.0; the L2 regularization coefficient on the logarithmic scale ranges from 3.0 to 50.0, and the learning rate on the logarithmic scale ranges from 0.02 to 0.10; the early stopping mechanism automatically terminates the training when the loss function on the verification set does not decrease for 50 consecutive rounds; The five theoretical probability distribution functions include Gamma distribution, Normal distribution, Lognormal distribution, Generalized Extreme Value distribution and Pearson-III distribution; Pixel-by-pixel statistical verification showed that the best-fit distribution of pixels in the study area is a normal distribution, and the standardized soil moisture index is: In the formula, μ and σ are the mean and standard deviation of the soil moisture time series SM, respectively.

10. An agricultural drought monitoring system based on spatiotemporal adaptive infilling and XGBoost downscaling, utilizing the agricultural drought monitoring method based on spatiotemporal adaptive infilling and XGBoost downscaling as described in any one of claims 1-9, characterized in that, comprises: a data preprocessing module, configured to acquire multi-source passive microwave soil moisture data, remote sensing data, meteorological data, and topographic and soil data, and complete spatiotemporal matching and preprocessing of the data; The missing value reconstruction module, connected to the data preprocessing module, is used to perform temporal reconstruction of improved Savitzky-Golay filtering, spatial reconstruction of the geographic weighted regression model that takes into account environmental similarity, and least squares spatiotemporal coupling to complete the missing value reconstruction of soil moisture data; and to perform pixel-by-pixel consistency correction and temporal stitching of multi-source passive microwave data to generate a long-term spatiotemporal continuous soil moisture dataset. The downscaling module, connected to the missing value reconstruction module, is used to construct the input feature set, perform two-stage feature selection, and train the Optuna-optimized XGBoost model to complete the downscaling of long-term spatiotemporal continuous soil moisture datasets and generate high spatial resolution spatiotemporal seamless soil moisture products. The drought identification module, connected to the downscaling module, is used to construct long-term soil moisture time series, calculate standardized soil moisture index, and complete drought level classification and drought event identification.

Images