High-resolution surface soil moisture estimation method fused with space-ground multivariate data

Abstract

The invention relates to the technical field of soil moisture monitoring, in particular to a high-resolution surface soil moisture estimation method fusing air-ground multivariate data, which comprises the following steps: acquiring unmanned aerial vehicle visible light image data, thermal infrared image data and ground auxiliary data of a research area; based on the unmanned aerial vehicle visible light image data, the thermal infrared image data and the ground auxiliary data, environment variable characteristics of temperature information, vegetation indexes and texture information are calculated, key environment variables are selected through a characteristic screening method, and a soil moisture inversion model is constructed; on the basis of the soil moisture inversion model, a soil moisture distribution diagram of a research area is generated, a set of environment variable extraction method based on unmanned aerial vehicle visible light and thermal infrared image and ground auxiliary data fusion is constructed, and the feature space of soil moisture inversion is enriched; a feature screening strategy combining a Boruta algorithm and Pearson's correlation coefficient analysis is adopted, an optimal feature subset is selected for different soil depths, and the model efficiency and precision are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of soil moisture monitoring, and particularly to a high-resolution surface soil moisture estimation method that integrates multi-source data from air and ground, and is particularly suitable for accurately retrieving the soil moisture of vegetation communities in semi-arid regions. Background Art

[0002] Soil water content plays an important role in agricultural, meteorological, climatic, and hydrological applications. At appropriate spatial and temporal scales, accurate soil moisture information contributes to efficient irrigation decision-making and ecological environment monitoring. Traditional soil moisture measurement methods, such as the oven-drying method, time domain reflectometry, neutron measurement method, heat pulse measurement method, and cosmic ray soil moisture sensor method, etc., are limited by disadvantages such as small detection scale, high cost, time-consuming and laborious, and it is difficult to meet the requirements of large-scale and high-precision soil moisture monitoring.

[0003] In recent years, the application of unmanned aerial vehicle (UAV) remote sensing technology in large-scale, high spatio-temporal resolution crop growth monitoring and non-destructive analysis of soil moisture has been increasing. Compared with traditional manual measurement methods, UAV remote sensing technology provides a more efficient method for estimating various vegetation parameters and soil properties. Through regression analysis based on machine learning algorithms using the feature information extracted from UAV images and the target variables of ground measured points, it has been widely used in the estimation of vegetation parameters and soil properties.

[0004] However, the existing technologies still have the following problems: (1) Most studies focus on homogeneous underlying surfaces or field experiments with artificial controlled irrigation, lacking application verification under complex natural vegetation conditions; (2) The synergistic effects of multi-source data such as temperature information, vegetation index, and texture information have not been fully explored; (3) There is insufficient research on feature screening and model optimization for different soil depths; (4) There is a lack of high spatial resolution soil moisture inversion methods. Therefore, developing a high-resolution surface soil moisture estimation method that can integrate multi-source data from air and ground is of great significance for accurately monitoring soil moisture in complex vegetation environments in semi-arid regions. Summary of the Invention

[0005] The purpose of the present invention is to provide a high-resolution surface soil moisture estimation method that integrates multi-source data from air and ground to solve the problems existing in the prior art. The present invention constructs a comprehensive environmental variable feature system of temperature information, vegetation index, and texture information by integrating UAV visible light and thermal infrared images with ground auxiliary data, selects key variables using a feature screening method, and realizes high-precision soil moisture inversion based on multiple machine learning models, and can provide high-resolution soil moisture spatial distribution information under complex vegetation environments.

[0006] The present invention proposes a high-resolution surface soil moisture estimation method that integrates multi-source data from air and ground, including: Acquisition step: acquire the visible light image data, thermal infrared image data of the study area and ground auxiliary data; Processing step: based on the visible light image data, thermal infrared image data of the unmanned aerial vehicle and ground auxiliary data of the study area, calculate the environmental variable characteristics of temperature information, vegetation index and texture information, select key environmental variables through a feature screening method, and construct a soil moisture inversion model; Output step: generate a soil moisture distribution map of the study area based on the soil moisture inversion model.

[0007] Preferably, the step of acquiring the visible light image data, thermal infrared image data of the unmanned aerial vehicle and ground auxiliary data of the study area specifically includes: Set the flight altitude, heading overlap rate and side overlap rate of the unmanned aerial vehicle, and collect the visible light image data and thermal infrared image data of the unmanned aerial vehicle; Acquire ground auxiliary data, and the ground auxiliary data includes leaf area index, soil volume water content, soil temperature and canopy air temperature data.

[0008] Preferably, the step of calculating the environmental variable characteristics of temperature information, vegetation index and texture information specifically includes: Calculate the vegetation coverage based on the visible light image of the unmanned aerial vehicle; Calculate the normalized canopy temperature index based on the thermal infrared image; Calculate the vegetation index and texture information based on the visible light image of the unmanned aerial vehicle; Calculate the leaf area index based on the ground auxiliary data and unmanned aerial vehicle data.

[0009] Preferably, the calculation of the vegetation index specifically includes: Calculate the normalized band values based on the RGB band values of the visible light image of the unmanned aerial vehicle; Calculate vegetation indices including vegetation factor index, Woebbecke index, Kawashima index, visible light atmospheric impedance index, normalized green-red difference index, improved green-red vegetation index, green leaf index, red-green-blue vegetation index, over-red index, over-green index and over-blue index based on the normalized band values.

[0010] Preferably, the calculation of the texture information specifically includes: Calculate the texture information based on the gray-level co-occurrence matrix including mean value, variance, co-occurrence, contrast, dissimilarity, entropy, second moment and correlation based on the visible light image of the unmanned aerial vehicle.

[0011] Preferably, the feature screening method includes the Boruta algorithm and the Pearson correlation coefficient analysis method, and key environmental variables at different soil depths are screened out through the feature screening method.

[0012] Preferably, the steps of constructing the soil moisture inversion model specifically include: Taking the key environmental variables as independent variable data; Dividing the training set data into model training data and validation data; Based on the model training data, constructing a soil moisture inversion model including a random forest regression model, a gradient boosting decision tree model, an extreme gradient boosting decision tree model, and a classification boosting decision tree model; Based on the validation data, evaluating the accuracy of the soil moisture inversion model and selecting the optimal soil moisture inversion model.

[0013] Preferably, the steps of evaluating the accuracy of the soil moisture inversion model specifically include: Calculating the coefficient of determination, root mean square error, and mean absolute error; Based on the coefficient of determination, root mean square error, and mean absolute error, determining the optimal soil moisture inversion model at different soil depths.

[0014] Preferably, the soil moisture distribution map includes soil moisture distribution maps at four depths of 3.8 cm, 7.5 cm, 12 cm, and 20 cm.

[0015] Preferably, the optimal soil moisture inversion model includes: A model for soil moisture inversion at a depth of 3.8 cm using the feature subset screened by CatBoost combined with the Boruta algorithm; A model for soil moisture inversion at a depth of 7.5 cm using the feature subset screened by CatBoost combined with the Boruta algorithm; A model for soil moisture inversion at a depth of 12 cm using the feature subset screened by CatBoost combined with the Pearson algorithm; A model for soil moisture inversion at a depth of 20 cm using the feature subset screened by GBDT combined with the Pearson algorithm.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) Constructed a set of environmental variable extraction methods based on the fusion of UAV visible light and thermal infrared images and ground auxiliary data, enriching the feature space of soil moisture inversion; (2) Adopted a feature screening strategy combining the Boruta algorithm and the Pearson correlation coefficient analysis, selecting the optimal feature subset for different soil depths, and improving the model efficiency and accuracy; (3) The performance of four machine learning models, namely random forest regression, gradient boosting decision tree, extreme gradient boosting decision tree, and classification boosting decision tree, in soil moisture inversion at different depths was compared, providing a basis for the optimal model selection for soil moisture inversion at different depths; (4) The generation of soil moisture distribution maps with a spatial resolution of 1 m was achieved, providing high-resolution soil moisture spatial information for precision agriculture and ecological environment monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flowchart of the high-resolution surface soil moisture estimation method that integrates multi-source data from air and ground of the present invention; Figure 2 It is a schematic diagram of the experimental design, including the UAV platform, sensor equipment, ground plot layout, and measurement methods, (a: DJI M300rtk UAV equipped with WIRS pro, b: WIRS pro thermal infrared imager, c: location of the ground test points, d: ground black and white board, e: plot layout, f: soil moisture measurement, g: plot LAI measurement); Figure 3 It is a flowchart for extracting environmental variable features; Figure 4 It is a schematic diagram of the screening results of the feature screening method; Figure 5 It is a comparison chart of the accuracy of soil moisture models at different depths; Figure 6 It is an analysis chart of the contribution degree of model features; Figure 7 It is a soil moisture distribution map of the study area for six days. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] Please refer to the attached Figure 1-7 , and the present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0019] Embodiment 1 As Figure 1 shown, the high-resolution surface soil moisture estimation method that integrates multi-source data from air and ground provided by the present invention includes an acquisition step, a processing step, and an output step.

[0020] In the acquisition step, the UAV visible light image data, thermal infrared image data, and ground auxiliary data of the study area are acquired. Specifically, as Figure 2 shown, the study area is located in the agro-pastoral ecotone of Hohhot City, Inner Mongolia Autonomous Region, which belongs to a typical semi-arid climate region. The study area contains various vegetation types, such as corn, Leymus chinensis, Elymus sibiricus, Elymus nutans, Lespedeza daurica, and Medicago sativa.

[0021] Preferably, the specific method for obtaining the visible light image data and thermal infrared image data of the study area by drone is as follows: Use the DJI M300 RTK drone platform, equipped with the WIRIS Pro high-performance airborne thermal infrared imager. Set the flight altitude of the drone to 100 m, the forward overlap rate to 75%, and the side overlap rate to 70%. Fly and collect data under clear and windless weather conditions. The present invention selects to collect data from 11:30 to 12:30 Beijing time to ensure relatively stable solar radiation and reduce the influence of shadows on the image quality.

[0022] In an embodiment of the present invention, the acquisition of ground auxiliary data includes: setting up ground quadrats in different vegetation coverage areas within the study area, with each quadrat sized 1 m × 1 m; using a TDR350 soil moisture meter to measure the volumetric soil water content at four depths of 3.8 cm, 7.5 cm, 12 cm, and 20 cm within the quadrat; using a LAI-2200C plant canopy analyzer to measure the leaf area index of the quadrat; recording the soil temperature of the quadrat and the air temperature 1.5 m above the canopy. At the same time, black and white boards are arranged around the study area for temperature calibration of thermal infrared images.

[0023] In the processing step, based on the acquired visible light image data, thermal infrared image data, and ground auxiliary data of the drone, environmental variable characteristics of temperature information, vegetation index, and texture information are calculated. Key environmental variables are selected through a feature screening method, and a soil moisture inversion model is constructed.

[0024] As Figure 3 shown, the calculation of environmental variable characteristics mainly includes four aspects: vegetation coverage calculation, normalized canopy temperature index calculation, vegetation index and texture information calculation, and leaf area index calculation.

[0025] In the calculation of vegetation coverage (FVC), the present invention uses the random forest algorithm in eCognition 9.0 software to classify the processed visible light images into vegetation and non-vegetation, and exports the classification vector. The classification vector is exported as a raster and reclassified into vegetation and bare soil, and then resampled into a raster with a size of 1 m × 1 m. The value of each raster is the proportion of vegetation. The calculation formula for vegetation coverage is: , where refers to the number of vegetation pixels within a range of 1 m × 1 m, refers to the total number of pixels, and the value range of FVC is between 0 and 1.

[0026] In the calculation of the Normalized Canopy Temperature Index (NRCT), the present invention extracts the calibrated temperature image using the derived vegetation vector mask to obtain the vegetation temperature raster data. To avoid the influence of mixed pixels, the first and last 2% of the pixels in the vegetation temperature image of each day are removed respectively as the vegetation canopy temperature of that day. Then, the Normalized Canopy Temperature Index is calculated through the band operation tool in ENVI in combination with the NRCT formula: , where, is the canopy temperature pixel, is the maximum value of all canopy temperature pixels in the six-day image, is the minimum value of all canopy temperature pixels. Finally, the result is resampled to a size of 1m×1m, and the resampling process is achieved by calculating the mean through the aggregation tool of Arcgis10.8.1, where the pixel coefficient is set to 20 and the method is set to MEAN.

[0027] In the calculation of the vegetation index, the present invention calculates the normalized band values using the R, G, and B bands of the visible light image: r = R / (R + G + B), g = G / (R + G + B), b = B / (R + G + B), where R, G, and B are the DN values of the red, green, and blue bands in each quadrat area of the UAV image respectively. Based on the normalized band values, 11 vegetation indices are calculated, including the Vegetation Factor Index (VEG), Woebbecke Index (WI), Kawashima Index (IKAW), Visible Atmospherically Resistant Index (VARI), Normalized Green-Red Difference Index (NGRDI), Modified Green-Red Vegetation Index (MGRVI), Green Leaf Index (GLI), Red-Green-Blue Vegetation Index (RGBVI), Excess Red Index (EXR), Excess Green Index (EXG), and Excess Blue Index (EXB).

[0028] In the calculation of the texture information, the present invention calculates 8 kinds of texture information based on the Gray Level Co-occurrence Matrix (GLCM) for the visible light image using ENVI / IDL software, including Mean, Variance, Homogeneity, Contrast, Dissimilarity, Entropy, Second Moment, and Correlation. The smallest window size (3×3 pixels) is used in the calculation of these textures.

[0029] In the calculation of leaf area index (LAI), the present invention uses the extracted texture information and color index as feature variables, and inversely analyzes the LAI distribution in the study area through the random forest algorithm. First, five variables most relevant to LAI (RMea, BMea, NGRDI, MGRVI, VARI) are selected, then a random forest model is constructed based on these variables and the measured LAI data, and finally the model is applied to the entire study area to obtain the LAI distribution map.

[0030] In the process of selecting key environmental variables by the feature screening method, the present invention adopts two complementary methods: the Boruta algorithm and the Pearson correlation coefficient analysis. The Boruta algorithm is a feature selection method based on random forest, which can effectively identify features that have a significant impact on the target variable; the Pearson correlation coefficient analysis is used to evaluate the linear relationship between two variables. Through these two methods, the present invention screens out key environmental variables for four soil depths of 3.8 cm, 7.5 cm, 12 cm and 20 cm respectively, forming eight groups of feature subsets, such as Figure 4 shown.

[0031] For example, for the soil moisture at a depth of 3.8 cm, the feature subsets screened by the Pearson correlation coefficient method include R(Mea), G(Mea), B(Mea), NRCT, LAI, FVC, Ta, Tc, Ts; while the feature subsets screened by the Boruta algorithm include G(Mea), B(Ent), WI, IKAW, NRCT, LAI, Ta, Tc, Ts. This differential feature screening strategy for different depths effectively improves the accuracy and efficiency of the model.

[0032] In the process of constructing the soil moisture inversion model, the present invention selects four machine learning algorithms based on tree models: random forest regression (RFR), gradient boosting decision tree (GBDT), extreme gradient boosting decision tree (XGBoost) and classification boosting decision tree (CatBoost). The eight groups of feature subsets obtained by screening are used as the independent variable data of the four models at different depths. Among 162 soil moisture datasets, the training set and the test set are divided in a ratio of 7:3, and the five-fold cross-validation and grid search methods are used to optimize the model parameters.

[0033] In the output step, based on the constructed soil moisture inversion model, a soil moisture distribution map of the study area is generated. By evaluating the performance of the four models under different feature subsets, the present invention determines the optimal model combination for soil moisture inversion at different depths and applies it to the entire study area to obtain a soil moisture distribution map with a resolution of 1 meter.

[0034] Example 2

[0035] This embodiment relates to the specific implementation of obtaining visible light image data, thermal infrared image data and ground auxiliary data of an unmanned aerial vehicle (UAV).

[0036] In a preferred embodiment of the present invention, a DJI M300 RTK UAV platform is used for UAV data collection. This platform is equipped with a WIRIS Pro high-performance airborne thermal infrared imager, which can collect visible light and thermal infrared data simultaneously. The flight altitude is set at 100 m, which can cover a sufficiently large research area while ensuring the image resolution (about 0.05 m). The forward overlap rate is set at 75%, and the side overlap rate is set at 70%. Such overlap rate settings can ensure the accuracy and integrity of image mosaicking.

[0037] The collection of ground auxiliary data mainly includes the following aspects: (1) Quadrat data collection: 27 quadrats of 1 m×1 m are set in the research area, covering different vegetation type areas. The data collected for each quadrat include leaf area index, soil volumetric water content, soil temperature and air temperature above the canopy.

[0038] (2) Leaf area index (LAI) collection: The LAI of the quadrat is measured using a LAI-2200C plant canopy analyzer. The collection method is in the form of ABBBB, that is, measuring 4 times (B) below the quadrat and 1 time (A) above the quadrat to obtain one set of LAI data. Each quadrat is measured 4 times repeatedly, and the mean value is taken as the measured LAI value of this quadrat.

[0039] (3) Soil moisture collection: The soil volumetric water content in the quadrat is measured using a TDR350 soil moisture meter. The five-point method is adopted, that is, measuring once at each of the four corners and the center of each quadrat, and the average value of the five measurements is taken as the measured soil moisture value. The measurement depths are 3.8 cm, 7.5 cm, 12 cm and 20 cm respectively. At the same time, the oven-drying method is used to calibrate the TDR data.

[0040] (4) Temperature data collection: The temperature of the black and white boards is measured using a Fluke 62 max+ infrared thermometer for thermal infrared image calibration; the air temperature 1.5 m above the quadrat canopy is measured using a mercury thermometer.

[0041] Through this detailed ground auxiliary data collection scheme, the present invention realizes the effective integration of UAV remote sensing data and ground measured data, providing reliable basic data for subsequent soil moisture inversion.

[0042] Embodiment 3 This embodiment relates to the specific implementation of calculating environmental variable characteristics of temperature information, vegetation index and texture information.

[0043] In the calculation of fractional vegetation cover (FVC), the present invention adopts a classification method based on the random forest algorithm, which classifies visible light images into two categories: vegetation and non-vegetation. Compared with the traditional threshold-based classification method, the random forest algorithm can better handle the vegetation classification problem under complex backgrounds and improve the classification accuracy. After classification, it is resampled to a resolution of 1m×1m, which can not only reflect the spatial variability of vegetation but also meet the actual application requirements.

[0044] In the calculation of the normalized canopy temperature index (NRCT), the present invention introduces a strategy of excluding the top and bottom 2% of pixels to eliminate the influence of extreme temperature values on the index calculation. The NRCT index can effectively reflect the water stress status of vegetation and is an important indicator for soil moisture inversion.

[0045] In the calculation of vegetation indices, the present invention calculates 11 visible light-based vegetation indices. These indices reflect the growth status and canopy characteristics of vegetation from different perspectives and provide rich characteristic information for soil moisture inversion. For example, the normalized green-red difference index (NGRDI) is sensitive to the photosynthetic activity of vegetation, while the visible atmosphere resistance index (VARI) has a certain resistance to the influence of the atmosphere.

[0046] In the calculation of texture information, the present invention selects the gray-level co-occurrence matrix (GLCM) as the basis for texture feature extraction and calculates 8 texture features. These texture features can describe the spatial structure characteristics of the vegetation canopy and indirectly reflect the soil moisture status. The window size is selected as 3×3 pixels to capture local texture changes while avoiding boundary effects caused by too large windows.

[0047] The calculation of leaf area index (LAI) adopts an inversion method based on random forest. By analyzing the correlation between each feature and LAI, the five variables with the highest correlation are selected for model construction. This method comprehensively considers spectral and texture information and can more accurately invert the spatial distribution of LAI.

[0048] Through the calculation of the above environmental variable features, the present invention constructs a complete feature system, providing multi-angle and multi-level information support for soil moisture inversion.

[0049] Example 4 This example relates to the specific implementation manner of selecting key environmental variables by the feature screening method.

[0050] The present invention adopts two complementary feature screening methods: the Boruta algorithm and Pearson correlation coefficient analysis. The Boruta algorithm is a feature selection method based on random forest. By introducing shadow features and performing multiple iterations, it can robustly identify significant features. This algorithm performs excellently in dealing with high-dimensional data, effectively reducing the model complexity and improving the prediction performance. Pearson correlation coefficient analysis is a classical linear correlation evaluation method, which can intuitively reflect the strength of the linear relationship between features and target variables.

[0051] For different soil depths, the present invention respectively applies these two methods for feature screening and obtains eight groups of feature subsets. Taking the depth of 3.8 cm as an example, the features screened by the Pearson correlation coefficient method include R(Mea), G(Mea), B(Mea), NRCT, LAI, FVC, Ta, Tc, Ts, mainly including texture mean, temperature information and vegetation structure parameters; while the features screened by the Boruta algorithm include G(Mea), B(Ent), WI, IKAW, NRCT, LAI, Ta, Tc, Ts. In addition to texture and temperature information, it also includes more vegetation indices.

[0052] At different depths, the key variables screened are also different. As the soil depth increases, the influence of surface environmental variables on soil moisture gradually weakens, and deeper soil moisture may be related to fewer environmental variables. For example, at a depth of 20 cm, the features screened by the Pearson correlation coefficient method are reduced to G(Mea), B(Mea), NRCT, LAI, FVC, Ta, Tc, while the features screened by the Boruta algorithm are NRCT, Tc, Ts, EXB, mainly temperature-related variables.

[0053] This differential feature screening strategy for different depths can effectively capture the relationship between soil moisture and environmental variables at different depths, improving the pertinence and accuracy of the model.

[0054] Example 5 This example relates to the specific implementation of constructing a soil moisture inversion model.

[0055] The present invention selects four machine learning algorithms based on tree models: Random Forest Regression (RFR), Gradient Boosting Decision Tree (GBDT), Extreme Gradient Boosting Decision Tree (XGBoost) and Classification Boosting Decision Tree (CatBoost). These algorithms all belong to the ensemble learning method, which improves the overall performance by combining multiple base learners and has strong non-linear fitting ability and anti-noise ability.

[0056] During the model training process, the present invention divides 162 soil moisture data into a training set and a test set at a ratio of 7:3. This is a commonly used division ratio in the field of machine learning, which can not only ensure sufficient data volume in the training set but also provide enough test samples. The training set adopts the five-fold cross-validation method, dividing the data into five parts. Each time, four parts are used for training and one part is used for validation. This process is repeated five times, and the average performance is taken as the model evaluation index. This method can make full use of limited data and improve the reliability of model evaluation.

[0057] To optimize the model parameters, the present invention adopts the grid search method (GridSearchCV), sets the search range for the key parameters of each model, and aims at the minimum root mean square error (RMSE) to determine the best parameter combination. For example, for the random forest model, the main parameters to be tuned include the number of decision trees (n_estimators), the maximum number of features (max_features), the minimum number of samples in a leaf node (min_samples_leaf), etc.; for the GBDT and XGBoost models, the parameters to be tuned include the learning rate (learning_rate), the maximum depth (max_depth), the subsampling ratio (subsample), etc.

[0058] By combining eight groups of feature subsets and four machine learning models, the present invention evaluates the performance of 32 model configurations at four soil depths. The evaluation indicators include the coefficient of determination (R²), the root mean square error (RMSE), and the mean absolute error (MAE), and their calculation formulas are as follows: , , , Among them, and are the measured value and the predicted value of soil moisture respectively, represents the average of the measurements, and n represents the number of samples.

[0059] Based on the evaluation results, the present invention determines the optimal model combinations for soil moisture inversion at different depths: for depths of 3.8 cm and 7.5 cm, the feature subset screened by CatBoost combined with the Boruta algorithm is used; for a depth of 12 cm, the feature subset screened by CatBoost combined with the Pearson correlation coefficient method is used; for a depth of 20 cm, the feature subset screened by GBDT combined with the Pearson correlation coefficient method is used. The performance of the test sets of these optimal combinations respectively reaches R² = 0.91, RMSE = 0.72%, MAE = 0.64% (3.8 cm); R² = 0.83, RMSE = 0.93%, MAE = 0.78% (7.5 cm); R² = 0.81, RMSE = 0.97%, MAE = 0.84% (12 cm); R² = 0.77, RMSE = 1.00%, MAE = 0.97% (20 cm).

[0060] Example 6 This example relates to the specific implementation manner of generating the soil moisture distribution map of the study area.

[0061] Based on the determined optimal model combinations, the present invention applies the models to the entire study area and generates soil moisture distribution maps at four depths of 3.8 cm, 7.5 cm, 12 cm, and 20 cm, with a spatial resolution of 1 meter. As Figure 7 shown, the present invention successfully inverses the spatial distribution of soil moisture on six collection dates (August 31, 2021; September 7, 2021; October 1, 2021; June 9, 2022; July 21, 2022; August 30, 2022).

[0062] The following characteristics can be observed from the inversion results: (1) The spatial distribution law of soil moisture at different depths: The soil moisture contents at depths of 3.8 cm, 7.5 cm, and 12 cm are relatively low, while the soil moisture content at a depth of 20 cm is relatively high. This is consistent with the actual situation because the surface soil is more affected by evaporation and the water loss is faster.

[0063] (2) The differences in soil moisture under different vegetation covers: The soil moisture under the vegetation covers such as corn, alfalfa, and Chinese wildrye is generally higher than that in the areas covered by old awnless brome, Elymus nutans, and Lespedeza davurica. This may be related to the different water use strategies of different vegetation. According to research, tall, small-leaved, and shallow-rooted plants on flat terrain are beneficial to soil moisture retention and replenishment.

[0064] (3) Temporal variation characteristics: There are obvious differences in soil moisture distribution on different dates, which reflects the influence of factors such as precipitation, evaporation, and vegetation growth on soil moisture. For example, the overall soil moisture on June 9, 2022 was relatively low, which may be related to the drought conditions during that period; while the soil moisture on September 7, 2021 was relatively high, probably due to abundant precipitation in the previous period.

[0065] To further verify the reliability of the model, the present invention separately evaluated the soil moisture inversion results for each acquisition date. The results showed that the inversion accuracies (R²) for the six days were 0.86 (August 31, 2021), 0.89 (September 7, 2021), 0.80 (October 1, 2021), 0.69 (June 9, 2022), 0.90 (July 21, 2022), and 0.72 (August 30, 2022). Except for June 9, 2022, the inversion accuracies on other dates were relatively high, indicating that the model has good temporal stability and adaptability.

[0066] By generating high-resolution soil moisture distribution maps, the present invention provides important spatial information support for applications such as precision agricultural irrigation decision-making and ecological environment monitoring.

[0067] Example 7 This example relates to the specific implementation methods of feature importance assessment and model interpretation.

[0068] To deeply understand the contributions of various environmental variables to soil moisture inversion, the present invention used the SHAP (Shapley Additive Explanations) analysis method to evaluate the importance of features. SHAP is a model interpretation method based on game theory that can fairly quantify the contribution of each feature to the model prediction and provide explanations at both the global and local levels.

[0069] The analysis results showed that the feature with the highest average importance contribution was NRCT (24.7%), followed by Tc (21.8%), Ts (20.4%), and Ta (11.8%). The sum of the average contributions of the four temperature-related features was as high as 78.7%, indicating that temperature information is the dominant factor in soil moisture inversion. This finding is consistent with the physical mechanism because the temperature states of the soil and vegetation are closely related to their moisture conditions: soils and vegetation with sufficient moisture will exhibit lower temperatures due to evaporation and transpiration, while those under water stress conditions will have higher temperatures.

[0070] The contribution degrees of the vegetation structure parameters LAI (11.8%) and FVC (6.9%) were relatively low but cannot be ignored, and the two together contributed 18.7% of the information. These parameters reflect the growth status and coverage degree of the vegetation and indirectly affect the evaporation and distribution of soil moisture.

[0071] Vegetation indices (such as WI, 4.7%) and texture information contribute the least in SHAP analysis. This may be because there is a certain information redundancy between these features and temperature and vegetation structure parameters, or the relationship between them and soil moisture is relatively complex and not easily captured directly by the model.

[0072] In the inversion of soil moisture at different depths, the importance of key variables also varies. As the depth increases, the influence of surface environmental variables (especially texture information and vegetation indices) gradually weakens, while the relative importance of temperature-related variables increases. This also explains why different depth requires different feature subsets and model combinations to obtain the best inversion effect.

[0073] Example 8 This example involves the evaluation of model performance under different vegetation types and environmental conditions.

[0074] To comprehensively evaluate the adaptability of the method of the present invention, the inversion accuracy of soil moisture under different vegetation type covers was analyzed. The results show that in relatively uniform vegetation cover areas such as corn and Chinese wildrye, the model performance is slightly better than that in non-uniformly distributed vegetation areas such as old awn grass and Elymus nutans. This may be because the radiation transfer and heat exchange processes are more stable under uniform vegetation, and the relationship between environmental variables and soil moisture is more consistent.

[0075] In addition, the method of the present invention also shows good stability under different weather conditions. For example, the inversion accuracies on July 21, 2022 (air temperature 33.1°C, higher temperature condition) and August 30, 2022 (air temperature 20.9°C, lower temperature condition) both reached an acceptable level (R² was 0.90 and 0.72 respectively). This indicates that the model constructed by the present invention has a certain adaptability to temperature changes.

[0076] It should be noted that under extremely arid conditions (such as on June 9, 2022, when soil moisture was generally low), the model performance was relatively weak (R² = 0.69). This may be because drought stress exacerbates the water regulation mechanism of vegetation, making the relationship between environmental variables and soil moisture more complex. Future research can consider adding specific environmental variables or modifying the model structure for drought conditions to further improve the inversion accuracy.

[0077] Generally speaking, the method of the present invention can achieve high-precision inversion of soil moisture under various vegetation type covers in semi-arid regions, showing good adaptability and robustness.

[0078] Example 9 This example involves the actual application scenarios and value of the method of the present invention.

[0079] The method of the present invention can be widely applied to the following scenarios: (1) Precision agriculture: Based on the high-resolution soil moisture distribution map, farmers can implement variable-rate irrigation, adjust irrigation strategies according to the actual moisture requirements of different regions, improve water resource utilization efficiency, and reduce water resource waste. For example, in the corn planting area, precise irrigation plans can be formulated according to the soil moisture conditions at different growth stages.

[0080] (2) Ecological environment monitoring: In semi-arid regions, soil moisture is a key factor affecting vegetation growth and ecosystem functions. The method of the present invention can be used to monitor the dynamic changes of soil moisture over a large range, evaluate the impact of drought on the ecosystem, and provide decision-making support for ecological protection and restoration.

[0081] (3) Hydrological model parameterization: The high-resolution soil moisture distribution map can be used as input or verification data for hydrological models, improving the accuracy and reliability of the models. This is of great significance for water resource management, flood prediction, etc.

[0082] (4) Climate change research: Long-term soil moisture monitoring data can be used to analyze the impact of climate change on the regional water cycle and evaluate the effectiveness of adaptation and mitigation measures.

[0083] Compared with traditional soil moisture monitoring methods, the present invention has the following advantages: (1) Non-destructive monitoring: It is possible to conduct long-term continuous monitoring without destructive sampling; (2) High resolution: It provides soil moisture distribution information with a spatial resolution of 1 meter, far superior to satellite remote sensing products; (3) Multi-depth monitoring: It provides soil moisture information at four depths of 3.8 cm, 7.5 cm, 12 cm, and 20 cm simultaneously, forming a three-dimensional monitoring system; (4) Adaptability to the natural environment: High-precision inversion has been achieved under complex natural vegetation conditions, breaking through the limitations of previous studies confined to homogeneous underlying surfaces or artificially controlled environments.

[0084] In summary, the present invention provides a high-resolution surface soil moisture estimation method that integrates multi-source data of air and ground. By comprehensively using visible and thermal infrared images of unmanned aerial vehicles and ground auxiliary data, combined with feature screening strategies and multi-model comparisons, high-precision soil moisture inversion in complex vegetation environments in semi-arid regions has been achieved. This method provides important technical support for fields such as precision agriculture and ecological environment monitoring.

[0085] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A high-resolution surface soil moisture estimation method integrating air-ground multivariate data, characterized in that: include: Acquisition step: obtaining UAV visible light image data, thermal infrared image data and ground auxiliary data of the study area; The processing step is to calculate the environmental variable characteristics of temperature information, vegetation index and texture information based on the visible light image data, thermal infrared image data and ground auxiliary data of the UAV, select key environmental variables through a feature screening method, and construct a soil moisture inversion model; The output step generates a soil moisture distribution map of the study area based on the soil moisture inversion model.

2. The method according to claim 1, characterized in that The steps of obtaining the UAV visible light image data, thermal infrared image data and ground auxiliary data of the study area specifically include: Set the UAV flight altitude, heading overlap rate and lateral overlap rate, and collect the UAV visible light image data and thermal infrared image data; Acquire ground auxiliary data, which include leaf area index, soil volume moisture content, soil temperature and air temperature above the canopy.

3. The method according to claim 2, characterized in that The step of calculating the environmental variable characteristics of temperature information, vegetation index and texture information specifically includes: Calculating vegetation coverage based on the visible light image of the drone; Calculating a normalized canopy temperature index based on the thermal infrared image; Calculating vegetation index and texture information based on the visible light image of the UAV; The leaf area index is calculated based on the ground auxiliary data and the UAV data.

4. The method according to claim 3, characterized in that The calculation of the vegetation index specifically includes: Based on the RGB band values ​​of the visible light image of the drone, a normalized band value is calculated; Based on the normalized band values, vegetation indices including vegetation factor index, Woebbecke index, Kawashima index, visible light atmospheric impedance index, normalized green-red difference index, improved green-red vegetation index, green leaf index, red-green-blue vegetation index, over-red index, over-green index and over-blue index are calculated.

5. The method according to claim 3, characterized in that: The computing texture information specifically includes: Based on the visible light image of the UAV, texture information based on the gray level co-occurrence matrix including mean value, variance, synergy, contrast, dissimilarity, information entropy, second-order moment and correlation is calculated.

6. The method according to claim 1, characterized in that The feature screening method includes a Boruta algorithm and a Pearson correlation coefficient analysis method, and the key environmental variables at different soil depths are screened out respectively by the feature screening method.

7. The method according to claim 1, characterized in that The steps of constructing the soil moisture inversion model specifically include: Using the key environmental variables as independent variable data; Divide the training set data into model training data and validation data; Based on the model training data, a soil moisture inversion model including a random forest regression model, a gradient boosting decision tree model, an extreme gradient boosting decision tree model and a classification boosting decision tree model is constructed; Based on the validation data, the accuracy of the soil moisture inversion model is evaluated, and an optimal soil moisture inversion model is selected.

8. The method according to claim 7, characterized in that The step of evaluating the accuracy of the soil moisture inversion model specifically includes: Calculate the coefficient of determination, root mean square error, and mean absolute error; Based on the determination coefficient, root mean square error and mean absolute error, the optimal soil moisture inversion model at different soil depths is determined.

9. The method according to claim 1, characterized in that: The soil moisture distribution map includes soil moisture distribution maps at four depths of 3.8 cm, 7.5 cm, 12 cm and 20 cm.

10. The method according to claim 7, characterized in that The optimal soil moisture inversion model includes: The soil moisture inversion at 3.8 cm depth uses a model based on a feature subset selected by CatBoost combined with the Boruta algorithm; The soil moisture inversion at 7.5 cm depth uses a model based on a feature subset selected by CatBoost combined with the Boruta algorithm; The soil moisture inversion at 12 cm depth uses a model based on a feature subset selected by CatBoost combined with the Pearson algorithm; The soil moisture inversion at 20 cm depth adopts the model of GBDT combined with the feature subset screened by Pearson algorithm.

Images