A soil moisture inversion method based on multi-source remote sensing data normalization processing

Abstract

The application discloses a soil humidity inversion method based on multi-source remote sensing data normalization processing and belongs to the technical field of soil data processing, and comprises the following steps: acquiring environment meteorological data of a high grass prairie area, terrain data of the high grass prairie area and measured soil humidity data of the high grass prairie area; acquiring corresponding multi-source remote sensing data of the high grass prairie area; performing space-time normalization on the multi-source remote sensing data through Z-Score standardization, a Loess curve and anti-Z-Score standardization; matching the environment meteorological data of the high grass prairie area, the terrain data of the high grass prairie area, the space-time normalized remote sensing data and the measured soil humidity data of the high grass prairie area to establish a database; and predicting soil humidity through an XGBoost model. The soil humidity inversion method solves the problem that the difference between different optical sensors in the prior art affects the consistency of data, thereby reducing the accuracy and reliability of a soil humidity inversion model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of soil data processing, and particularly relates to a soil moisture inversion method based on multi-source remote sensing data normalization processing. BACKGROUND

[0002] Soil moisture is a core parameter in hydrology and ecology, and has a double regulatory effect on carbon cycle, energy flow and water, carbon and energy exchange between land surface and atmosphere. As an important indicator and driving factor of climate change, its spatial heterogeneity is significant, which not only regulates the runoff response of the basin (such as the infiltration and runoff distribution of rainfall), but also deeply affects the soil carbon cycle, vegetation productivity, and the occurrence and development of extreme climate events (drought, flood, heat wave). In the grassland ecosystem, this parameter shows multi-dimensional value: it is not only a key indicator for evaluating biomass, monitoring drought and disaster impact, but also dominates the evolution of grazing systems, the long-term interaction between large herbivores and vegetation, and has strategic significance for agriculture and animal husbandry production, ecological protection and sustainable development. The breakthrough of high-resolution soil moisture inversion technology will directly promote the application of precise environmental monitoring, crop evaluation, flood warning, and the establishment of long-term dynamic monitoring system will provide a scientific basis for analyzing the coordinated evolution law of "soil-vegetation-animal" system under the background of climate change, and has irreplaceable decision support value for global ecological governance and resource management.

[0003] It is a classic problem in the field of remote sensing detection to understand and build soil moisture inversion models. Traditional soil moisture measurement methods, such as constant temperature weighing method, time domain reflectometer and resistance method, although have high measurement accuracy, are usually regarded as "ground true value", but the measurement range is very small, which is difficult to meet the needs of environmental monitoring and disaster prevention. In contrast, satellite remote sensing technology can provide large-scale, high-temporal and spatial resolution soil moisture data. Among them, red, green and blue (RGB) and infrared methods are essential in soil moisture monitoring, which can capture high-resolution temporal and spatial data, thereby providing detailed understanding of the dynamics of soil moisture in the study area. Usually in space, soil moisture data in high grass prairie areas are often measured in the form of single-point measurement or regional measurement, which are usually sparsely distributed in space. Therefore, long-term monitoring in time can make up for the small amount of measured sample data, and more continuous long-time soil moisture samples can be obtained. In addition, multi-source remote sensing data corresponding to high grass prairie areas can make up for the problem of low time resolution of a single sensor. However, for multi-source remote sensing data corresponding to high grass prairie areas, different optical sensors have different return periods in time, and different measurement time intervals for a certain point or region. There are obvious differences in spectral wavelength range and spectral response function curve between different optical sensors due to differences in spectral response between different optical sensors and internal differences of the sensor, as well as imaging time and environmental differences. These differences will affect the consistency of the data, thereby reducing the accuracy and reliability of the soil moisture inversion model. SUMMARY

[0004] In order to overcome the defects existing in the prior art, the present application provides a soil moisture inversion method based on multi-source remote sensing data normalization processing to solve the above problems.

[0005] The technical scheme adopted by the present application to solve its technical problems is: a soil moisture inversion method based on multi-source remote sensing data normalization processing, comprising the following steps:

[0006] S1: obtaining environmental meteorological data of high grass prairie area, terrain data of high grass prairie area and measured soil moisture data of high grass prairie area;

[0007] S2: obtaining multi-source remote sensing data corresponding to high grass prairie area;

[0008] S3: performing spatio-temporal normalization on the multi-source remote sensing data corresponding to the high grass prairie area by Z-Score standardization, Loess curve and inverse Z-Score standardization to obtain spatio-temporally normalized remote sensing data;

[0009] S4: Match the environmental meteorological data of the tallgrass prairie area, the terrain data of the tallgrass prairie area, the remote sensing data after temporal and spatial normalization, and the measured soil moisture data of the tallgrass prairie area to establish a database, and predict the soil moisture of the tallgrass prairie area through the XGBoost model.

[0010] Preferably, in step S2, remote sensing data corresponding to the tallgrass prairie area on the corresponding date is obtained according to the acquisition time of the measured soil moisture data of the tallgrass prairie area; and the vegetation index NDVI, dual-band enhanced vegetation index EVI2 and water body index NDWI of each remote sensing data are calculated.

[0011] Optionally, multi-source remote sensing data corresponding to the tallgrass prairie area include Landsat-5, Landsat-7, Landsat-8, Sentinel-2, and Planetscope multi-band optical satellite image data. After obtaining the remote sensing data, images with less than 30% cloud cover are screened and resampled to a spatial resolution of 30 meters.

[0012] Specifically, in step S3, the mean value μ and standard deviation σ of all pixel values ​​of a single image corresponding to the vegetation index NDVI, a single image corresponding to the dual-band enhanced vegetation index EVI2, and a single image corresponding to the water body index NDWI calculated from multi-source remote sensing data corresponding to the original tallgrass prairie area in the same year are calculated, and each image is standardized using Z-Score standardization;

[0013] Then, the Loess curves of the mean value μ and standard deviation σ changing with time are established respectively.

[0014] It is worth noting that in step S3, the time-space normalized average value is obtained according to the Loess curve corresponding to the average value μ of the date. , according to the Loess curve corresponding to the standard deviation σ of the date, the standard deviation after time and space normalization is obtained ; Use inverse Z-Score standardization to obtain the single image corresponding to the vegetation index NDVI after spatiotemporal normalization correction, the single image corresponding to the dual-band enhanced vegetation index EVI2, and the single image corresponding to the water body index NDWI; use the Loess-Z-Score spatiotemporal normalization correction of the single image corresponding to the original vegetation index NDVI, the single image corresponding to the dual-band enhanced vegetation index EVI2, and the single image corresponding to the water body index NDWI as the remote sensing data after spatiotemporal normalization.

[0015] Optionally, the step S32 is specifically as follows: dividing all samples consisting of the mean value μ or all samples consisting of the standard deviation σ into multiple intervals, performing polynomial fitting on the samples in the intervals, repeating this process to obtain weighted regression curves in different intervals, and finally connecting the centers of these weighted regression curves together to synthesize a complete Loess curve.

[0016] Specifically, in step S4, the database is randomly divided into a training set and a validation set in a ratio of 7:3, wherein the training set is used to train the XGBoost model; then the training set and the validation set are input into the trained XGBoost model to obtain the predicted soil moisture value of the tallgrass prairie area, and the correlation coefficient R, root mean square error RMSE, bias and unbiased root mean square error ubRMSE between the predicted soil moisture value of the tallgrass prairie area and the measured soil moisture data of the tallgrass prairie area are calculated for accuracy assessment.

[0017] Specifically, in step S4, the SHAP value is used to analyze the influence of the environmental meteorological data of the tallgrass grassland area, the terrain data of the tallgrass grassland area and the remote sensing data after time and space normalization in the trained XGBoost model on the measured soil moisture data of the tallgrass grassland area; the importance of the influence is represented by the size of the SHAP value, a positive value indicates that the feature contributes to the measured soil moisture data of the tallgrass grassland area, for positive values, the larger the SHAP value, the greater the contribution to the measured soil moisture data of the tallgrass grassland area, and the negative value indicates that the feature has a negative contribution to the measured soil moisture data of the tallgrass grassland area, and for negative values, the smaller the SHAP value, the greater the negative contribution to the measured soil moisture data of the tallgrass grassland area.

[0018] The beneficial effect of the present invention is that in the soil moisture inversion method based on multi-source remote sensing data normalization processing, by adopting the idea of ​​Z-Score standardization combined with Loess curve regression, the obvious differences in spectral wavelength ranges and spectral response function curves between different sensor images caused by spectral responses and sensor internal differences between different optical sensors, as well as differences in imaging time and environment are eliminated. Combined with the characteristics of the XGBoost model that can capture and learn complex patterns and nonlinear relationships in data, the remote sensing data after topography, environmental climate, temporal and spatial normalization and the measured tallgrass grassland area soil moisture data are input into the XGBoost model for training to obtain high-precision, high temporal and spatial resolution, and significant spatial heterogeneity predicted soil moisture data. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is the basic framework of a soil moisture inversion method based on normalization processing of multi-source remote sensing data in one embodiment of the present invention;

[0020] Figure 2Flowchart of a soil moisture inversion method based on normalization processing of multi-source remote sensing data in one embodiment of the present invention;

[0021] Figure 3 This is a Loess-Z-Score spatiotemporal normalization preprocessing process using NDVI as an example in one embodiment of the present invention;

[0022] Figure 4 Schematic diagram of modeling accuracy assessment after Loess-Z-Score spatiotemporal normalization preprocessing in one embodiment of the present invention;

[0023] Figure 5 Schematic diagram of modeling accuracy assessment without Loess-Z-Score spatiotemporal normalization preprocessing in one embodiment of the present invention;

[0024] Figure 6 Schematic diagram of modeling hierarchical accuracy assessment after Loess-Z-Score spatiotemporal normalization preprocessing in one embodiment of the present invention;

[0025] Figure 7 Schematic diagram of modeling hierarchical accuracy assessment without Loess-Z-Score spatiotemporal normalization preprocessing in one embodiment of the present invention;

[0026] Figure 8 is a Landsat 8 OLI true color image in one embodiment of the present invention;

[0027] Figure 9 This is a comparative example of soil moisture distribution maps at a depth of 10 cm without and with Loess-Z-Score spatiotemporal normalization preprocessing in one embodiment of the present invention;

[0028] Figure 10 This is a comparative example of soil moisture distribution maps at a depth of 30 cm without and with Loess-Z-Score spatiotemporal normalization preprocessing in one embodiment of the present invention;

[0029] Figure 11 This is a SHAP analysis diagram of an embodiment of the present invention without Loess-Z-Score spatiotemporal normalization preprocessing modeling;

[0030] Figure 12 This is a SHAP analysis diagram of Loess-Z-Score spatiotemporal normalization preprocessing modeling in one embodiment of the present invention. DETAILED DESCRIPTION

[0031] The following is a further description of specific embodiments of the present invention in conjunction with the accompanying drawings. It should be noted that the description of these embodiments is intended to facilitate understanding of the present invention and does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0032] like Figure 1-12 As shown in FIG, a soil moisture inversion method based on normalization processing of multi-source remote sensing data includes the following steps:

[0033] S1: Acquire environmental meteorological data, topographic data, and measured soil moisture data for the tallgrass prairie region. In this embodiment, the environmental meteorological data for the tallgrass prairie region include air temperature, relative humidity, wind speed, wind direction, maximum wind speed, solar radiation, soil temperature, vegetation evapotranspiration, rainfall data, and digital elevation model (DEM) data corresponding to the tallgrass prairie region. The DEM data is analyzed using ArcGIS 10.8.1 to obtain elevation, slope, aspect, and terrain wetness index (TWI), and the spatial resolution is unified to 30 meters. Simultaneously, the measured soil moisture data for the tallgrass prairie region is downloaded for subsequent model accuracy verification analysis.

[0034] S2: Acquire multi-source remote sensing data corresponding to the tallgrass prairie area;

[0035] S3: Perform spatiotemporal normalization on the multi-source remote sensing data corresponding to the tallgrass prairie area through Z-Score normalization, Loess curve and inverse Z-Score normalization to obtain spatiotemporal normalized remote sensing data;

[0036] S4: Match the environmental meteorological data of the tallgrass prairie area, the terrain data of the tallgrass prairie area, the remote sensing data after temporal and spatial normalization, and the measured soil moisture data of the tallgrass prairie area to establish a database, and predict the soil moisture of the tallgrass prairie area through the XGBoost model.

[0037] In the soil moisture inversion method based on multi-source remote sensing data normalization processing, the idea of ​​Z-Score standardization combined with Loess curve regression is adopted to eliminate the obvious differences in spectral wavelength range and spectral response function curves between different sensor images caused by spectral responses and sensor internal differences between different optical sensors, as well as differences in imaging time and environment. Combined with the characteristics of the XGBoost model that can capture and learn complex patterns and nonlinear relationships in data, the remote sensing data after topography, environmental climate, temporal and spatial normalization and the measured tallgrass grassland area soil moisture data are input into the XGBoost model for training to obtain high-precision, high temporal and spatial resolution, and spatially heterogeneous predicted tallgrass grassland area soil moisture data.

[0038] It is worth noting that in step S2, remote sensing data corresponding to the tallgrass prairie region for the corresponding date is obtained based on the acquisition time of the measured tallgrass prairie region soil moisture data. The vegetation index (NDVI), dual-band enhanced vegetation index (EVI2), and water index (NDWI) are calculated for each remote sensing data set. The vegetation index (NDVI) = (NIR-RED) / (NIR+RED); the dual-band enhanced vegetation index (EVI2) = 2.5*(NIR-RED) / (NIR+2.4*RED+1); and the water index (NDWI) = (GREEN-NIR) / (GREEN+NIR).

[0039] Preferably, the multi-source remote sensing data corresponding to the tallgrass prairie region includes multi-band optical satellite imagery data from Landsat-5, Landsat-7, Landsat-8, Sentinel-2, and Planetscope. In this embodiment, the multi-band optical satellite imagery data from Landsat-5, Landsat-7, Landsat-8, Sentinel-2, and Planetscope is specifically downloaded from the USGS, Copernicus Data Center, and Planet Labs websites. After obtaining the remote sensing data, images with less than 30% cloud cover are screened and resampled to a spatial resolution of 30 meters. These remote sensing data are then preprocessed, including radiometric calibration and atmospheric correction, to obtain GREEN data, RED data, and NIR data corresponding to each remote sensing data. The preprocessing process is completed using ENVI5.6 software.

[0040] Optionally, in step S3, the mean value μ and standard deviation σ of all pixel values ​​of a single image corresponding to the vegetation index NDVI, a single image corresponding to the dual-band enhanced vegetation index EVI2, and a single image corresponding to the water body index NDWI of the multi-source remote sensing data corresponding to the tall grass grassland area in the same year are calculated, and each image is standardized using Z-Score standardization; specifically, the calculation formula of Z-Score standardization is: Where X is the pixel value of the original single image, and Z is the median value after Z-Score normalization, that is, the image pixel value after Z-Score normalization. In this embodiment, the vegetation index NDVI, the dual-band enhanced vegetation index EVI2, and the water body index NDWI can be used to calculate the NDVI single image, EVI2 single image, and NDWI single image corresponding to the multi-source remote sensing data corresponding to the tallgrass prairie area. Then, Loess curves are established for the time-varying mean value μ and standard deviation σ (x-axis is year, month, and day, y-axis is pixel value). Specifically, if remote sensing data from multiple sensors is available on a given day, the mean value μ and standard deviation σ of the remote sensing data from all sensors are averaged as the mean value μ and standard deviation σ for that day.

[0041] Specifically, in step S3, Figure 3 As shown, the average value after time and space normalization is obtained according to the Loess curve corresponding to the average value μ of the date , according to the Loess curve of the date in the standard deviation σ, the standard deviation after time and space normalization is obtained ; Use inverse Z-Score standardization to obtain the single image corresponding to the vegetation index NDVI after spatiotemporal normalization correction, the single image corresponding to the dual-band enhanced vegetation index EVI2, and the single image corresponding to the water body index NDWI; use the single image corresponding to the vegetation index NDVI after spatiotemporal normalization correction, the single image corresponding to the dual-band enhanced vegetation index EVI2, and the single image corresponding to the water body index NDWI as the spatiotemporal normalized remote sensing data; specifically, the calculation formula for inverse Z-Score standardization is: ,in, is the image pixel value obtained after Loess-Z-Score correction, that is, the remote sensing data after temporal and spatial normalization.

[0042] It is worth noting that the step S32 is specifically as follows: dividing all samples composed of the mean value μ or all samples composed of the standard deviation σ into multiple intervals, performing polynomial fitting on the samples in the intervals, and repeating this process to obtain weighted regression curves in different intervals, and finally connecting the centers of these weighted regression curves together to synthesize a complete Loess curve.

[0043] Loess is a nonparametric method for local regression analysis. Its specific process is:

[0044] S321: Determine the number and positions of fitting points;

[0045] S322: Determine k closest points with the fitting point as the center;

[0046] S323: Calculate the weights of the k closest points using a weight function;

[0047] S324: Polynomial fitting (linear or quadratic) by weighted linear regression.

[0048] S325: Repeat the above steps for all fitted points.

[0049] For a fitted point , first determine its neighborhood, usually controlled by a user-defined smoothing parameter (use span or frac to control neighborhood size, range from 0 to 1, the smaller the smoothing parameter, the more dependent on local data, the more volatile the curve, the larger the smoothing parameter, the more significant the smoothing effect, but may ignore the details of the trend). The neighborhood of the fitted point is determined by the distance between the fitted point and the data points in the dataset. The data points within the neighborhood are assigned weights, and the weight function is calculated using the Tricube Kernel function. When , where is the weighted value of the fitted point ; d is the distance from the farthest point in the neighborhood to the fitted point , determined by the smoothing parameter; the weight decreases with distance, and the weight of points outside the neighborhood is 0. Within the neighborhood, a low-order polynomial (usually linear or quadratic) is fitted using weighted least squares. Taking linear regression as an example, the model is:

[0050] ; the objective is to minimize the weighted residual sum of squares: ; by solving the objective, we get the local parameter estimates and , and then predict the smoothed value of the fitted point . This smoothed value is the center of a weighted regression curve corresponding to the fitted point and y, which is a small straight line, and the final Loess curve is the weighted regression curve formed by combining all the weighted regression curves. Repeat the above steps for all fitted points in the dataset, and connect the smoothed values corresponding to each fitted point to form the complete Loess curve.

[0051]

[0052] ​Preferably, in step S4, the database is randomly divided into a training set and a validation set in a ratio of 7:3, wherein the training set is used to train the XGBoost model; the training set and the validation set are then input into the trained XGBoost model to obtain the predicted soil moisture value of the tallgrass prairie area, and the correlation coefficient R, root mean square error RMSE, bias, and unbiased root mean square error ubRMSE between the predicted soil moisture value of the tallgrass prairie area and the measured soil moisture data of the tallgrass prairie area are calculated for accuracy assessment. The trained XGBoost model is then applied to draw soil moisture distribution maps at different depths at a spatial resolution of 30 meters.

[0053] In the process of training the XGBoost model, the environmental meteorological data of the tall grass prairie area, the terrain data of the tall grass prairie area and the remote sensing data after time and space normalization are used as the input of the XGBoost model, and the measured soil moisture data of the tall grass prairie area is used as the output of the XGBoost model. In this embodiment, the correlation coefficient ; Root mean square error ;deviation ; Unbiased root mean square error ;in is the average operator, To measure soil moisture data in tallgrass grassland areas, To predict soil moisture values ​​in tallgrass prairie areas, To predict the standard deviation and is the standard deviation of the measured soil moisture data in the tallgrass prairie area.

[0054] The correlation coefficient (R), root mean square error (RMSE), bias, and unbiased root mean square error (UBRMSE) are all common metrics used to evaluate the performance of regression models. They provide information on the goodness of fit of the model from different perspectives. The correlation coefficient (R) is a statistical indicator that measures the closeness of the correlation between variables. Its value ranges from 0 to 1, with higher values ​​indicating a better fit of the model to the data. The root mean square error (RMSE) is a standard measure of the difference between the model's predicted values ​​and the actual observed values. Its value is always non-negative, and values ​​closer to 0 indicate smaller prediction errors and better model performance. Bias is the average error between the predicted values ​​and the true values, that is, the average of all prediction errors. The smaller the absolute value of the bias, the closer the model's predictions are to the true values ​​overall. The smaller the bias, the better the model's performance. The unbiased root mean square error (ubRMSE) is the root mean square error after bias correction is taken into account. It eliminates the impact of the model's systematic bias on error assessment and more accurately reflects the random error part of the model's prediction. Similar to the root mean square error (RMSE), the smaller the value of the unbiased root mean square error (ubRMSE), the higher the accuracy of the model's prediction and the more reliable the prediction result.

[0055] Optionally, in the step S4, the SHAP value is used to analyze the influence of the high grass steppe area environmental meteorological data, the high grass steppe area terrain data and the spatio-temporally normalized remote sensing data on the measured high grass steppe area soil humidity data in the trained XGBoost model; the SHAP value framework regards the high grass steppe area environmental meteorological data, the high grass steppe area terrain data and the spatio-temporally normalized remote sensing data as members of the XGBoost model, and measures the influence degree of the high grass steppe area environmental meteorological data, the high grass steppe area terrain data and the spatio-temporally normalized remote sensing data on the prediction of the XGBoost model by calculating the average marginal contribution of the high grass steppe area environmental meteorological data, the high grass steppe area terrain data and the spatio-temporally normalized remote sensing data; wherein the importance of the influence is indicated by the size of the SHAP value, a positive value indicates that the feature contributes to the measured high grass steppe area soil humidity data, for a positive value, the larger the SHAP value, the greater the contribution to the measured high grass steppe area soil humidity data, and a negative value indicates that the feature has a negative contribution to the measured high grass steppe area soil humidity data, for a negative value, the smaller the SHAP value, the greater the negative contribution to the measured high grass steppe area soil humidity data.

[0056] In the calculation of the SHAP value, all feature combinations are traversed, and the newly added contribution of the feature after joining the combination is calculated, wherein the features include the high grass steppe area environmental meteorological data, the high grass steppe area terrain data and the spatio-temporally normalized remote sensing data. Then, the newly added contributions of all combinations are weighted and averaged, wherein the weight is determined according to the size of the combination.

[0057] wherein the SHAP value of the feature i can be calculated as: ; wherein the SHAP value of the feature i represents the contribution degree of the feature i; N represents the form of all combinations of the feature i; wherein the benefit of all combinations S is represented by wherein the benefit of the combination is represented by removing the feature i, wherein the marginal contribution of the feature i in the alliance represents the marginal benefit of the feature i; wherein the weighting factor is represented by the weight, ; finally, the contribution degree of the feature i can be arranged as: .

[0058] In this scenario, the Konza tallgrass prairie, which covers an area of ​​34.87 km2 in the Flint Mountains of Manhattan, Kansas, USA, is taken as an example. The environmental meteorological data, topographic data and measured soil moisture data of the tallgrass prairie area are provided by the Konza Prairie Biological Station. The main data include: (1) Daily recorded environmental meteorological data of the tallgrass prairie area, including air temperature (℃), relative air humidity (%), wind speed (m / s), wind direction (angle), maximum wind speed (m / s, recorded every ten seconds), solar radiation (joules / m2), soil temperature (℃), the most recent rainfall (mm), the time interval from the last rainfall (days) and the daily evapotranspiration of the prairie grass (mm / d), collected by the meteorological micro-recorder at the Konza Prairie Biological Station headquarters. (2) The topographic data of the tallgrass prairie region was generated from a 2-meter resolution digital elevation model (DEM) using LiDAR DEM data collected according to USGS standards. Data analysis was performed using Arcgis 10.8.1 to obtain elevation, slope, aspect, and terrain wetness index (TWI), and the spatial resolution was unified to 30 meters. (3) The measured soil moisture data for the tallgrass prairie region were daily soil moisture data (m3 / m3) at depths of 10 and 30 cm, measured using the neutron probe method. (4) The multi-source remote sensing data corresponding to the tallgrass prairie region were multi-band optical satellite image data from Landsat-5, Landsat-7, Landsat-8, Sentinel-2, and Planetscope, downloaded from the USGS, Copernicus Data Center, and Planet Labs websites, respectively. Radiometric calibration, atmospheric correction, and flat field processing were performed using ENVI5.6, and the vegetation index NDVI, the dual-band enhanced vegetation index EVI2, and the water body index NDWI were calculated. (5) Perform Loess-Z-Score spatiotemporal normalization preprocessing on multi-source and multi-band optical sensor images.(6) The remote sensing data, environmental meteorological data and topographic data of the tall grass prairie area after time and space normalization of the same date were matched with the measured soil moisture data of the tall grass prairie area to establish a database. The multi-source remote sensing data, environmental meteorological data and topographic data of the tall grass prairie area corresponding to the original tall grass prairie area without time and space normalization preprocessing were matched with the measured soil moisture data of the tall grass prairie area to establish a database (there are two sets of databases, one corresponding to the database with time and space normalization preprocessing and the other corresponding to the database without time and space normalization preprocessing). The final sorted air temperature (T_air), air relative humidity (RHUM), wind speed (WSPEE D), wind direction (WDIR), maximum wind speed (WMAX), solar radiation (SRAD), soil temperature (STEMP), the most recent rainfall (precipitation (mm)), the time interval from the last rainfall (rain_time (day)), grassland daily evapotranspiration (ET), elevation (elevation), slope (slope), aspect (aspect), terrain wetness index (TWI), NDVI, EVI2, NDWI and soil depth (depth) data are used as the input data for machine learning, and the target variable is the daily soil moisture data (SM) at a depth of 10 and 30 cm.

[0059] Machine learning model building and accuracy assessment:

[0060] Hyperparameter tuning is crucial to the performance of machine learning models, directly controlling the behavior and effectiveness of the training algorithm. This example uses a PSO particle swarm optimization method based on five-fold cross-validation to automatically adjust the optimal hyperparameters of the model. The PSO particle swarm optimization algorithm uses the Python language pyswarm library for hyperparameter tuning. The particle optimization swarm is influenced by multiple control parameters, namely the fitness criterion (e.g., RMSE, MSE, and R2), the local coefficient (c1), the global coefficient (c2), the inertia coefficient (ω), and the population size (s). The specific parameters used in this example's PSO particle swarm optimization algorithm are as follows: fitness criterion (R2), local coefficient (c1 = 1.49618), global coefficient (c2 = 1.49618), inertia coefficient (ω = 0.7298), and population size (s = 100).

[0061] Both databases were divided into training and validation sets in a ratio of 7:3. The training set was input into the XGBoost model for training, and the validation set was used to verify the effectiveness of the model prediction results. R, RMSE, bias, and ubRMSE were used as the accuracy evaluation indicators of the model on the validation set. The accuracy evaluation results of the predicted values ​​and measured values ​​predicted by the XGBoost model for the two databases are shown in the figure below. Figure 4 and Figure 5shown.

[0062] like Figure 3 and 4 As shown in Figure 2, soil moisture prediction using PSO-XGBoost has high accuracy on both the training set and the validation set, but the accuracy of the two databases on the validation set does not change much. This example also uses June 28, 2022 as an example to predict soil moisture, and the accuracy evaluation results are as follows: Figure 6 and 7 As shown in the figure, it can be seen that PSO-XGBoost has the same high accuracy when predicting soil moisture at different depths and with different extraction methods. However, the accuracy of the validation set of the two databases in the stratified accuracy evaluation does not change much.

[0063] Machine learning model plots soil moisture distribution map:

[0064] This example uses the XGBoost models trained by the two sets of databases to draw a soil moisture distribution map with a spatial resolution of 30 meters for the data on June 28, 2022, and conducts a comparative analysis. The results are as follows: Figure 8-10 As shown in the figure, it can be found that the XGBoost model trained on the database corresponding to the Loess-Z-Score spatiotemporal normalization preprocessing has a significantly better prediction effect on some details than the XGBoost model trained on the database corresponding to the non-Loess-Z-Score spatiotemporal normalization preprocessing. Combined with optical imaging, compared with the XGBoost model trained on the database corresponding to the non-Loess-Z-Score spatiotemporal normalization preprocessing, for example, on bare soil pavement, the XGBoost model trained on the database corresponding to the Loess-Z-Score spatiotemporal normalization preprocessing can well reflect the distribution of soil moisture on the bare soil pavement.

[0065] In some areas where fires are more severe (such as Figure 8 (2) In Figure 2, 2D watersheds affected by fire have significantly less vegetation cover and more exposed soil than 1D watersheds. The spatiotemporal normalization model is more responsive to these areas and more accurately predicts differences in these regions. Furthermore, the difference in soil moisture before and after correction is more pronounced in areas with higher slopes and more severe fire conditions, as seen in Figure 2. This suggests that preprocessing multi-source optical remote sensing data using the Loess-Z-Score spatiotemporal normalization method can effectively help the XGBoost model eliminate the problem of weakening spatial heterogeneity caused by temporal variations in optical imagery from different sensors, which obscures the true surface spatial distribution information.

[0066] SHAP value analysis:

[0067] This example performs SHAP value analysis on the XGBoost models trained with the two databases. The overall SHAP value contribution analysis results are shown in the figure below: Figure 11 and 12 As shown in the figure, due to the differences in the spectral responses of different optical sensors, the values ​​of NDVI, EVI2 and NDWI images are uneven. The rankings of NDVI and EVI2 show a clear upward trend in the corrected model, especially NDVI. This shows that the Loess-Z-Score spatiotemporal normalization preprocessing method can enable XGBoost to more effectively learn the differences between NDVI, EVI2 and NDWI at different time points, thereby better explaining the mechanism by which NDVI, EVI2 and NDWI affect soil moisture.

[0068] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. It is apparent to those skilled in the art that various changes, modifications, substitutions, and variations to these embodiments may be made without departing from the principles and spirit of the present invention, and these changes and modifications still fall within the scope of protection of the present invention.

Claims

1. A soil moisture inversion method based on normalization processing of multi-source remote sensing data, characterized in that: The following steps are involved: S1: Acquire the environmental meteorological data, topographic data and measured soil moisture data of the tallgrass prairie area; S2: Acquire multi-source remote sensing data corresponding to the tallgrass prairie area; In step S2, remote sensing data corresponding to the tallgrass prairie area on the corresponding date is obtained according to the acquisition time of the measured soil moisture data of the tallgrass prairie area; the vegetation index NDVI, the dual-band enhanced vegetation index EVI2 and the water body index NDWI of each remote sensing data are calculated; S3: Perform spatiotemporal normalization on the multi-source remote sensing data corresponding to the tallgrass prairie area through Z-Score normalization, Loess curve and inverse Z-Score normalization to obtain spatiotemporal normalized remote sensing data; In step S3, the average value μ and standard deviation σ of all pixel values ​​of a single image corresponding to the vegetation index NDVI, a single image corresponding to the dual-band enhanced vegetation index EVI2, and a single image corresponding to the water body index NDWI calculated from multi-source remote sensing data corresponding to the original tallgrass prairie area in the same year are calculated, and each image is standardized using Z-Score standardization; Then, the Loess curves of the mean value μ and standard deviation σ according to time are established respectively; Obtain the time-space normalized average value based on the Loess curve corresponding to the average value μ of the date , according to the Loess curve corresponding to the standard deviation σ of the date, the standard deviation after time and space normalization is obtained ; Use the inverse Z-Score standardization to obtain the single image corresponding to the vegetation index NDVI after spatiotemporal normalization correction, the single image corresponding to the dual-band enhanced vegetation index EVI2, and the single image corresponding to the water body index NDWI; use the single image corresponding to the original vegetation index NDVI corrected by Loess-Z-Score spatiotemporal normalization, the single image corresponding to the dual-band enhanced vegetation index EVI2, and the single image corresponding to the water body index NDWI as the spatiotemporal normalized remote sensing data; S4: Match the environmental meteorological data of the tallgrass prairie area, the terrain data of the tallgrass prairie area, the remote sensing data after temporal and spatial normalization, and the measured soil moisture data of the tallgrass prairie area to establish a database, and predict the soil moisture of the tallgrass prairie area through the XGBoost model.

2. The soil moisture inversion method based on normalization processing of multi-source remote sensing data according to claim 1 is characterized in that: The multi-source remote sensing data corresponding to the tallgrass prairie area includes Landsat-5, Landsat-7, Landsat-8, Sentinel-2 and Planetscope multi-band optical satellite image data. After obtaining the remote sensing data, images with less than 30% cloud cover were screened and resampled to a spatial resolution of 30 meters.

3. The soil moisture inversion method based on normalization processing of multi-source remote sensing data according to claim 2 is characterized in that: Divide all samples consisting of mean μ or all samples consisting of standard deviation σ into multiple intervals, perform polynomial fitting on the samples in the intervals, repeat this process to obtain weighted regression curves in different intervals, and finally connect the centers of these weighted regression curves together to synthesize a complete Loess curve.

4. The soil moisture inversion method based on normalization processing of multi-source remote sensing data according to claim 3 is characterized in that: In step S4, the database is randomly divided into a training set and a validation set in a ratio of 7:3, wherein the training set is used to train the XGBoost model; then the training set and the validation set are input into the trained XGBoost model to obtain the predicted soil moisture value of the tallgrass prairie area, and the correlation coefficient R, root mean square error RMSE, bias and unbiased root mean square error ubRMSE between the predicted soil moisture value of the tallgrass prairie area and the measured soil moisture data of the tallgrass prairie area are calculated for accuracy assessment.

5. The soil moisture inversion method based on normalization processing of multi-source remote sensing data according to claim 4 is characterized in that: In step S4, the SHAP value is used to analyze the influence of the environmental meteorological data of the tallgrass grassland area, the terrain data of the tallgrass grassland area and the remote sensing data after spatiotemporal normalization in the trained XGBoost model on the measured soil moisture data of the tallgrass grassland area; the importance of the influence is represented by the size of the SHAP value, a positive value indicates that the feature contributes to the measured soil moisture data of the tallgrass grassland area, for positive values, the larger the SHAP value, the greater the contribution to the measured soil moisture data of the tallgrass grassland area, and the negative value indicates that the feature has a negative contribution to the measured soil moisture data of the tallgrass grassland area, and for negative values, the smaller the SHAP value, the greater the negative contribution to the measured soil moisture data of the tallgrass grassland area.

Images