Method and system for high-resolution inversion of soil moisture based on multi-source remote sensing and deep regression

Abstract

The application belongs to the technical field of remote sensing data processing and agricultural environment monitoring, and particularly relates to a soil moisture high-resolution inversion method and system based on multi-source remote sensing and deep regression, which comprises the following steps: multi-source remote sensing data acquisition and preprocessing; spatio-temporal resolution unification; driving variable and reference variable construction; Transformer regression network construction; model training and inversion; and high spatio-temporal resolution soil moisture data output. Through the cooperation of multi-source heterogeneous remote sensing data, the application can realize the complementary advantages of brightness temperature data and SAR backscattering data under different spatio-temporal accuracies, develop a high spatio-temporal resolution SM inversion framework suitable for large-scale agricultural areas, and be applied to agricultural area SM extraction, so as to guide large-scale precision water resource management.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of remote sensing data processing and agricultural environmental monitoring technology, and particularly relates to a high-resolution soil moisture inversion method and system based on multi-source remote sensing and depth regression. Background Technology

[0002] Soil moisture (SM) is a key environmental parameter affecting crop growth, agricultural water use efficiency, and regional hydrological cycles. Accurate acquisition of soil moisture information with high spatiotemporal resolution is of great significance for agricultural irrigation management, drought monitoring, and crop yield prediction.

[0003] Currently, soil moisture acquisition methods mainly include ground station measurements and remote sensing inversion. Ground station measurements offer high accuracy but have limited coverage, failing to meet the needs of large-scale, continuous monitoring. Remote sensing technology provides an effective means for large-scale soil moisture monitoring, but existing remote sensing data products suffer from a trade-off between spatial and temporal resolution: active microwave remote sensing (such as Sentinel-1) offers high spatial resolution (up to 10 meters) but low temporal resolution (long revisit intervals); passive microwave remote sensing (such as SMAP) offers high temporal resolution (daily coverage) but low spatial resolution (above 9 kilometers).

[0004] To overcome the aforementioned problems, this invention attempts to fuse multi-source remote sensing data. Existing fusion methods are mostly based on physical models or traditional machine learning models (such as random forests and support vector regression), but these methods have limited ability to process high-dimensional heterogeneous data, making it difficult to fully exploit the complementary advantages of multi-source data. In recent years, deep learning models have shown potential in remote sensing inversion, but large-scale, high spatiotemporal resolution soil moisture inversion methods based on the Transformer structure have not yet been systematically studied.

[0005] Based on the above analysis, the urgent technical problems that need to be solved in the existing technology are:

[0006] There is an urgent need for a method that can integrate active and passive microwave data, optical data, and environmental parameters, combined with a deep regression network, to achieve high spatiotemporal resolution soil moisture retrieval. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides a high-resolution soil moisture inversion method and system based on multi-source remote sensing and depth regression, which is applicable to large-scale agricultural areas.

[0008] This invention is implemented as follows: a high-resolution soil moisture inversion method based on multi-source remote sensing and depth regression, the method comprising:

[0009] S1: Acquire multi-source heterogeneous remote sensing data, including active microwave data, passive microwave data, optical remote sensing data, and environmental parameter data;

[0010] S2: Perform spatiotemporal resolution unification processing on the multi-source heterogeneous remote sensing data to construct a set of driving variables for the target spatiotemporal resolution;

[0011] S3: Construct a reference variable set based on ground-measured soil moisture data;

[0012] S4: Construct a soil moisture inversion model based on a Transformer regression network. The model includes an input layer, multiple encoding layers, and a regression output layer.

[0013] S5: Train the inversion model using the set of driving variables and the set of reference variables, and use the mean squared error as the loss function;

[0014] S6: Input the multi-source remote sensing data of the area to be inverted into the trained model, and output high spatiotemporal resolution soil moisture data.

[0015] Furthermore, the active microwave data includes synthetic aperture radar (SAR) data, particularly its backscattering coefficient; the passive microwave data includes spaceborne radiometer data, particularly its brightness temperature data (SMAP).

[0016] Furthermore, the optical remote sensing data includes multispectral data used to calculate the normalized vegetation index.

[0017] Furthermore, the environmental parameter data includes one or more of the following: surface temperature, precipitation, evapotranspiration, soil roughness, and topography.

[0018] Furthermore, the spatiotemporal resolution unification process includes: unifying the temporal and spatial resolutions of different data sources to a preset target spatiotemporal resolution, wherein the target spatiotemporal resolution is set according to application requirements, and low-resolution data is downscaled using interpolation or resampling methods.

[0019] Furthermore, the Transformer regression network structure includes an input layer, six fully connected layers, six Transformer encoding layers, and a regression output layer. Each Transformer encoding layer includes a four-head self-attention mechanism and a feedforward neural network.

[0020] Furthermore, the inversion model is trained using mean squared error as the loss function, with preset iteration rounds, learning rate, and batch size.

[0021] Furthermore, the set of driving variables includes active microwave features, passive microwave features, vegetation index, surface temperature, precipitation, evapotranspiration, soil roughness, and topographic information, constituting a multidimensional time series for each pixel.

[0022] Another objective of this invention is to provide a high-resolution soil moisture retrieval system based on multi-source remote sensing and depth regression, which can be divided into two modules:

[0023] The building blocks for driving and reference variables are used to construct driving and reference variable datasets, including active and passive microwave features, optical features, environmental parameters, and field-measured SM data.

[0024] The SM inversion module is used to construct an SM inversion model using the driving and reference variable datasets. The SM inversion model is constructed based on the Transformer regression method.

[0025] In combination with the above technical solutions and the technical problems solved, the significant advancements and unexpected technical effects of the technical solution to be protected by this invention are as follows:

[0026] This invention acquires multi-source heterogeneous remote sensing data, fusing active and passive microwave data, optical data, and environmental parameters to achieve spatiotemporal complementarity. It performs unified spatiotemporal resolution processing on the multi-source heterogeneous remote sensing data, achieving for the first time daily soil moisture retrieval with a 10-meter spatial resolution, suitable for fine-grained management at the farmland scale. This invention constructs a deep regression model for soil moisture retrieval, introducing a Transformer network structure, which outperforms traditional machine learning and convolutional neural network methods, achieving high accuracy.

[0027] This invention does not simply combine active microwave data, passive microwave data, optical remote sensing data, environmental parameter data, and deep regression networks in parallel. Instead, it uses a unified target spatial grid and target time interval as a foundation to match remote sensing data from different sources, at different scales, and with different observation periods to the same target pixel and corresponding observation time. Furthermore, it organizes the multi-source features of continuous observation times into a pixel-level multidimensional time series. This processing method establishes a correspondence between multi-source remote sensing observation information in terms of spatial location, temporal order, and variable dimensions, reducing the interference caused by inconsistencies in the spatiotemporal scales of different data sources on soil moisture inversion results. This ensures that the model input maintains clear pixel assignments and temporal evolution relationships.

[0028] This invention matches ground-measured soil moisture data to corresponding target pixels and observation times, forming a reference variable set corresponding to the driving variable set, and uses this correspondence to supervise the training of the Transformer regression network. Compared to technical solutions that only perform static pixel regression based on remote sensing features, this invention utilizes a multi-head self-attention mechanism to encode the feature associations of the same target pixel at consecutive observation times. This ensures that active microwave features, passive microwave features, vegetation information, and environmental parameters are no longer treated as independent instantaneous inputs, but rather form a joint feature representation with temporal correlation.

[0029] Active microwave data can characterize surface scattering characteristics, passive microwave data can characterize surface microwave radiation characteristics, optical remote sensing data can characterize vegetation cover, and environmental parameters can reflect surface energy exchange, water replenishment, and topographic conditions. This invention, through unified spatiotemporal matching and temporal correlation coding, enables data with different observation mechanisms to collaboratively participate in soil moisture inversion within the same pixel time series. This reduces inversion biases caused by limitations of vegetation cover, surface roughness, weather conditions, and spatial resolution imposed by single remote sensing data.

[0030] This invention also sets the training and actual inversion stages as the same data processing chain, so that the data of the area to be inverted is input into the trained model according to the spatial matching, temporal matching, resolution uniformity, and time series construction method of the training data, reducing the data structure difference between the training samples and the actual inversion samples. Based on the synergistic effect of the above technical features, this invention can improve the spatial integrity, temporal continuity, and local change identification ability of soil moisture inversion results by utilizing multi-source observation information at continuous time while maintaining the refinement of the target spatial grid. Especially when the information from a single data source is insufficient or there are spatiotemporal resolution differences between different data sources, it can still obtain inversion results with a good correspondence to the ground-measured soil moisture. This effect cannot be directly expected by simply using data resampling or conventional regression networks.

[0031] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0032] The high spatiotemporal resolution soil moisture inversion model constructed by the technical solution of this invention, after being put into practical application, can directly generate daily soil moisture monitoring data products with a spatial resolution of 10 meters. This product can provide core data support for precision agricultural irrigation decisions, crop growth assessment, and regional drought early warning. Its target users include large-scale agricultural planting enterprises, agricultural insurance institutions, and agricultural and water conservancy management departments at all levels. Compared with the spatiotemporal resolution constraints that are common in existing commercial remote sensing soil moisture products, the product provided by this invention has achieved significant improvements in both inversion accuracy and spatial detail characterization. Based on the differentiated competitive barrier formed by this technological advantage, it can achieve continuous revenue through various business models such as data product licensing, platform subscription services, and customized solutions. It is expected to have broad commercial application prospects and considerable economic returns in the domestic and international remote sensing agricultural monitoring and service market.

[0033] (2) The technical solution of this invention fills a technical gap in the industry both domestically and internationally:

[0034] For a long time, the field of remote sensing soil moisture retrieval has faced the inherent contradiction of the incompatibility between the high spatial resolution of active microwave remote sensing and the high temporal resolution of passive microwave remote sensing. Existing single remote sensing sources or simple fusion methods cannot simultaneously achieve continuous soil moisture monitoring at daily temporal resolution and 10-meter spatial resolution over large areas. This invention constructs a unified spatiotemporal resolution processing framework for multi-source heterogeneous remote sensing data and introduces a Transformer deep regression network on this basis. For the first time, it deeply fuses active microwave backscattering coefficients, passive microwave brightness temperature data, optical vegetation indices, and multiple environmental parameters in the spatiotemporal dimensions, successfully achieving stable daily soil moisture retrieval at 10-meter resolution in large-area agricultural areas. This effectively fills the technological gap in high spatiotemporal resolution soil moisture remote sensing products both domestically and internationally.

[0035] (3) The technical solution of the present invention solves a technical problem that people have long wanted to solve but have never been able to solve successfully:

[0036] Acquiring soil moisture information with high temporal and spatial resolution simultaneously over a large area has been a long-standing and unresolved technical challenge in this field. The difficulty lies in the fundamental differences in spatiotemporal spectral characteristics among different remote sensing data sources, making it difficult for traditional physical models and shallow machine learning methods to fully exploit the complementary information between multi-source data. This invention addresses this challenge by uniformly processing heterogeneous multi-source data into multi-dimensional time-series inputs for each pixel. It effectively extracts deep correlations between features in the temporal and spatial dimensions using a multi-head self-attention mechanism in a multi-layer Transformer coding structure. Combined with an end-to-end regression training strategy, this substantially solves the aforementioned problem. Verification experiments show that the coefficient of determination between the inversion results and ground-measured values ​​reaches 0.9220, and the root mean square error is only 0.0319 cm³ / cm³, fully demonstrating the practical ability of this technical solution to solve this long-standing technical problem.

[0037] (4) The technical solution of the present invention overcomes technical bias:

[0038] In this field, it is generally believed that the inherent differences in spatiotemporal resolution between active and passive microwave remote sensing cannot be fundamentally bridged through data processing, making high spatiotemporal resolution soil moisture inversion technically infeasible. Furthermore, most researchers tend to use traditional machine learning methods such as random forests and support vector regression for remote sensing parameter inversion, believing these methods already meet application requirements, while holding reservations about introducing more complex deep network structures. This invention, through systematic comparative experiments, demonstrates that the inversion model built based on a Transformer regression network significantly outperforms traditional machine learning methods and convolutional neural network methods across all evaluation metrics, and successfully achieves daily soil moisture output at a 10-meter resolution. This provides solid experimental data to overcome the aforementioned technical biases, offering a new technical path and empirical evidence for subsequent research. Attached Figure Description

[0039] Figure 1 This is a flowchart of the high-resolution soil moisture inversion method based on multi-source remote sensing and depth regression provided in this embodiment of the invention;

[0040] Figure 2 This is a diagram of the high-resolution soil moisture inversion system architecture based on multi-source remote sensing and depth regression provided in this embodiment of the invention.

[0041] Figure 3 This is a diagram of the Transformer regression network structure provided in an embodiment of the present invention;

[0042] Figure 4 This is a scatter plot comparing the accuracy of different inversion methods provided in the embodiments of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0044] The study area was the agricultural region of the Pacific Northwest of the United States, spanning from March 1, 2021 to July 31, 2021. This region includes winter wheat, spring wheat, and corn growing areas, covering approximately 230,000 square kilometers.

[0045] like Figure 1 As shown, this embodiment of the invention provides a high-resolution soil moisture inversion method based on multi-source remote sensing and depth regression. The method includes:

[0046] S1: Acquisition and Preprocessing of Multi-Source Remote Sensing Data

[0047] Obtain Sentinel-1 backscattering coefficient, SMAP brightness temperature, NDVI, surface temperature, precipitation, evapotranspiration, soil roughness, and topographic data:

[0048] (1) Backscattering coefficients of VH and VV of Sentinel-1 GRD product, with a spatial resolution of 10 meters and a temporal resolution of 12 days;

[0049] (2) SMAP enhanced L1C brightness temperature data, with a spatial resolution of 9 km and a temporal resolution of 1 day;

[0050] (3) Calculate daily NDVI using MODIS near-infrared and red band data;

[0051] (4) Environmental parameters: surface temperature, precipitation, evapotranspiration, soil roughness, and topography. Evapotranspiration: 9 km, 1 day; soil roughness: 9 km; topography: 30 m.

[0052] (5) Ground measured data: Soil moisture data at a depth of 0–5 cm from 32 stations provided by ISMN.

[0053] S2: Data Processing

[0054] (1) All data are uniformly set to daily and 10-meter resolution. In order to maintain the original data distribution characteristics and avoid introducing false values, nearest neighbor interpolation resampling is adopted.

[0055] (2) Perform time-dimensional linear interpolation on the NDVI cloud-covered pixels.

[0056] S3: Construction of driving and reference variables:

[0057] The driving variables are multidimensional sequences for each pixel, including active and passive microwave features, optical features, and environmental parameters. The reference variable is the measured soil moisture data from the ground station at a depth of 0-5 cm.

[0058] S4: Model Building and Training

[0059] (1) Construct a 6-layer Transformer encoding layer with 4 self-attention heads in each layer, and use a sliding window to input the multidimensional time series into the Transformer;

[0060] (2) Loss function: MSE; Number of training rounds: 1000; Learning rate: 0.0001; Batch size: all samples;

[0061] (3) Randomly select 30% of the samples as the training set and the rest as the test set.

[0062] S5: Model Training and Inversion

[0063] The model is trained using driving and reference variables, and outputs daily soil moisture data at a 10-meter resolution.

[0064] S6: Accuracy Evaluation

[0065] Calculate the coefficient of determination (R²) 2 The average bias (Bais), relative prediction bias (RPD), root mean square error (RMSE), and uncalibrated root mean square prediction error (ubRMSE) are also included.

[0066] like Figure 2 As shown, this embodiment of the invention provides a high-resolution soil moisture inversion system based on multi-source remote sensing and depth regression. This system can be divided into two modules. Module 1 constructs the driving and reference variables, and module 2 performs SM inversion. Module 1 constructs the driving and reference variable datasets, including active and passive microwave features, optical features, environmental parameters, and field-measured SM data. In module 2, this invention utilizes the driving and reference variable datasets to construct an SM inversion model based on the Transformer regression method.

[0067] Driving variables are the input feature variables of the model, used to drive the model to obtain SM results; reference variables are the true values ​​of the model's output results, mainly used for training the model and evaluating its performance. Driving variables include active and passive microwave characteristics, optical characteristics, and environmental parameters, while reference variables are field-measured SM data, including 32 stations.

[0068] The structure of the Transformer regression model is shown below. Figure 3 The model consists of an input layer, six fully connected layers, six Transformer encoding layers, and a regression output layer. Each Transformer encoding layer contains a multi-head self-attention network and a feedforward neural network. Figure 3 The feature sequence is composed of multi-source remote sensing data, with a total of 8 feature types: Sentinel-1 backscattering coefficient, SMAP brightness temperature, land surface temperature (LST), soil roughness, evapotranspiration, precipitation, and vegetation index (NDVI). The Sentinel-1 backscattering coefficient includes vertical transmit / horizontal receive (VH) and vertical transmit / vertical receive (VV); the SMAP brightness temperature includes horizontal polarization brightness temperature (TBH) and vertical polarization brightness temperature (TBV).

[0069] The process of SM inversion using this network structure can be described as follows: First, driving variables such as active and passive microwave features, optical features, and environmental parameters are input as model parameters by the input layer; then, the input model parameters are processed through a 6-layer Transformer encoding layer to extract features; finally, the SM inversion result is output by the regression output layer. This method utilizes the Transformer model to improve the accuracy and robustness of SM inversion across the entire study area. The parameter settings for training the Transformer regression network are as follows: the loss function is the mean squared error function, the number of training epochs is 1000, the learning rate is 0.0001, and the batch size of a single input is all samples. By training the Transformer regression model using driving and reference variables, an SM inversion model with a spatiotemporal resolution of 1 day and 10 meters is established. In addition, the reference variable will also be used to evaluate the accuracy of SM inversion.

[0070] Evidence related to the technical effects obtained by the embodiments of the present invention.

[0071] The inversion results are as follows:

[0072] (1) Comparison between inverted soil moisture and measured values: R²=0.9220, Bias=0.0176 cm³ / cm³, RPD=3.5696, RMSE=0.0319 cm³ / cm³, ubRMSE=0.0266 cm³ / cm³; N represents the number of samples;

[0073] (2) Compared with traditional methods (SVR, RF, XGBoost, GRNN, CNNR), the Transformer method has the highest accuracy.

[0074] Table 1. Comparison of SM inversion evaluation metrics between Transformer method and traditional method

[0075] N 3166 3166 3166 3166 3166 3166 <![CDATA[R 2 ]]> 0.8480 0.8886 0.9019 0.8456 0.9064 0.9220 Bias 0.0299 0.0278 0.0229 0.0243 0.0176 0.0176 RPD 2.5206 2.8156 3.1714 2.5447 3.2667 3.5696 RMSE 0.0446 0.0403 0.0354 0.0445 0.0345 0.0319 ubRMSE 0.0331 0.0292 0.0270 0.0373 0.0297 0.0266

[0076] (3) Compared with existing products (AMSR-U, ESACCI, SMAP L4), the Transformer method has significantly better spatial continuity and temporal variation inversion capabilities;

[0077] Table 2. Comparison of SM inversion evaluation indicators between the Transformer method and existing technology products.

[0078] N 3166 3404 2110 2751 4738 <![CDATA[R 2 ]]> 0.9220 0.0614 0.2349 0.6798 0.1936 Bias 0.0176 0.1001 0.0880 0.1006 0.0936 RPD 3.5696 0.3862 0.8746 0.4146 0.7642 RMSE 0.0319 0.1216 0.1140 0.1189 0.1177 ubRMSE 0.0266 0.0691 0.0724 0.0634 0.0713

[0079] like Figure 4As shown, the Transformer regression provided in this embodiment of the invention is compared with other mainstream regression methods to analyze the performance of the SM inversion framework built based on Transformer regression in large-scale spatiotemporal continuous SM inversion. 30% of the sample data was randomly selected as training data for constructing the SM inversion model, which was then used to invert the SM of the remaining sample points. Finally, the accuracy of the inversion results was evaluated. By comparing the performance of different regression methods, the advantages of Transformer regression in SM inversion are discussed. These regression methods include Support Vector Regression (SVR), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Generalized Recurrent Neural Network (GRNN), and Convolutional Neural Network Regression (CNNR).

[0080] Figure 4 In the diagram, N represents the number of samples. The red solid line is the linear fit line between the reference result and the inversion result. The black solid line from the lower left to the upper right is the reference line. The color bars show the number of data points within each 0.05 × 0.05 cm³ / cm³ SM range in the scatter plot. The inversion results show that traditional machine learning regression methods, traditional neural network regression methods, and the Transformer regression method (used for the first time in large-scale SM inversion) can all obtain relatively ideal SMs. From the evaluation metrics, Transformer and CNNR outperform other regression methods. Among neural network regression methods, Transformer outperforms CNNR. Specifically, in the SM results inverted by CNNR, R... 2 The values ​​of , Bais, RPD, RMSE, and ubRMSE are 0.9064, 0.0176 cm³ / cm³, 3.2667, 0.0345 cm³ / cm³, and 0.0297 cm³ / cm³, respectively. However, in the SM results obtained through Transformer inversion, R… 2 The values ​​of BAIS, RPD, RMSE, and ubRMSE reached 0.9220, 0.0176 cm³ / cm³, 3.5696, 0.0319 cm³ / cm³, and 0.0266 cm³ / cm³, respectively. Among traditional machine learning regression methods, XGBoost performed best, while SVR yielded the worst results. In the SM results inverted by XGBoost, R… 2 The values ​​of , Bais, RPD, RMSE, and ubRMSE were 0.9019, 0.0229 cm³ / cm³, 3.1714, 0.0354 cm³ / cm³, and 0.0270 cm³ / cm³, respectively. However, in the SM results retrieved from the SVR, R… 2The values ​​of Bais, RPD, RMSE and ubRMSE were 0.8480, 0.0299 cm³ / cm³, 2.5206, 0.0446 cm³ / cm³ and 0.0331 cm³ / cm³, respectively.

[0081] In the description of this invention, unless otherwise stated, "a plurality of" means two or more; the terms "upper," "lower," "left," "right," "inner," "outer," "front end," "rear end," "head," "tail," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0082] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A high-resolution soil moisture inversion method based on multi-source remote sensing and depth regression, characterized in that, include: Acquire active microwave data, passive microwave data, optical remote sensing data, environmental parameter data, and ground-measured soil moisture data for the target area at multiple observation times; According to the target spatial grid and target time interval, the active microwave data, passive microwave data, optical remote sensing data and environmental parameter data are spatially matched, temporally matched and resolution unified. The active microwave features, passive microwave features, optical features and environmental parameters of the same target pixel at continuous observation time are combined into a multi-dimensional time series to form a set of driving variables. The measured soil moisture data on the ground is matched with the corresponding target pixels and observation times to form a reference variable set corresponding to the driving variable set; A Transformer regression network is constructed. The Transformer regression network receives the multi-dimensional time series, performs feature mapping on the multi-source features at each observation time, and encodes the feature association between consecutive observation times based on the multi-head self-attention relationship. Based on the encoding result, the network outputs the soil moisture inversion value of the corresponding target pixel and the observation time. The Transformer regression network is trained using the set of driving variables as model input and the set of reference variables as supervision target. The active microwave data, passive microwave data, optical remote sensing data, and environmental parameter data of the area to be inverted are processed according to the spatial matching, temporal matching, and resolution unification methods. The resulting multidimensional time series is input into the trained Transformer regression network, and the output is soil moisture data with the target spatial grid and target time interval.

2. The method according to claim 1, characterized in that, The active microwave data includes synthetic aperture radar data, and the active microwave features include backscattering coefficients. The passive microwave data includes spaceborne radiometer data, and the passive microwave characteristics include brightness temperature data. The optical remote sensing data includes multispectral data, and the optical features include the normalized vegetation index calculated based on the multispectral data. The environmental parameter data includes at least one of the following: surface temperature, precipitation, evapotranspiration, soil roughness, and topography.

3. The method according to claim 1, characterized in that, The spatial matching, temporal matching, and resolution unification include: Determine the target spatial resolution and target temporal resolution based on the application requirements of soil moisture data; According to the target spatial resolution, different data sources are mapped to the same target spatial grid; According to the target time resolution, different data sources are matched to the corresponding target observation time. Perform interpolation or resampling on data that is below the target spatial resolution or target temporal resolution; Based on the target pixel and the target observation time, the processed active microwave features, passive microwave features, optical features and environmental parameters are combined accordingly.

4. The method according to claim 1, characterized in that, The Transformer regression network comprises an input layer, six fully connected layers, six Transformer encoding layers, and a regression output layer connected in sequence. The input layer receives a multidimensional time series of the same target pixel at consecutive observation times; The six fully connected layers perform feature mapping on the active microwave features, passive microwave features, optical features, and environmental parameters at each observation time. Each Transformer coding layer includes four self-attention units and a feedforward neural network. The four self-attention units determine the correlation between the mapping features at consecutive observation times and combine the processing results of their respective attention units. The feedforward neural network performs feature transformation on the combined processing results. The regression output layer generates soil moisture inversion values ​​based on the feature transformation results output by the last Transformer encoding layer.

5. The method according to claim 1, characterized in that, The Transformer regression network uses mean squared error as the loss function, which is determined by the average value of the squared difference between the soil moisture inversion value and the corresponding ground measured soil moisture value in the training samples. The Transformer regression network is trained according to a preset number of iteration rounds, a preset learning rate, and a preset batch size; Each training sample in the driving variable set includes active microwave features, passive microwave features, normalized vegetation index, surface temperature, precipitation, evapotranspiration, soil roughness, and topographic information of the same target pixel at continuous observation times. The reference variable set includes the measured soil moisture values ​​on the ground corresponding to the target pixel and the target observation time of the training sample.

6. A high-resolution soil moisture inversion system based on multi-source remote sensing and depth regression, implementing the method of any one of claims 1-5, characterized in that, Includes modules for constructing driving and reference variables, as well as a soil moisture inversion module; The driving variable and reference variable construction module is used to acquire active microwave data, passive microwave data, optical remote sensing data, environmental parameter data, and ground-measured soil moisture data of the target area at multiple observation times. According to the target spatial grid and target time interval, the active microwave data, passive microwave data, optical remote sensing data, and environmental parameter data are spatially matched, temporally matched, and resolution unified. The active microwave features, passive microwave features, optical features, and environmental parameters of the same target pixel at continuous observation times are combined into a multi-dimensional time series to form a driving variable set. The ground-measured soil moisture data are matched with the corresponding target pixels and observation times to form a reference variable set. The soil moisture inversion module is used to construct a Transformer regression network. The driving variable set is used as the model input, and the reference variable set is used as the supervision target to train the Transformer regression network. The Transformer regression network performs feature mapping on the multi-source features at each observation time, encodes the feature association between consecutive observation times based on the multi-head self-attention relationship, and outputs the soil moisture inversion value of the corresponding target pixel and observation time according to the encoding result. The soil moisture inversion module is also used to receive the multidimensional time series of the region to be inverted, which is processed according to the spatial matching, temporal matching and resolution unification method, and to output soil moisture data with the target spatial grid and target time interval using the trained Transformer regression network.

7. The high-resolution soil moisture inversion system based on multi-source remote sensing and depth regression according to claim 6, characterized in that, The active microwave data received by the driving variable and reference variable construction module includes synthetic aperture radar data, and the backscattering coefficients are obtained from the synthetic aperture radar data. The passive microwave data includes spaceborne radiometer data, and the driving variable and reference variable construction module obtains brightness temperature data from the spaceborne radiometer data; The optical remote sensing data includes multispectral data, and the driving variable and reference variable construction module calculates the normalized vegetation index based on the multispectral data. The environmental parameter data includes at least one of the following: surface temperature, precipitation, evapotranspiration, soil roughness, and topography.

8. The high-resolution soil moisture retrieval system based on multi-source remote sensing and depth regression according to claim 6, characterized in that, The driving variable and reference variable construction module is used for: Determine the target spatial resolution and target temporal resolution based on the application requirements of soil moisture data; Map different data sources to the same target space grid according to the target spatial resolution; Match different data sources to the corresponding target observation time according to the target time resolution; Perform interpolation or resampling on data that is below the target spatial resolution or target temporal resolution; Based on the target pixel and the target observation time, the processed active microwave features, passive microwave features, optical features and environmental parameters are combined into a multi-dimensional time series; The measured soil moisture data on the ground is matched with the corresponding target pixels and target observation time.

9. The high-resolution soil moisture inversion system based on multi-source remote sensing and depth regression according to claim 6, characterized in that, The Transformer regression network comprises an input layer, six fully connected layers, six Transformer encoding layers, and a regression output layer connected in sequence. The input layer is used to receive multidimensional time series of the same target pixel at consecutive observation times; The six fully connected layers are used to perform feature mapping on the active microwave features, passive microwave features, optical features, and environmental parameters at each observation time. Each Transformer coding layer includes four self-attention units and a feedforward neural network. The four self-attention units are used to determine the correlation between the mapping features at consecutive observation times and combine the processing results of their respective attention units. The feedforward neural network is used to perform feature transformation on the combined processing results. The regression output layer is used to generate soil moisture inversion values ​​based on the feature transformation results output by the last Transformer encoding layer.

10. The high-resolution soil moisture inversion system based on multi-source remote sensing and depth regression according to claim 6, characterized in that, The soil moisture inversion module uses mean square error as the loss function. The mean square error is determined based on the average value of the squared difference between the soil moisture inversion value and the corresponding ground measured soil moisture value in the training sample. The soil moisture inversion module trains the Transformer regression network according to a preset number of iterations, a preset learning rate, and a preset batch size. Each training sample in the set of driving variables includes the active microwave features, passive microwave features, normalized vegetation index, surface temperature, precipitation, evapotranspiration, soil roughness, and topographic information of the same target pixel at consecutive observation times. The reference variable set includes the ground-measured soil moisture values ​​corresponding to the target pixels and target observation times of the training samples.

Images