A method and system for multi-source remote sensing inversion of farmland soil water based on crop physiological perception

Abstract

This invention discloses a method and system for retrieving farmland soil water based on crop physiological perception through multi-source remote sensing. The method includes: acquiring synthetic aperture radar data and multispectral optical remote sensing data and preprocessing them to obtain remote sensing data for the target date and daily-scale time-series data; constructing a multi-branch neural network model, including a main branch for soil water retrieval and three auxiliary branches for vegetation structure corresponding to crop type, crop height, and phenological stage information, respectively; employing a two-stage training strategy, first training the auxiliary branches with measured vegetation parameters, then freezing their parameters and training the main branch with measured soil moisture content; using the trained model to perform pixel-level prediction, and combining it with land use data masks to obtain farmland soil water products. This invention introduces crop physiological parameters as constraints to explicitly characterize vegetation structure features, effectively reducing vegetation interference and improving the accuracy, stability, and generalization ability of soil water retrieval in farmland with dense crop canopies.

Description

Technical Field

[0001] This invention relates to the fields of agricultural water resources management and agricultural remote sensing, and in particular to a method and system for multi-source remote sensing inversion of farmland soil water based on crop physiological perception. Background Technology

[0002] Soil moisture content is a crucial component of the terrestrial water cycle and one of the most critical environmental variables in agricultural production systems. It directly affects crop root water absorption capacity, photosynthetic efficiency, and nutrient transport processes, and significantly determines crop yield and quality. Therefore, obtaining high-precision, high-spatiotemporal resolution soil moisture information is of great significance for precision irrigation, agricultural drought early warning, and regional ecological environment assessment. Traditional soil moisture monitoring mainly relies on ground-based station observations or manual sampling and analysis. While this method offers high accuracy, it suffers from high equipment costs, maintenance difficulties, and sparse spatial distribution, making it unsuitable for large-scale, continuous, dynamic monitoring.

[0003] With the development of Earth observation technology, remote sensing methods have gradually become an important means of obtaining soil moisture information. Among them, microwave remote sensing, especially synthetic aperture radar, has significant advantages in soil water retrieval due to its all-weather, all-day observation capabilities and sensitivity to soil moisture changes. However, in farmland environments, the propagation process of electromagnetic waves in the soil-vegetation system becomes extremely complex due to the long-term presence of vegetation cover. Radar backscattered signals include not only scattering from the soil surface but also scattering from vegetation and the coupling effect of multiple scattering between vegetation and soil. During the vigorous growth period of crops, the complex canopy structure, large biomass, and high water content cause significant attenuation of microwave signals during penetration, severely masking soil information and making soil water retrieval a typical nonlinear ill-conditioned problem.

[0004] To address the aforementioned issues, researchers have proposed various inversion methods, including physical models based on electromagnetic scattering theory, semi-empirical models (such as water cloud models), change detection methods, and machine learning methods that have emerged in recent years. However, these methods all have significant limitations in complex farmland scenarios: physical model parameters are difficult to obtain and computationally complex; semi-empirical models oversimplify vegetation structure, making it difficult to characterize different crop types and phenological changes; change detection methods rely on stability assumptions and struggle to obtain absolute water content; data-driven methods, while highly accurate, lack physical constraints, are prone to spurious correlations, and have limited generalization ability. Furthermore, existing methods generally treat vegetation as a "distraction term" for weakening or removal, failing to effectively utilize the important information contained in the vegetation structure. Therefore, there is an urgent need to propose a new method that can explicitly model vegetation structure and integrate multi-source remote sensing information to improve the accuracy and stability of farmland soil water inversion. Summary of the Invention

[0005] This invention addresses the problems of severe vegetation interference, insufficient model generalization ability, and lack of physical constraints in existing farmland soil moisture remote sensing inversion processes. It proposes a multi-source remote sensing inversion method for farmland soil moisture based on crop physiological perception. This method breaks away from the traditional approach of treating vegetation as a "distraction term" to be weakened or removed. Instead, it introduces vegetation structure and physiological information as important constraints into the inversion model. Through a multi-branch neural network structure, it achieves collaborative modeling of vegetation information and soil signals, thereby significantly improving the accuracy and stability of soil moisture inversion.

[0006] According to one aspect of the present invention, a method for retrieving farmland soil water based on multi-source remote sensing of crop physiological perception is provided, comprising: Synthetic aperture radar data and multispectral optical remote sensing data of the study area were acquired and preprocessed to obtain target date remote sensing data and daily time series data. The target date remote sensing data and daily-scale time series data are input into the trained multi-branch neural network model, which outputs the spatial distribution results of soil moisture. Combined with land use data, farmland masking is performed to obtain farmland soil moisture products. The training of the multi-branch neural network model further includes: The training dataset was constructed, including ground-measured data covering information on topsoil moisture content, crop height, crop type and phenological stage, as well as target date remote sensing data and daily time-series data obtained by processing synthetic aperture radar data and multispectral optical remote sensing data, respectively. A multi-branch neural network model is constructed, including a main branch for soil moisture retrieval and three auxiliary branches for vegetation structure. The three auxiliary branches correspond to crop type, crop height, and phenological stage information, respectively. Each auxiliary branch includes a first feature extraction module for processing remote sensing data for the target date and a second feature extraction module for processing diurnal time-series data. The output features of the two feature extraction modules are concatenated to form the high-dimensional feature of the current auxiliary branch. The main branch for soil moisture only inputs multi-source remote sensing data for the target date to extract features related to soil moisture. The high-dimensional features of the three auxiliary branches are concatenated with the features of the main branch and then input into a multilayer perceptron to output the predicted soil moisture content. A two-stage training strategy is used to train the multi-branch neural network model to obtain a trained multi-branch neural network model.

[0007] As a further technical solution, the two-stage training strategy includes: In the first stage, the crop type, crop height and phenological stage information in the ground measured data are used as supervision signals to train the corresponding vegetation structure auxiliary branches and save the optimal parameters. In the second stage, the auxiliary branch parameters obtained from the first stage training are loaded into the model and frozen. The surface soil moisture content in the ground measured data is used as the supervision signal to train the parameters of the soil moisture main branch and the multilayer perceptron.

[0008] As a further technical solution, the inputs of the first feature extraction module include: the VV polarization linearity value, VH polarization linearity value, and incident angle of the synthetic aperture radar data, as well as the multispectral band reflectance of the multispectral optical remote sensing data.

[0009] As a further technical solution, the preprocessing of multispectral optical remote sensing data includes: using time interpolation method on a pixel-by-pixel basis to complete multi-temporal data in order to construct continuous daily-scale multi-band time series data; the starting date for the extraction of the daily-scale time series data is a date that is artificially specified and is at least a preset number of days earlier than the first sampling date.

[0010] According to one aspect of the present invention, a multi-source remote sensing inversion system for farmland soil water based on crop physiological perception is provided, comprising: The data acquisition module is used to acquire synthetic aperture radar data and multispectral optical remote sensing data of the study area and perform preprocessing to obtain target date remote sensing data and daily time series data; The model prediction module is used to input the remote sensing data of the target date and the daily-scale time series data into the trained multi-branch neural network model and output the spatial distribution results of soil moisture. The masking module is used to perform farmland masking by combining land use data to obtain farmland soil moisture products; The trained multi-branch neural network model is pre-trained in the following manner: A training dataset is constructed, which includes ground-measured data covering information on topsoil moisture content, crop height, crop type and phenological stage, as well as target date remote sensing data and daily time-series data obtained by processing synthetic aperture radar data and multispectral optical remote sensing data, respectively. A multi-branch neural network model is constructed, comprising a main branch for soil moisture retrieval and three auxiliary branches for vegetation structure. These three auxiliary branches correspond to crop type, crop height, and phenological stage information, respectively. Each auxiliary branch includes a first feature extraction module for processing remote sensing data for the target date and a second feature extraction module for processing diurnal time-series data. The output features of the two feature extraction modules are concatenated to form the high-dimensional feature of the current auxiliary branch. The main branch for soil moisture is input only with multi-source remote sensing data for the target date and is used to extract features related to soil moisture. The high-dimensional features of the three auxiliary branches are concatenated with the features of the main branch and input into a multilayer perceptron to output a predicted soil moisture value. A two-stage training strategy is used to train the multi-branch neural network model, and the trained multi-branch neural network model is output.

[0011] As a further technical solution, the two-stage training strategy adopted in the model training module includes: In the first stage, the crop type, crop height and phenological stage information in the ground measured data are used as supervision signals to train the corresponding vegetation structure auxiliary branches and save the optimal parameters. In the second stage, the auxiliary branch parameters obtained from the first stage training are loaded into the model and frozen. The surface soil moisture content in the ground measured data is used as the supervision signal to train the parameters of the soil moisture main branch and the multilayer perceptron.

[0012] As a further technical solution, the first feature extraction module of the vegetation structure auxiliary branch is configured to input the VV polarization linearity, VH polarization linearity, and incident angle of synthetic aperture radar data, as well as the multispectral band reflectivity of multispectral optical remote sensing data.

[0013] As a further technical solution, the data acquisition module is specifically used for: performing radiometric calibration, geometric correction, terrain correction, and speckle filtering on synthetic aperture radar data; extracting VV polarization, VH polarization, and incident angle parameters; and converting VV and VH into linear scales; performing atmospheric correction, cloud removal, and spatial resampling on multispectral optical remote sensing data; and using temporal interpolation to complete multi-temporal data pixel by pixel to construct continuous daily-scale multi-band time series data; the starting date for extracting the daily-scale time series data is a date that is artificially specified and is at least a preset number of days earlier than the first sampling date.

[0014] According to one aspect of the present invention, a non-transitory computer-readable storage medium is provided, the non-transitory computer-readable storage medium storing computer instructions that cause the computer to perform the method described.

[0015] According to one aspect of the present invention, an electronic device for retrieving farmland soil water based on multi-source remote sensing of crop physiological perception is provided, comprising: at least one processor; a memory communicatively connected to the at least one processor; the memory storing a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the method described thereon.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) By explicitly modeling vegetation structure information, the information loss problem caused by “de-vegetation” in traditional methods is avoided, and the model can make full use of the structural information contained in vegetation scattering. (2) By introducing physiological parameters such as crop type, plant height and phenological stage as constraints, the physical interpretability and cross-regional generalization ability of the model are improved; (3) By integrating time series data, the dynamic changes in crop growth process are effectively characterized, thereby enhancing the model's adaptability to complex farmland environments; (4) The present invention adopts a multi-branch structure design, which enables different types of information to be learned in independent subspaces, reducing the interference caused by feature coupling and improving the stability and generalization ability of the model; (5) The method of the present invention can realize high-resolution and high-precision soil moisture inversion under dense crop canopy cover at the regional scale, and has good application prospects. It can be widely used in precision agriculture, smart irrigation, water resource management and agricultural ecological monitoring. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating the construction and application of a multi-branch neural network model in a multi-source remote sensing inversion method for farmland soil water based on crop physiological perception, as provided in an embodiment of the present invention. Figure 2 This is a structural diagram of a multi-branch neural network model based on crop physiological perception provided in an embodiment of the present invention; Figure 3 The graph shows the accuracy results of each branch of the model based on the measured data validation set provided in the embodiments of the present invention. Detailed Implementation

[0019] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0021] This invention provides a method for retrieving farmland soil water based on crop physiological perception through multi-source remote sensing. The method first acquires synthetic aperture radar (SAR) data and multispectral optical remote sensing data of the study area and preprocesses them to obtain remote sensing data with consistent target date spatial resolution and daily-scale time-series data. Next, the target date remote sensing data and daily-scale time-series data are input into a trained multi-branch neural network model to output the spatial distribution of soil moisture. This data is then combined with land use data for farmland masking to obtain farmland soil moisture products.

[0022] See Figure 1 As shown, the training and application of a multi-branch neural network model are explained.

[0023] Step 1: Obtain ground measurement data through field sampling This embodiment selects a typical irrigation district as the research object and conducts random sampling on five synthetic aperture radar (SAR) satellite (specifically Sentinel-1 satellite) transit dates: June 3, June 15, July 9, August 2, and August 26, 2025. Sampling points are evenly distributed within the irrigation district, and their layout covers as many different crop types (corn, sunflower, wheat, pepper, and alfalfa, etc.) and their various growth stages as possible to improve the representativeness and diversity of the samples. Larger fields are selected for sampling, and the sampling location is at least 10 meters from the field edge to avoid uncertainties caused by mixed pixels. The sampling experiment is arranged to be completed on the day of satellite transit or within one day before or after, and there are no interference factors such as rainfall or irrigation during the sampling period to ensure the consistency of the observation data with the remote sensing observation conditions. At least 10 samples are collected for each sampling date and each crop type.

[0024] At each sampling point, the topsoil moisture content (SM) in the 0–5 cm layer and the crop height (H) were measured, and the crop type and phenological stage information were recorded. Crop type was identified using numerical coding, and phenological stage was recorded using BBCH coding to reflect the structural differences in different growth stages of the crop. The collected data also included the latitude and longitude coordinates of the sampling points and the sampling date.

[0025] Step 2: Acquire and preprocess remote sensing data Sentinel-2 multispectral optical remote sensing data for the study area from May 20 to August 30, 2025 were acquired. Atmospheric correction, cloud detection, and removal were performed on the data, and spatial resampling was used to ensure its spatial resolution was consistent with Sentinel-1 data (10m in this embodiment). Furthermore, linear interpolation was used pixel-by-pixel to complete the multi-temporal data, constructing a continuous diurnal multi-band time series dataset to mitigate the impact of data loss due to cloud cover and other factors on the model input. The time series data for each band were stored as independent CSV files, with row indices representing station numbers, column indices representing dates, and values ​​representing the reflectance data for the corresponding band.

[0026] Simultaneously, Sentinel-1 synthetic aperture radar data corresponding to the experimental dates were acquired, and radiometric calibration, geometric correction, terrain correction, and speckle filtering were performed on them. VV polarization, VH polarization, and incident angle parameters were extracted, and VV and VH were converted from logarithmic to linear scales for subsequent modeling.

[0027] The interpolated Sentinel-2 data for the experimental day were extracted and integrated with the corresponding Sentinel-1 data (VV linear value, VH linear value, incident angle) and ground-measured data (soil moisture content, plant height, crop type, phenological stage, station number, sampling date) to form a unified data file data2025.csv. Each row corresponds to a sample and includes station number, experiment number, soil moisture content, plant height, crop type, phenological stage information, and VV, VH, incident angle, and multispectral reflectance data for the day.

[0028] For each sample, the corresponding current date is determined according to its experiment number, and all historical data from the preset early stage of crop growth (in this embodiment, it is set to May 20, 2025, at least 10 days earlier than the first sampling date) to that date are extracted from the Sentinel-2 time series as the daily-scale time series data input for that sample, so as to ensure that the model can utilize complete crop growth information.

[0029] Step 3: Construct a multi-branch neural network model based on crop physiological perception The specific structure of the multi-branch neural network model constructed in this embodiment is as follows: (a) Auxiliary branches of vegetation structure The model includes three vegetation structure auxiliary branches, corresponding to crop type, crop height, and phenological stage information, respectively. Each vegetation structure auxiliary branch consists of two sub-feature extraction modules: The first feature extraction module processes the remote sensing data for the current day. Its inputs are: the VV polarization linearity and VH polarization linearity of Sentinel-1, the incident angle, and the multispectral reflectance of Sentinel-2. This module extracts the feature vector for the current day using a multilayer perceptron (MLP, containing two hidden layers, each with 64 neurons).

[0030] The second feature extraction module processes diurnal time-series data. Its input is Sentinel-2 multispectral time-series data (diurnal reflectance values ​​for each band) from the early stages of crop growth to the current observation date. This module learns the dynamic characteristics of vegetation growth through another multilayer perceptron (containing two hidden layers, each with 128 neurons, processing temporal information) and outputs a temporal feature vector, thereby implicitly representing the crop canopy structure, water content, and growth stage.

[0031] The output features of the two sub-modules are then concatenated and fused through a linear layer to form a high-dimensional vegetation structure representation (64 dimensions) for this auxiliary branch.

[0032] (II) Main Branch of Soil Moisture Inversion The main branch only inputs the remote sensing data for the current day (including the VV polarization linearity value, VH polarization linearity value, incident angle, and multispectral reflectance of Sentinel-1), and does not input time-series data to avoid interference from redundant time-series information on the soil signal. This branch extracts feature vectors (64-dimensional) closely related to soil moisture using a multilayer perceptron (containing two hidden layers, each with 64 neurons).

[0033] (III) Feature Fusion and Output The high-dimensional features (64-dimensional each) output from the three vegetation auxiliary branches are concatenated with the features (64-dimensional) output from the main branch through a linear layer to form a high-dimensional fused feature (128-dimensional). This fused feature is then input into the final multilayer perceptron (containing two hidden layers, each with 32 neurons, and an output layer with 1 neuron), which outputs the predicted soil moisture content value through a fully connected network.

[0034] The above process essentially introduces a "soft physics constraint mechanism" into the deep learning framework, that is, by explicitly learning vegetation structure parameters to help decouple the contributions of vegetation and soil in radar backscattering, thereby reducing the uncertainty of inversion.

[0035] Step 4: Two-stage model training See Figure 2 As shown, this embodiment adopts a two-stage training strategy. All input variables are standardized to improve the convergence speed and stability of the model. The mean squared error (MSE) is used as the loss function, and the parameters are updated iteratively through the Adam optimization algorithm.

[0036] Phase 1: Training Vegetation Structure Auxiliary Branches Ground-based measured data were divided into a training set (80%) and a validation set (20%). Measured crop type, crop height, and phenological stage information were used as supervision signals to construct loss functions for the corresponding auxiliary branches. Backpropagation was used to optimize the parameters of each auxiliary branch, enabling the model to learn stable and physically meaningful vegetation structure features from multi-source remote sensing data. An early stopping mechanism was employed during training: training was stopped when the validation set loss did not decrease for 10 consecutive rounds, and the optimal model parameters for each auxiliary branch were saved. After the first stage of training was completed, the optimal model parameters for each auxiliary branch were also saved.

[0037] Phase 2: Training the main branch and the final multilayer perceptron The auxiliary branch parameters saved in the first stage are loaded into the model and frozen (i.e., these parameters are not updated during training). Only measured soil moisture content is used as the supervision signal to train the parameters of the soil moisture main branch and the final multilayer perceptron. Mean squared error is also used as the loss function, and the parameters of the main branch and the final multilayer perceptron are optimized through backpropagation. In this stage, the high-dimensional features output by the frozen vegetation auxiliary branch serve as fixed physical constraints, and the auxiliary main branch separates the soil moisture contribution from the daily remote sensing data, realizing a nonlinear mapping between remote sensing observation data and soil moisture content.

[0038] Figure 3 The results of the validation set for soil water, crop height, crop type, and growth period are shown when using measured data for model training.

[0039] in, Figure 3 (a) in the figure is a scatter plot of the soil moisture content validation set results. The correlation coefficient R between the predicted and measured values ​​is 0.79 and the RMSE is 0.0513, indicating that the model can perform high-precision soil moisture content inversion.

[0040] Figure 3(b) in the figure is a scatter plot of the crop plant height branch validation set results. The crop plant height ranges from 0 to 400 cm. The correlation coefficient R between the predicted and measured values ​​is 0.97, and the RMSE is 26.06 cm, indicating that the plant height estimation accuracy is very high.

[0041] Figure 3 (c) in the matrix represents the confusion matrix of the model's crop type branch validation set results. The coordinate axes, labeled 1-5, correspond to corn, sunflower, wheat, chili pepper, and alfalfa, respectively. The numbers in the matrix represent the number of samples corresponding to the true and predicted combinations; darker colors and larger values ​​indicate a greater number of accurately predicted samples. The results show that the vast majority of samples are concentrated on the diagonal. The overall accuracy of the crop type branch is 0.92. The model performs exceptionally well in predicting corn, sunflower, and alfalfa, while its prediction accuracy is slightly lower for wheat and chili pepper, which have fewer samples, with a small number of misclassifications.

[0042] Figure 3 In the matrix (d), the confusion matrix represents the validation set results of the model's phenological branch. The coordinate axis labels (0-9) represent the BBCH codes of the main stage. The numbers in the matrix indicate the number of samples corresponding to the true and predicted combinations. The darker the color and the larger the value, the more samples are accurately predicted. The results show that most samples are concentrated on the diagonal, with an overall accuracy of 0.79. When the BBCH code is small, the prediction accuracy is high. When the BBCH code is large, the crop vegetation is dense, making it difficult to accurately predict the growth period, and a small number of samples are misclassified to adjacent growth periods.

[0043] Step 5: Model predicts and generates farmland soil and aquatic products. After completing two-stage training, the model was applied to pixel-level prediction of the target area. First, Sentinel-1 and Sentinel-2 data for the target area corresponding to the specified date were downloaded and processed according to the preprocessing procedure described above. This yielded daily Sentinel-1 and Sentinel-2 data with consistent spatial resolution and corresponding temporal phases, as well as Sentinel-2 time-series data within the corresponding time range. Then, the remote sensing data was input pixel-by-pixel into the trained model: a vegetation structure feature vector was generated by an auxiliary branch, which was then concatenated and fused with the soil-sensitive features extracted by the main branch. This fusion was then input into the final multilayer perceptron for forward inference calculations, obtaining the predicted soil moisture content value for each pixel, thereby generating a high spatial resolution soil moisture distribution map of the study area.

[0044] To further enhance the application value of the results, land use type data was introduced to perform farmland masking on the inversion results. The specific steps were as follows: the land use raster data was resampled to the same spatial resolution as the remote sensing data; farmland categories were extracted based on the land use classification results; non-farmland areas such as water bodies, impermeable surfaces, forests, bare land, deserts, and shrubs were masked and removed; and only the soil moisture content prediction results corresponding to farmland areas were retained, thus obtaining a farmland soil moisture spatial distribution product that only includes farmland areas.

[0045] Finally, by comparing and analyzing the inversion results of five experimental dates, it can be verified that the method of the present invention has high inversion accuracy and stability under different crop types and different growth stages, can effectively overcome the interference problem caused by vegetation cover, and realize high-precision remote sensing inversion of farmland soil moisture.

[0046] This embodiment is only a preferred embodiment of the present invention. For those skilled in the art, appropriate adjustments or replacements can be made to the model structure, input variables or training methods without departing from the spirit and substance of the present invention. All such improvements should be included within the protection scope of the present invention.

[0047] Based on the same inventive concept as the foregoing embodiments, this invention also provides a multi-source remote sensing inversion system for farmland soil water based on crop physiological perception, comprising the following modules: Data Acquisition Module: This module acquires synthetic aperture radar (SAR) data, multispectral optical remote sensing data, and ground-measured data (including topsoil moisture content, crop height, crop type, and phenological stage information) for the study area. The acquired data is then transmitted to the preprocessing module.

[0048] The preprocessing module performs radiometric calibration, geometric correction, terrain correction, and speckle filtering on synthetic aperture radar (SAR) data, extracts VV polarization, VH polarization, and incident angle parameters, and converts VV and VH to linear scales. It also performs atmospheric correction, cloud removal, and spatial resampling on multispectral optical remote sensing data to achieve the same spatial resolution as the SAR data. Furthermore, it uses temporal interpolation to complete multi-temporal data pixel by pixel, constructing continuous diurnal multi-band time series data. The starting date for extracting the diurnal time series data is a manually specified date at least 10 days earlier than the first sampling date.

[0049] Model prediction module: This module is used to input remote sensing data and daily time-series data for the target date into the trained model and output the spatial distribution results of soil moisture.

[0050] Masking module: Used to perform farmland masking by combining land use data to obtain farmland soil moisture products.

[0051] The trained multi-branch neural network model is pre-trained in the following manner: A training dataset is constructed, which includes ground-measured data covering information on topsoil moisture content, crop height, crop type and phenological stage, as well as target date remote sensing data and daily time-series data obtained by processing synthetic aperture radar data and multispectral optical remote sensing data, respectively. A multi-branch neural network model is constructed, comprising a main branch for soil moisture retrieval and three auxiliary branches for vegetation structure. These three auxiliary branches correspond to crop type, crop height, and phenological stage information, respectively. Each auxiliary branch includes a first feature extraction module for processing remote sensing data for the target date and a second feature extraction module for processing diurnal time-series data. The output features of the two feature extraction modules are concatenated to form the high-dimensional feature of the current auxiliary branch. The main branch for soil moisture is input only with multi-source remote sensing data for the target date and is used to extract features related to soil moisture. The high-dimensional features of the three auxiliary branches are concatenated with the features of the main branch and input into a multilayer perceptron to output a predicted soil moisture value. A two-stage training strategy is used to train the multi-branch neural network model, and the trained multi-branch neural network model is output.

[0052] The two-stage training strategy employed includes: In the first stage, the crop type, crop height and phenological stage information in the ground measured data are used as supervision signals to train the corresponding vegetation structure auxiliary branches and save the optimal parameters. In the second stage, the auxiliary branch parameters obtained from the first stage training are loaded into the model and frozen. The surface soil moisture content in the ground measured data is used as the supervision signal to train the parameters of the soil moisture main branch and the multilayer perceptron.

[0053] Based on the same inventive concept as the foregoing embodiments, this embodiment of the invention also provides a non-transitory computer-readable storage medium storing computer instructions thereon. When these computer instructions are executed by a processor, they can implement the methods described in the foregoing embodiments. This storage medium may be a ROM, RAM, magnetic disk, optical disk, USB flash drive, or solid-state drive, etc.

[0054] Based on the same inventive concept as the foregoing embodiments, this invention also provides an electronic device for retrieving farmland soil water using multi-source remote sensing based on crop physiological perception, including at least one processor and a memory communicatively connected to the at least one processor. The memory stores a computer program executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the methods described in the foregoing embodiments. This electronic device can be a desktop computer, server, embedded device, or mobile terminal, etc.

[0055] In summary, this invention discloses a method for retrieving farmland soil water based on multi-source remote sensing of crop physiological perception. Specifically, it includes the following steps: acquiring ground-based measured data through field sampling, including soil moisture content, crop height, crop type, phenological stage information, and the latitude, longitude, and date of sampling points; acquiring Sentinel-1 radar data and Sentinel-2 optical remote sensing data of the study area and preprocessing them to obtain target date remote sensing data and Sentinel-2 diurnal time-series data with consistent spatial resolution; constructing a multi-branch neural network model, including a main branch for soil moisture retrieval and multiple auxiliary branches for vegetation structure; training the model using a two-stage training strategy; using the trained model to perform pixel-level prediction of the target area to generate spatial distribution results of soil moisture; and combining land use data for farmland masking processing to obtain farmland soil moisture products. Compared with traditional methods, this invention constrains the remote sensing inversion process by introducing vegetation physiological parameters such as crop type, plant height, and phenological stage, explicitly characterizing the vegetation structure features and their dynamic changes. This effectively reduces the interference of vegetation cover on microwave signals and improves the accuracy and stability of soil moisture inversion. Simultaneously, this method can achieve high-precision, high-spatiotemporal-resolution farmland soil moisture inversion results at the regional scale, enhancing the model's applicability and generalization ability under different crop types and growth stages. This provides reliable data support for crop growth status assessment, farmland moisture dynamic monitoring, and precision agricultural management.

[0056] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for retrieving farmland soil water based on multi-source remote sensing of crop physiological perception, characterized in that, include: Synthetic aperture radar data and multispectral optical remote sensing data of the study area were acquired and preprocessed to obtain target date remote sensing data and daily time series data. The target date remote sensing data and daily-scale time series data are input into the trained multi-branch neural network model, which outputs the spatial distribution results of soil moisture. Combined with land use data, farmland masking is performed to obtain farmland soil moisture products. The training of the multi-branch neural network model further includes: The training dataset was constructed, including ground-measured data covering information on topsoil moisture content, crop height, crop type and phenological stage, as well as target date remote sensing data and daily time-series data obtained by processing synthetic aperture radar data and multispectral optical remote sensing data, respectively. A multi-branch neural network model is constructed, including a main branch for soil moisture retrieval and three auxiliary branches for vegetation structure. The three auxiliary branches correspond to crop type, crop height, and phenological stage information, respectively. Each auxiliary branch includes a first feature extraction module for processing remote sensing data for the target date and a second feature extraction module for processing diurnal time-series data. The output features of the two feature extraction modules are concatenated to form the high-dimensional feature of the current auxiliary branch. The main branch for soil moisture retrieval only inputs multi-source remote sensing data for the target date to extract features related to soil moisture. The high-dimensional features of the three auxiliary branches are concatenated with the features of the main branch and then input into a multilayer perceptron to output the predicted soil moisture content. A two-stage training strategy is adopted to train the multi-branch neural network model to obtain a trained multi-branch neural network model. The two-stage training strategy includes: in the first stage, using only the crop type, crop height and phenological stage information in the ground measured data as supervision signals, the corresponding vegetation structure auxiliary branches are trained and the optimal parameters are saved; in the second stage, the auxiliary branch parameters obtained in the first stage are loaded into the model and frozen, and the parameters of the soil moisture inversion main branch and the multilayer perceptron are trained using only the surface soil moisture content in the ground measured data as supervision signals.

2. The method for retrieving farmland soil water based on crop physiological perception using multi-source remote sensing according to claim 1, characterized in that, The inputs to the first feature extraction module include: the VV polarization linearity, VH polarization linearity, and incident angle of the synthetic aperture radar data, as well as the multispectral band reflectance of the multispectral optical remote sensing data.

3. The method for retrieving farmland soil water based on crop physiological perception using multi-source remote sensing according to claim 1, characterized in that, The multispectral optical remote sensing data is Sentinel-2 satellite data, and the multispectral reflectance includes the reflectance of bands B1 to B12. The preprocessing of multispectral optical remote sensing data includes: using time interpolation to complete multi-temporal data pixel by pixel in order to construct continuous diurnal multi-band time series data; The starting date for extracting the daily-scale time-series data is a date that is manually specified and is at least a preset number of days earlier than the first sampling date.

4. A multi-source remote sensing inversion system for farmland soil water based on crop physiological perception, characterized in that, include: The data acquisition module is used to acquire synthetic aperture radar data and multispectral optical remote sensing data of the study area and perform preprocessing to obtain target date remote sensing data and daily time series data; The model prediction module is used to input the remote sensing data of the target date and the daily-scale time series data into the trained multi-branch neural network model and output the spatial distribution results of soil moisture. The masking module is used to perform farmland masking by combining land use data to obtain farmland soil moisture products; The trained multi-branch neural network model is pre-trained in the following manner: A training dataset is constructed, which includes ground-measured data covering information on topsoil moisture content, crop height, crop type and phenological stage, as well as target date remote sensing data and daily time-series data obtained by processing synthetic aperture radar data and multispectral optical remote sensing data, respectively. A multi-branch neural network model is constructed, comprising a main branch for soil moisture retrieval and three auxiliary branches for vegetation structure. These three auxiliary branches correspond to crop type, crop height, and phenological stage information, respectively. Each auxiliary branch includes a first feature extraction module for processing remote sensing data for the target date and a second feature extraction module for processing diurnal time-series data. The output features of the two feature extraction modules are concatenated to form the high-dimensional feature of the current auxiliary branch. The main branch for soil moisture retrieval only inputs multi-source remote sensing data for the target date to extract features related to soil moisture. The high-dimensional features of the three auxiliary branches are concatenated with the features of the main branch and input into a multilayer perceptron to output a predicted soil moisture value. A two-stage training strategy is adopted to train the multi-branch neural network model, and the trained multi-branch neural network model is output. The two-stage training strategy includes: in the first stage, the crop type, crop height and phenological stage information in the ground measured data are used as supervision signals to train the corresponding vegetation structure auxiliary branches and save the optimal parameters; in the second stage, the auxiliary branch parameters obtained in the first stage are loaded into the model and frozen, and the parameters of the soil moisture inversion main branch and the multilayer perceptron are trained using only the surface soil moisture content in the ground measured data as supervision signals.

5. The multi-source remote sensing inversion farmland soil water system based on crop physiological perception according to claim 4, characterized in that, The first feature extraction module of the vegetation structure auxiliary branch is configured to input the VV polarization linearity, VH polarization linearity, and incident angle of synthetic aperture radar data, as well as the multispectral band reflectance of multispectral optical remote sensing data.

6. The multi-source remote sensing inversion farmland soil water system based on crop physiological perception according to claim 4, characterized in that, The multispectral optical remote sensing data is Sentinel-2 satellite data, and the multispectral band reflectance includes the reflectance of bands B1 to B12. The data acquisition module is specifically used for: performing radiometric calibration, geometric correction, terrain correction, and speckle filtering on the synthetic aperture radar data; extracting VV polarization, VH polarization, and incident angle parameters; and converting VV and VH into linear values. The module also performs atmospheric correction, cloud removal, and spatial resampling on the multispectral optical remote sensing data, and uses temporal interpolation to complete the multi-temporal data pixel by pixel to construct continuous diurnal multi-band time series data. The starting date for the diurnal time series data is a manually specified date that is at least a preset number of days earlier than the first sampling date.

7. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that cause the computer to perform the method described in any one of claims 1 to 3.

8. An electronic device for retrieving farmland soil water based on multi-source remote sensing of crop physiological perception, characterized in that, include: At least one processor; A memory that is communicatively connected to the at least one processor; The memory stores a computer program that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 3.

Images