A farmland soil moisture remote sensing monitoring method, device, equipment and medium

Abstract

The application discloses a farmland soil moisture remote sensing monitoring method, device, equipment and medium, effectively solves the problem that the prior art lacks soil moisture accurate monitoring capable of deeply fusing remote sensing information and plant water transmission physical mechanism. The method provided by the application comprises the following steps: collecting multi-modal data, and constructing a water-sensitive crown layer comprehensive feature index based on the multi-modal data; within a preset time window, a crown layer-root water transmission model is constructed based on thermal infrared remote sensing images and meteorological data to quantize the contribution rate of each root layer soil to root water absorption; a standardized multi-task machine learning model is trained to obtain an inversion model of soil moisture of different soil layers; the trained inversion model is used to invert soil moisture at different depths of a crop root zone, and a soil moisture spatiotemporal distribution map is generated, so that farmland soil moisture remote sensing monitoring is completed based on the soil moisture spatiotemporal distribution map.

Description

Technical Field

[0001] This invention relates to the field of remote sensing monitoring technology, and in particular to a method, device, equipment and medium for remote sensing monitoring of farmland soil moisture. Background Technology

[0002] With the development of smart agriculture, remote sensing technology has become an important means of large-scale, non-contact monitoring of farmland soil moisture. However, current mainstream soil moisture remote sensing inversion methods still have the following significant limitations:

[0003] 1. Shallow monitoring depth makes it difficult to reflect the true water stress of crops: Existing remote sensing methods (such as optical, thermal infrared, and microwave remote sensing) mainly perceive information from the surface or shallow (0-5 cm) soil or vegetation canopy, while crop water absorption mainly occurs in the root layer (usually up to 60 cm or even deeper). Shallow water information often has a weak correlation with root layer water status, making it difficult to accurately diagnose physiological drought in crops, especially when deep soil is water-deficient while the surface layer is moist, which can easily lead to misjudgment.

[0004] 2. Lack of constraints on the physical mechanisms of root water uptake: Most inversion models based on statistics or machine learning only establish empirical relationships between remote sensing features and shallow soil moisture. These models are "black boxes" and fail to incorporate the physical mechanisms of water transport in the soil-root-canopy continuum (SPAC). This results in poor model interpretability, weak generalization ability, and insufficient applicability and stability across different crop types, growth stages, climatic regions, or soil conditions.

[0005] 3. Multimodal data fusion fails to effectively link aboveground and underground processes: Although existing studies have attempted to fuse multi-source remote sensing data such as multispectral and thermal infrared data, the fusion often remains at the data or feature level, failing to establish a causal relationship between canopy response signals (such as spectral indices and canopy temperature) and underground root zone water distribution and root water absorption dynamics through effective physical models. This results in inversion results that cannot distinguish the contribution of different soil depths to crop water supply, making it difficult to meet the needs of precise irrigation based on soil layers.

[0006] 4. There is a lag and confusion in the indication of crop water stress: the response of canopy indicators (such as vegetation index) to water stress is delayed, and canopy temperature is easily affected by meteorological conditions (such as solar radiation, wind speed and humidity). A single indicator can easily misjudge environmental stress as water stress, reducing the accuracy and timeliness of monitoring.

[0007] Therefore, there is an urgent need for an inversion method that can deeply integrate remote sensing information with the physical mechanisms of plant water transport to achieve accurate and dynamic monitoring of soil moisture in the crop root zone (especially in stratified areas), in order to overcome the above-mentioned defects and provide a reliable basis for true precision irrigation. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the present invention aims to provide a method, device, equipment, and medium for remote sensing monitoring of farmland soil moisture, effectively solving the problem of the lack of accurate soil moisture monitoring in existing technologies that can deeply integrate remote sensing information with the physical mechanisms of plant water transport.

[0009] The technical solution provided in this application is: a remote sensing monitoring method for farmland soil moisture, the method comprising:

[0010] Multimodal data was obtained by simultaneously collecting farmland ground data using multispectral and thermal infrared remote sensing from UAVs. Based on the multispectral and thermal infrared remote sensing images in the multimodal data, a moisture-sensitive comprehensive canopy characteristic index was constructed. The multimodal data also included meteorological data.

[0011] Within a preset time window, a canopy-root water transport model is constructed based on thermal infrared remote sensing images and meteorological data to quantify the contribution rate of each root layer soil to root water absorption.

[0012] Using the comprehensive canopy feature index and the contribution rate as physical constraints, a standardized multi-task machine learning model is trained to obtain an inversion model of soil moisture for different soil layers.

[0013] Using a trained inversion model, soil moisture at different depths in the crop root zone is inverted, and a spatiotemporal distribution map of soil moisture is generated, so as to complete remote sensing monitoring of farmland soil moisture based on the spatiotemporal distribution map of soil moisture.

[0014] Furthermore, a canopy-root water transport model was constructed based on thermal infrared remote sensing imagery and meteorological data to quantify the contribution rate of soil in each root layer to root water uptake, including:

[0015] The soil moisture content of different root layers in a continuous time series was calculated using the central difference method to obtain the root water absorption rate of each root layer.

[0016] Based on the root water absorption rate of each root layer and meteorological data, a canopy-root water transport model was constructed to determine the contribution rate of each root layer soil to root water absorption.

[0017] Furthermore, the canopy-root water transport model is constructed based on the root water absorption rate and meteorological data of each root layer to determine the contribution rate of each root layer soil to root water absorption, including:

[0018] A canopy water balance network is constructed based on the root water absorption rate of each root layer to process the root water absorption rate of each root layer and meteorological data to obtain the simulated change in canopy water volume data.

[0019] By minimizing the error between the simulated and actual changes, a set of root hydraulic characteristic parameters are obtained to determine the contribution rate.

[0020] Furthermore, the canopy integrated feature index and the contribution rate are used as physical constraints to train a standardized multi-task machine learning model, including:

[0021] A training dataset was constructed based on the comprehensive characteristic index of the canopy and the actual measured values ​​of soil moisture in each soil layer.

[0022] Using the training dataset, the multi-task machine learning model is iteratively trained with the objective of minimizing the composite loss function of the multi-task machine learning model.

[0023] Furthermore, the objective is to minimize the composite loss function of the multi-task machine learning model, including:

[0024] Data fitting terms and physical constraint terms based on contribution rates were constructed to measure the total error between the predicted and actual soil moisture values ​​for each soil layer.

[0025] By combining the data fitting term and the physical constraint term, the composite loss function of the multi-task machine learning model is obtained.

[0026] Furthermore, based on the multispectral remote sensing image and the thermal infrared remote sensing image, a moisture-sensitive comprehensive canopy characteristic index is constructed, including:

[0027] At least one water-sensitive spectral index, vegetation index, and canopy temperature are extracted from the multispectral remote sensing image and the thermal infrared remote sensing image, respectively.

[0028] A moisture-sensitive thermal infrared index is constructed by processing the canopy temperature and vegetation index using a wet-dry edge fitting method, and then fused with the spectral index to obtain a comprehensive canopy characteristic index.

[0029] Furthermore, using the trained inversion model, soil moisture at different depths in the crop root zone is retrieved, and a spatiotemporal distribution map of soil moisture is generated, including:

[0030] Using the inversion model, the predicted soil moisture values ​​at different depths corresponding to each spatial location point within the farmland area are output simultaneously.

[0031] Based on the predicted soil moisture values ​​at different spatial locations, a spatial distribution map of soil moisture in the farmland area is generated.

[0032] This application provides another solution: a remote sensing monitoring device for farmland soil moisture, the device comprising:

[0033] The data acquisition module is used to simultaneously acquire farmland ground data using UAV multispectral and thermal infrared remote sensing to obtain multimodal data, and to construct a moisture-sensitive comprehensive canopy characteristic index based on the multispectral and thermal infrared remote sensing images in the multimodal data; the multimodal data also includes meteorological data;

[0034] The quantification module is used to construct a canopy-root water transport model based on thermal infrared remote sensing images and meteorological data within a preset time window, so as to quantify the contribution rate of each root layer soil to root water absorption.

[0035] The training module is used to train a standardized multi-task machine learning model by using the comprehensive feature index of the canopy and the contribution rate as physical constraints, so as to obtain the soil moisture inversion model for different soil layers.

[0036] The monitoring module is used to invert soil moisture at different depths in the crop root zone using a trained inversion model and generate a spatiotemporal distribution map of soil moisture, so as to complete remote sensing monitoring of farmland soil moisture based on the spatiotemporal distribution map of soil moisture.

[0037] This application also provides a solution: an electronic device, comprising: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the memory via the bus, and when the machine-readable instructions are executed by the processor, the steps of any one of the methods for remote sensing monitoring of farmland soil moisture are performed.

[0038] This application also provides another solution: a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of any one of the methods for remote sensing monitoring of farmland soil moisture.

[0039] This application provides a method for remote sensing monitoring of farmland soil moisture. The method first acquires multimodal data of farmland ground using simultaneous multispectral and thermal infrared remote sensing data from an unmanned aerial vehicle (UAV). Based on the multispectral and thermal infrared remote sensing images in the multimodal data, a moisture-sensitive canopy comprehensive characteristic index is constructed. The multimodal data also includes meteorological data. Second, within a preset time window, a canopy-root water transport model is constructed based on the thermal infrared remote sensing images and meteorological data to quantify the contribution rate of each root layer soil to root water absorption. Then, the canopy comprehensive characteristic index and the contribution rate are used as physical constraints to train a standardized multi-task machine learning model to obtain soil moisture inversion models for different soil layers. Finally, the trained inversion model is used to invert soil moisture at different depths in the crop root zone and generate a spatiotemporal distribution map of soil moisture, thereby completing remote sensing monitoring of farmland soil moisture based on the spatiotemporal distribution map. Based on the above methods, this application achieves precise stratified inversion of soil moisture in the crop root zone, breaking through the limitation of traditional remote sensing that can only monitor surface moisture. It directly obtains effective water information in the root zone that is closely related to crop water uptake, providing unprecedented data support for on-demand, stratified irrigation. The canopy-root water transport model ensures that the inversion results are consistent with the physiological mechanism of plant water transport, significantly improving the model's interpretability, physical consistency, and mechanism generalization ability, and enhancing the physical basis and interpretability of the inversion model. Based on the comprehensive canopy characteristic index, it can effectively distinguish between real water stress caused by soil water shortage and apparent stress caused by meteorological conditions such as high temperature, dryness, and strong winds, significantly reducing the misjudgment rate, improving the accuracy of drought identification, and enhancing monitoring precision and anti-interference ability. The use of soil moisture spatiotemporal distribution maps also enhances the universality and operational application potential of the method. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating a remote sensing monitoring method for farmland soil moisture provided in an embodiment of this application.

[0041] Figure 2 This is a schematic diagram illustrating the framework of a remote sensing monitoring method for farmland soil moisture provided in an embodiment of this application.

[0042] Figure 3 This is a schematic diagram illustrating the construction of dry and wet edge information for the thermal infrared index provided in an embodiment of this application.

[0043] Figure 4 This is a structural block diagram of a farmland soil moisture remote sensing monitoring device provided in an embodiment of this application.

[0044] Figure 5 A structural block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0045] For the purposes of this invention, the foregoing and other technical contents, features and effects are described in conjunction with the appendix below. Figures 1-5 The detailed description of the embodiments will make this clear. All structural details mentioned in the following embodiments are based on the accompanying drawings.

[0046] Current remote sensing monitoring of farmland soil moisture largely relies on satellites or drones to acquire surface reflectance or temperature information, and the inversion targets are usually limited to the surface or shallow (0-5 cm) soil moisture. However, actual crop water absorption mainly occurs in the deep root distribution area, and shallow moisture information often fails to reflect the actual water stress state of crops, leading to inaccurate measurements and imprecise irrigation. Existing methods are mostly purely data-driven models, lacking the integration of the physical mechanisms of water transport in the "soil-root-canopy" system. These models have poor interpretability and weak generalization ability, making it difficult to accurately identify crop water deficit by soil layer, thus hindering the precise implementation of smart irrigation.

[0047] Based on this, the present application provides a method, apparatus, equipment and medium for remote sensing monitoring of farmland soil moisture, which will be described below with reference to the embodiments and accompanying drawings.

[0048] Example 1

[0049] To facilitate understanding of this embodiment, a remote sensing method for monitoring farmland soil moisture disclosed in this application will first be described in detail. For example... Figure 1 The diagram shows a flowchart of a remote sensing monitoring method for farmland soil moisture. Figure 2 The diagram shows a framework of a remote sensing monitoring method for farmland soil moisture. The method provided in this application includes:

[0050] S101. Farmland ground is simultaneously collected by UAV multispectral and thermal infrared remote sensing to obtain multimodal data, and a moisture-sensitive canopy comprehensive characteristic index is constructed based on the multispectral remote sensing images and thermal infrared remote sensing images in the multimodal data; the multimodal data also includes meteorological data;

[0051] S102. Within a preset time window, construct a canopy-root water transport model based on thermal infrared remote sensing images and meteorological data to quantify the contribution rate of each root layer soil to root water absorption.

[0052] S103. Using the comprehensive characteristic index of the canopy and the contribution rate as physical constraints, train a standardized multi-task machine learning model to obtain an inversion model of soil moisture for different soil layers.

[0053] S104. Using the trained inversion model, invert the soil moisture at different depths in the crop root zone and generate a spatiotemporal distribution map of soil moisture, so as to complete the remote sensing monitoring of farmland soil moisture based on the spatiotemporal distribution map of soil moisture.

[0054] In step S101, this application simultaneously equips a hyperspectral imager (spectral range covering 400-1700 nm, including visible, near-infrared, and short-wave infrared bands) and a thermal infrared imager (spectral response range typically 8-14 μm) on a UAV flight platform. The hyperspectral imager and the thermal infrared imager are controlled by a hardware synchronization trigger device or a high-precision time synchronization module to ensure that when the UAV flies over the same plot of land, it can acquire multispectral remote sensing images (HS) and thermal infrared remote sensing images (TIR) ​​of the ground features almost simultaneously. The multispectral remote sensing images are used to extract moisture-sensitive spectral information, and the thermal infrared remote sensing images are used to represent canopy temperature. Simultaneously, soil moisture data is also acquired. Ground data, including water (soil moisture content in each root layer) and meteorological data (temperature, humidity, radiation, etc.), are collected synchronously by ground sensors and weather stations. Flight operations are conducted during periods of stable weather, ample sunshine, and low wind (typically within two hours before or after local noon) to minimize the impact of changes in light and atmospheric conditions. Before and after the flight operations, the hyperspectral and thermal infrared spectrometers are calibrated for radiometric and temperature data, respectively, to ensure the physical accuracy of the data and comparability between data from different periods. The multispectral remote sensing image is generated by radiometrically calibrating, atmospherically correcting, geometrically correcting, and orthorectifying the original image, producing a multispectral remote sensing image with accurate geographic coordinates and reflectance information. The thermal infrared remote sensing image is generated by radiometrically calibrating and atmospherically correcting the original thermal infrared data, and then... The received radiation values ​​are converted into true surface temperature. Subsequently, image registration technology is used to spatially align thermal infrared images and multispectral remote sensing images with high precision. Spatiotemporal reference and interpolation methods are also used to address the spatiotemporal scale differences of multimodal data, including multispectral remote sensing images (HS), thermal infrared remote sensing images (TIR), and ground data containing soil moisture and meteorological data. Based on the above-mentioned synchronously processed multispectral and thermal infrared remote sensing images, feature indicators sensitive to crop moisture status are extracted and fused from the spectral and temperature dimensions, respectively, to construct a comprehensive canopy moisture indicator vector. This comprehensive feature construction method, which combines spectral and temperature dimensions and multiple indicators, can more comprehensively and robustly capture early and weak physiological and physical response signals of crop canopy caused by water deficit, laying a reliable data foundation for subsequent accurate inversion of deep root zone soil moisture.

[0055] In a specific implementation of step S101, one embodiment is as follows: A moisture-sensitive comprehensive canopy characteristic index is constructed based on the multispectral remote sensing image and the thermal infrared remote sensing image, including:

[0056] S1011. Extract at least one water-sensitive spectral index, vegetation index, and canopy temperature from the multispectral remote sensing image and the thermal infrared remote sensing image, respectively.

[0057] S1012. The canopy temperature and vegetation index are processed by the dry-wet edge fitting method to construct a moisture-sensitive thermal infrared index, which is then fused with the spectral index to obtain a comprehensive canopy characteristic index.

[0058] In steps S1011-S1012, this application selects specific band combinations sensitive to leaf water content and canopy water status from multispectral remote sensing images, and calculates a series of recognized water-sensitive spectral indices, mainly including: Normalized Differential Water Index (NDWI), which uses near-infrared (NIR) and short-wave infrared (SWIR) bands to be sensitive to canopy liquid water content; Water Stress Index (MSI), which uses the ratio of short-wave infrared (SWIR) to near-infrared (NIR) bands, and an increase in the ratio usually indicates water stress; Normalized Differential Infrared Index (NDII), which is another index using NIR and SWIR, and is related to the equivalent water thickness of leaves; Near-infrared Water Stress Index (NWSI): calculated based on the ratio or normalization of reflectance of specific near-infrared bands. By calculating these indices, a series of spectral feature layers reflecting canopy water status from different angles are obtained. The above water-sensitive spectral indices are expressed based on formulas (1)-(4):

[0059] (1)

[0060] (2)

[0061] (3)

[0062] (4)

[0063] In the formula, NWSI, NDII, MSI, and NDWI are moisture-sensitive spectral indices, B1 and B2 are reflectance in the 400-1000nm band, and SB1 is shortwave infrared fluctuating reflectance. This application also obtains canopy temperature data from thermal infrared remote sensing images. The canopy temperature data is fitted with dry / wet edge information using NDVI, which is used to delineate canopy growth gradients. Figure 3 As shown, the crop moisture-sensitive thermal infrared index (TVDI) was constructed using formulas (5)-(7):

[0064] (5)

[0065] (6)

[0066] (7)

[0067] (5) In the formula, TVDI is the moisture-sensitive thermal infrared index; T s T represents the canopy temperature. min T represents the wet edge temperature of the canopy. max Canopy dry edge temperature, NDVI is normalized differential vegetation index, a wet ,b wet ,a dry ,b dry To obtain fitting coefficients, this application spatially overlays and feature-level fuses one or more water-sensitive spectral index layers (such as NDWI, MSI, etc.) with a Temperature Vegetation Drought Index (TVDI) layer to obtain a multi-dimensional feature vector corresponding to each spatial pixel location, i.e., the comprehensive canopy feature index. Different NDVI values ​​correspond to different canopy temperature baselines for wet edges (sufficient water) and dry edges (scarce water). Only through NDVI calibration can the water stress assessment of TVDI be more accurate, thus making the comprehensive canopy feature index, which integrates TVDI and water-sensitive spectral indices, more closely reflect the actual crop water conditions. This comprehensive canopy feature index serves as input to subsequent machine learning models. It simultaneously contains information reflecting canopy structure water content (from spectral indices) and canopy energy balance / transpiration state (from TVDI), forming a comprehensive canopy feature index with high sensitivity and high information content regarding crop water stress.

[0068] In step S102, within a preset time window, this application constructs a canopy-root water transport model based on thermal infrared remote sensing images and meteorological data using the water balance method. The quantitative relationship between soil volumetric water content θ, root water absorption rate R, canopy water content W, and crop transpiration Tp is clarified through the canopy-root water transport model, as shown in formula (8):

[0069] (8);

[0070] Where θ is the soil volumetric water content, q is the water flux, and R is the root water absorption rate. Formula (8) represents the change in soil volumetric water content θi of a certain root layer i within a preset time window. It is determined by the vertical change in soil water flux and the root water absorption rate R. That is, the decrease in water in the root zone is equal to the amount of water absorbed by the root system. The root water absorption rate R is shown by formula (9):

[0071] (9);

[0072] By analyzing the dynamic water balance among soil, roots, and canopy within a time window, the contribution rate of root water uptake was determined. The quantization is transformed into a least-squares problem with physical constraints; the core of the entire process is the unification of spatiotemporal scales within the time window and the strict integration of physical constraints, ensuring... The quantitative results not only conform to the model fitting rules, but also match the actual physical and physiological characteristics of crop root water absorption, ultimately providing reliable physical constraint parameters for subsequent soil moisture inversion.

[0073] In the specific implementation of step S102, one embodiment is as follows: A canopy-root water transport model is constructed based on thermal infrared remote sensing images and meteorological data to quantify the contribution rate of each root layer soil to root water absorption, including:

[0074] S1021. The soil moisture content of different root layers in a continuous time series is calculated using the central difference method to obtain the root water absorption rate of each root layer.

[0075] S1022. Based on the root water absorption rate of each root layer and meteorological data, construct a canopy-root water transport model to determine the contribution rate of each root layer soil to root water absorption.

[0076] In steps S1021-S1022, the root water absorption rate is the amount of water absorbed by the roots per unit time and unit volume of soil in a certain root layer, and it is the core indicator characterizing the water absorption capacity of each root layer. Based on the principle of water balance, this method uses the central difference method to calculate the soil volumetric water content time series data of each root layer separately, so as to achieve accurate quantification of the root water absorption rate at the root layer scale. Based on the root water absorption rate of each root layer and meteorological data, a canopy-root water transport model is constructed, specifically formula (8), and the contribution rate of each root layer soil to root water absorption is determined by formulas (9)-(10):

[0077] (10);

[0078] (11);

[0079] In the above two formulas, W represents the canopy water content. This refers to crop transpiration. The depth of the root layer. ...... This refers to the contribution rate of soil in each root layer to root water absorption, and it is also a root hydraulic characteristic parameter. It characterizes the weight of the contribution of root water absorption to soil moisture in a certain root layer. The contribution rate is also verified in three dimensions: model fit, parameter physical consistency, and physiological rationality. Once the verification is passed, it can be used as the contribution rate. By analyzing the distribution of root water absorption at different growth stages, the root hydraulic characteristic phenotypic parameters are quantified. and through This indicates that soil moisture content provides physical constraints for inversion, enabling precise soil moisture inversion through root-shoot coordination, and providing targeted guidance for precision irrigation. Root zone irrigation should be prioritized for areas with a high proportion of root tissue.

[0080] In a specific implementation of step S1022, one embodiment is as follows: The canopy-root water transport model is constructed based on the root water absorption rate of each root layer and meteorological data to determine the contribution rate of each root layer soil to root water absorption, including:

[0081] S10221. Based on the root water absorption rate of each root layer, a canopy water balance network is constructed to process the root water absorption rate of each root layer and meteorological data to obtain the simulated change in canopy water volume data.

[0082] S10222. By minimizing the error between the simulated change and the actual change, a set of root hydraulic characteristic parameters are solved to obtain the contribution rate.

[0083] In steps S10221-S10222, the canopy water balance network is a quantitative relationship network connecting root water absorption, meteorological transpiration, and canopy water volume changes, with water conservation as its physical core. Essentially, it transforms the processes of root water absorption, canopy water replenishment, canopy water consumption, and canopy water volume changes into mathematical equations.

[0084] (12);

[0085] in, The total amount of water replenished to the canopy by water absorption from each root layer; The simulated rate of change of canopy water volume; The simulated value of canopy water volume output by the network. For time step, To calculate transpiration, the simulated change in canopy water volume is obtained by processing the root water absorption rate of each root layer and meteorological data using the above formula. This application provides a constraint framework for subsequent parameter solutions. To solve for the optimal contribution rate, an error objective function needs to be constructed, with the goal of minimizing the total error between the simulated and actual changes. The solution for the contribution rate is transformed into a mathematical optimization problem, including minimizing the mean squared error (MSE) and... and Physical constraints ≥ 0; the above constrained minimization problem is solved using a numerical optimization algorithm to obtain the solution that minimizes the error. It still needs to be verified. The distribution of these elements matches the growth patterns of crop roots, such as the shallow layer during the seedling stage. High proportion, deep grouting period High proportion; if mismatched, adjust the initial value based on measured root biomass data. The final contribution rate of root zone water absorption can be obtained by resolving the problem.

[0086] In step S103, after obtaining the contribution rate, this application uses the canopy comprehensive feature index and the contribution rate as physical constraints to simultaneously acquire measured values ​​of soil volumetric water content (θ) at different soil depths. This is typically achieved using soil moisture sensors (such as a time-domain reflectometry (TDR) or frequency-domain reflectometry (FDR)) at ground measurement points simultaneously or nearly simultaneously with the UAV flight. Soil layers are defined, for example, as 0-10 cm, 10-20 cm, 20-40 cm, and 40-60 cm. A standardized multi-task machine learning model is trained. This standardized multi-task machine learning model uses a multi-task learning framework suitable for regression tasks, such as a multi-output regression model based on random forest (RF), gradient boosting decision tree (XGBoost / LightGBM), or deep neural network (DNN). The model has a shared feature input layer (receiving the canopy comprehensive feature vector X) and multiple independent output nodes in the output layer, each node corresponding to a predicted soil moisture value for a specific soil layer. This involves deeply integrating comprehensive canopy characteristics as data input and root water absorption contribution rate as physical constraints into the training process of a standardized multi-task machine learning model. This process trains inversion models of soil moisture in different soil layers. The introduction of physical constraints provides additional and powerful prior knowledge, helping the model learn more fundamental laws even with limited or noisy training data. As a result, the model exhibits stronger generalization ability and robustness when facing new environments, new crop varieties, or different management measures. The final inversion model can simultaneously invert soil moisture in different soil layers, and the prediction results conform to physical laws, providing reliable data support for precision irrigation.

[0087] In a specific implementation of step S103, one embodiment involves using the canopy comprehensive feature index and the contribution rate as physical constraints to train a standardized multi-task machine learning model, including:

[0088] S1031. A training dataset is constructed based on the comprehensive characteristic index of the canopy and the actual measured values ​​of soil moisture in each soil layer.

[0089] S1032. Using the training dataset, the multi-task machine learning model is iteratively trained with the goal of minimizing the composite loss function of the multi-task machine learning model.

[0090] In steps S1031-S1032, this application constructs a multi-dimensional feature vector based on the comprehensive canopy feature index and standardizes each feature dimension in the comprehensive canopy feature index, such as NDWI, MSI, TVDI, etc., using Z-score to eliminate the influence of differences in feature dimensions and numerical ranges on model training. At the same time, the measured soil moisture values ​​at different soil depths are normalized by Min-Max or standardized by Z-score to map them to similar numerical intervals (such as [0,1] or standard normal distribution) to balance the gradient scale of each output task and avoid the problem of task learning imbalance caused by the small value of deep soil moisture. Thus, a standardized complete dataset is obtained. The standardized complete dataset is randomly divided into a training set and an independent test set in a 7:3 ratio, or spatiotemporal stratified sampling is used to ensure that samples from different fields and different growth stages are representative in both the training and test sets. To enhance the robustness of the model, data augmentation can be performed on the training set, such as by adding Gaussian noise or performing small-scale random linear combinations of features, to simulate noise and variation that may exist in remote sensing observations. The training process, which includes training initialization, forward propagation and loss calculation, back propagation and parameter update, as well as iterative loops and convergence judgment, aims to minimize the composite loss function. The model parameters are updated iteratively, as shown in formula (13).

[0091] (13);

[0092] In the formula, Let be the soil moisture in the i-th crop root zone; X be the set of input variables consisting of multi-band spectral indices, canopy-temperature difference, etc.; f(⋅) be the machine learning regression function; The contribution rate of soil moisture in the crop root zone to root water uptake in that layer is calculated. The training process not only minimizes prediction error but also explicitly optimizes an interpretable physical consistency metric (weighted sum approximation target), ensuring the final model's output has clear physical meaning and higher reliability. After training, the final model is evaluated on a reserved independent test set. Key evaluation metrics include: root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) for each soil layer. Simultaneously, the contribution of physical constraints during training can be analyzed to verify the weighted sum of predictions. Whether the correlation with the target value f(X) is significantly improved, the trained model parameters, the mean and standard deviation used for data standardization, and the value table of physical constraint parameters are saved. The model then becomes a root zone soil moisture stratification inversion engine that can be directly deployed.

[0093] In a specific implementation of step S1032, one embodiment is as follows: minimizing the composite loss function of the multi-task machine learning model as the objective includes:

[0094] S10321. Construct data fitting terms and physical constraint terms based on contribution rate to measure the total error between the predicted and actual soil moisture values ​​for each soil layer.

[0095] S10322. The data fitting term and the physical constraint term are fused to obtain the composite loss function of the multi-task machine learning model.

[0096] In steps S10321-S10322, this application constructs a data fitting term and a physical constraint term based on the contribution rate to measure the total error between the predicted and actual measured values ​​of soil moisture in each soil layer. That is, based on the data fitting term, the model's predicted values ​​of soil moisture in each soil layer are made as close as possible to the actual measured values ​​on the ground. Its construction needs to consider the characteristics of multi-task output. The physical constraint term is to transform the physical knowledge that the root hydraulic characteristic parameters represent the contribution rate of soil moisture in each soil layer to root water absorption into a soft constraint on the model's prediction results. The composite loss function of the multi-task machine learning model is expressed by formula (14):

[0097] (14);

[0098] MSE stands for data fitting term. In other words, by adjusting the coefficient λ and defining f(X), the intensity and method of injecting physical knowledge can be flexibly controlled to adapt to different data conditions and application scenarios. Based on the aforementioned composite loss function and combined with a spatiotemporal cross-validation strategy, this application enhances the model's generalization ability and reliability, ensuring that the final inversion model's predicted output not only closely approximates actual measurements numerically but also conforms to the basic principles of plant water physiology in terms of the combination relationships of moisture in each soil layer, making the inversion results more physically meaningful and credible.

[0099] In step S104, the trained inversion model is used to infer the model input matrix, which is composed of standardized canopy comprehensive characteristic indicators, meteorological data, and auxiliary data. This outputs normalized predicted values ​​of soil volumetric water content for four root layers: 0-10 / 10-20 / 20-40 / 40-60cm. The extreme values ​​(θi,max, θi,min) of soil water content from the training set are used to convert the normalized predicted values ​​into actual soil volumetric water content. The resulting matrix of actual predicted soil moisture values ​​for each root layer is the inversion result. The inversion result is then quickly validated to eliminate obvious errors, such as reasonable range verification, physical consistency verification, and outlier removal, thus obtaining the final inversion result. The corresponding soil moisture data for each root layer, combined with spatial grid and temporal information, generates a spatiotemporal dynamic distribution map, realizing the visualization and intuitive display of soil moisture. Based on the aforementioned spatiotemporal distribution map of soil moisture, remote sensing monitoring of farmland soil moisture is completed. The generated spatiotemporal distribution map of soil moisture is applied to remote sensing monitoring of farmland soil moisture, realizing full-process monitoring of state assessment, problem identification, and decision support. Problems discovered during the monitoring process (such as model inversion error areas) can also be fed back to the model training stage to iteratively optimize the inversion model. For example, for areas with large inversion errors, ground measurement data is supplemented and the model is retrained. The entire process realizes a closed loop from model inversion to actual monitoring application, which is the core implementation link of soil moisture remote sensing monitoring in smart agriculture.

[0100] In the specific implementation of step S104, one embodiment is as follows: using a trained inversion model, soil moisture at different depths in the crop root zone is inverted, and a spatiotemporal distribution map of soil moisture is generated, including:

[0101] S1041. Using the inversion model, synchronously output the predicted soil moisture values ​​at different depths corresponding to each spatial location point within the farmland area.

[0102] S1042. Based on the predicted soil moisture values ​​at the different spatial locations, generate a spatial distribution map of soil moisture in the farmland area.

[0103] In steps S1041-S1042, this application assigns corresponding pixel geographic coordinates to the soil moisture values ​​SMi′ obtained from the inversion model for each soil layer. Using Geographic Information System (GIS) software or programming libraries (such as GDAL, ArcPy), a separate thematic map of soil moisture spatial distribution is created for each soil layer. This map is in raster format, with the pixel value representing soil volumetric water content. It is typically visualized using continuous color bands (e.g., from blue to red), where blue represents high moisture and red represents low moisture. The map should include necessary legends, scale bars, a north arrow, and labeling information (such as region, date, and soil depth). Based on the predicted soil moisture values ​​at different spatial locations, a spatial distribution map of soil moisture in the farmland area is generated. When inversion results for multiple dates from the same region are available, time series analysis can be performed. For specific key points or areas within farmland, curves showing the changes in soil moisture over time in different soil layers can be plotted, visually demonstrating water consumption, replenishment, and migration processes within the profile. Furthermore, spatial distribution maps of multiple periods and soil layers can be stacked chronologically to form a spatiotemporal data cube of soil moisture. Dynamic changes in moisture across time and space (including vertical depth) can be displayed through slices or animations. The inverted moisture values ​​for each period and soil layer can also be compared with the optimal soil moisture threshold for that crop growth stage, generating a water deficit depth map or a spatiotemporal distribution map of water deficit, directly indicating irrigation demand. The final output of the inversion model in this application is not a single data point, but rather a thematic map and prescription information that seamlessly integrates with agricultural decision-making processes such as irrigation and drought management, achieving a closed loop for remote sensing monitoring value. Based on the inversion model, only new UAV aerial survey data is needed for each subsequent monitoring process, keeping costs relatively controllable. Continuous and dynamic monitoring of the entire growing season or even several years can be achieved.

[0104] Example 2

[0105] This application also provides a remote sensing monitoring device for farmland soil moisture, such as Figure 4 The diagram shows a block diagram of a farmland soil moisture remote sensing monitoring device. The functions of this device correspond to the steps of executing a farmland soil moisture remote sensing monitoring method on a terminal device as described above. This device can be understood as a server component including a processor. The farmland soil moisture remote sensing monitoring device described in this application includes:

[0106] The acquisition module 401 is used to simultaneously acquire farmland ground data through multispectral and thermal infrared remote sensing by UAV, obtain multimodal data, and construct a moisture-sensitive canopy comprehensive characteristic index based on the multispectral remote sensing images and thermal infrared remote sensing images in the multimodal data; the multimodal data also includes meteorological data;

[0107] The quantification module 402 is used to construct a canopy-root water transport model based on thermal infrared remote sensing images and meteorological data within a preset time window, so as to quantify the contribution rate of each root layer soil to root water absorption.

[0108] Training module 403 is used to train a standardized multi-task machine learning model by using the comprehensive feature index of the canopy and the contribution rate as physical constraints, so as to train an inversion model of soil moisture for different soil layers.

[0109] The monitoring module 404 is used to invert soil moisture at different depths in the crop root zone using a trained inversion model and generate a spatiotemporal distribution map of soil moisture, so as to complete remote sensing monitoring of farmland soil moisture based on the spatiotemporal distribution map of soil moisture.

[0110] In one feasible implementation, the quantization module includes:

[0111] The soil moisture content of different root layers in a continuous time series was calculated using the central difference method to obtain the root water absorption rate of each root layer.

[0112] Based on the root water absorption rate of each root layer and meteorological data, a canopy-root water transport model was constructed to determine the contribution rate of each root layer soil to root water absorption.

[0113] In one feasible implementation, the quantization module further includes:

[0114] A canopy water balance network is constructed based on the root water absorption rate of each root layer to process the root water absorption rate of each root layer and meteorological data to obtain the simulated change in canopy water volume data.

[0115] By minimizing the error between the simulated and actual changes, a set of root hydraulic characteristic parameters are obtained to determine the contribution rate.

[0116] In one feasible implementation, the training module includes:

[0117] A training dataset was constructed based on the comprehensive characteristic index of the canopy and the actual measured values ​​of soil moisture in each soil layer.

[0118] Using the training dataset, the multi-task machine learning model is iteratively trained with the objective of minimizing the composite loss function of the multi-task machine learning model.

[0119] In one feasible implementation, the training module further includes:

[0120] Data fitting terms and physical constraint terms based on contribution rates were constructed to measure the total error between the predicted and actual soil moisture values ​​for each soil layer.

[0121] By combining the data fitting term and the physical constraint term, the composite loss function of the multi-task machine learning model is obtained.

[0122] In one feasible implementation, the acquisition module includes:

[0123] At least one water-sensitive spectral index, vegetation index, and canopy temperature are extracted from the multispectral remote sensing image and the thermal infrared remote sensing image, respectively.

[0124] A moisture-sensitive thermal infrared index is constructed by processing the canopy temperature and vegetation index using a wet-dry edge fitting method, and then fused with the spectral index to obtain a comprehensive canopy characteristic index.

[0125] In one feasible implementation, the monitoring module includes:

[0126] Using the inversion model, the predicted soil moisture values ​​at different depths corresponding to each spatial location point within the farmland area are output simultaneously.

[0127] Based on the predicted soil moisture values ​​at different spatial locations, a spatial distribution map of soil moisture in the farmland area is generated.

[0128] Example 3

[0129] This application also provides an electronic device, such as Figure 5 As shown, it includes: a processor 501, a memory 502, and a bus 503. The memory 502 stores machine-readable instructions that can be executed by the processor 501. When the electronic device is running, the processor 501 and the memory 502 communicate through the bus 503. When the machine-readable instructions are executed by the processor 501, the steps of any one of the methods for remote sensing monitoring of farmland soil moisture are performed.

[0130] Example 4

[0131] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of any one of the methods for remote sensing monitoring of farmland soil moisture.

Claims

1. A method for remote sensing monitoring of farmland soil moisture, characterized in that, The method includes: Multimodal data was obtained by simultaneously collecting farmland ground data using multispectral and thermal infrared remote sensing from UAVs. Based on the multispectral and thermal infrared remote sensing images in the multimodal data, a moisture-sensitive comprehensive canopy characteristic index was constructed. The multimodal data also included meteorological data. Within a preset time window, a canopy-root water transport model is constructed based on thermal infrared remote sensing images and meteorological data to quantify the contribution rate of each root layer soil to root water absorption. Using the comprehensive canopy feature index and the contribution rate as physical constraints, a standardized multi-task machine learning model is trained to obtain an inversion model of soil moisture for different soil layers. Using a trained inversion model, soil moisture at different depths in the crop root zone is inverted, and a spatiotemporal distribution map of soil moisture is generated, so as to complete remote sensing monitoring of farmland soil moisture based on the spatiotemporal distribution map of soil moisture.

2. The method as described in claim 1, characterized in that, A canopy-root water transport model was constructed based on thermal infrared remote sensing imagery and meteorological data to quantify the contribution of each root layer soil layer to root water uptake, including: The soil moisture content of different root layers in a continuous time series was calculated using the central difference method to obtain the root water absorption rate of each root layer. Based on the root water absorption rate of each root layer and meteorological data, a canopy-root water transport model was constructed to determine the contribution rate of each root layer soil to root water absorption.

3. The method as described in claim 2, characterized in that, The canopy-root water transport model is constructed based on the root water absorption rate and meteorological data of each root layer to determine the contribution rate of each root layer soil to root water absorption, including: A canopy water balance network is constructed based on the root water absorption rate of each root layer to process the root water absorption rate of each root layer and meteorological data to obtain the simulated change in canopy water volume data. By minimizing the error between the simulated and actual changes, a set of root hydraulic characteristic parameters are obtained to determine the contribution rate.

4. The method as described in claim 1, characterized in that, Using the canopy comprehensive feature index and the contribution rate as physical constraints, a standardized multi-task machine learning model is trained, including: A training dataset was constructed based on the comprehensive characteristic index of the canopy and the actual measured values ​​of soil moisture in each soil layer. Using the training dataset, the multi-task machine learning model is iteratively trained with the objective of minimizing the composite loss function of the multi-task machine learning model.

5. The method as described in claim 4, characterized in that, The objective is to minimize the composite loss function of the multi-task machine learning model, including: Data fitting terms and physical constraint terms based on contribution rates were constructed to measure the total error between the predicted and actual soil moisture values ​​for each soil layer. By combining the data fitting term and the physical constraint term, the composite loss function of the multi-task machine learning model is obtained.

6. The method as described in claim 1, characterized in that, Based on the aforementioned multispectral remote sensing imagery and thermal infrared remote sensing imagery, a moisture-sensitive comprehensive canopy characteristic index is constructed, including: At least one water-sensitive spectral index, vegetation index, and canopy temperature are extracted from the multispectral remote sensing image and the thermal infrared remote sensing image, respectively. A moisture-sensitive thermal infrared index is constructed by processing the canopy temperature and vegetation index using a wet-dry edge fitting method, and then fused with the spectral index to obtain a comprehensive canopy characteristic index.

7. The method as described in claim 1, characterized in that, Using a trained inversion model, soil moisture at different depths in the crop root zone is retrieved, and a spatiotemporal distribution map of soil moisture is generated, including: Using the inversion model, the predicted soil moisture values ​​at different depths corresponding to each spatial location point within the farmland area are output simultaneously. Based on the predicted soil moisture values ​​at different spatial locations, a spatial distribution map of soil moisture in the farmland area is generated.

8. A remote sensing monitoring device for farmland soil moisture, characterized in that, The device includes: The data acquisition module is used to simultaneously acquire farmland ground data using UAV multispectral and thermal infrared remote sensing to obtain multimodal data, and to construct a moisture-sensitive comprehensive canopy characteristic index based on the multispectral and thermal infrared remote sensing images in the multimodal data; the multimodal data also includes meteorological data; The quantification module is used to construct a canopy-root water transport model based on thermal infrared remote sensing images and meteorological data within a preset time window, so as to quantify the contribution rate of each root layer soil to root water absorption. The training module is used to train a standardized multi-task machine learning model by using the comprehensive feature index of the canopy and the contribution rate as physical constraints, so as to obtain the soil moisture inversion model for different soil layers. The monitoring module is used to invert soil moisture at different depths in the crop root zone using a trained inversion model and generate a spatiotemporal distribution map of soil moisture, so as to complete remote sensing monitoring of farmland soil moisture based on the spatiotemporal distribution map of soil moisture.

9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of a remote sensing monitoring method for farmland soil moisture as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of a remote sensing monitoring method for farmland soil moisture as described in any one of claims 1 to 7.

Images