Soil deep humidity prediction system and method based on near-infrared spectroscopy and federated learning

Abstract

The application belongs to the technical field of soil humidity prediction methods, and in particular relates to a soil deep humidity prediction system and method based on near-infrared spectroscopy and federated learning. The system comprises a near-infrared spectroscopy acquisition device, an auxiliary information acquisition module, an edge data preprocessing module, a federated learning client module, a cloud federated aggregation server and a human-computer interaction and device control module. The method establishes a surface humidity estimation model by acquiring near-infrared reflectivity of the soil surface layer and environmental auxiliary information. A deep humidity prediction model is constructed by combining a heat conduction equation and machine learning, and the accuracy is improved by using weighted fusion. The federated learning framework is introduced to realize local training of each agricultural machine and cloud aggregation, which protects data privacy and improves model generalization ability. The system does not need to destroy the soil structure, has the advantages of non-contact, high efficiency, safety and scalability, and is suitable for intelligent prediction of large-scale deep humidity in farmland and linkage control of agricultural operations.

Description

Technical Field

[0001] The present invention relates to the technical field of soil moisture prediction methods, and specifically to a deep soil moisture prediction system and method based on near-infrared spectroscopy and federated learning. Background Art

[0002] Soil moisture is a critical environmental parameter in agricultural production, directly influencing crop emergence rates, growth cycles, and agronomic management decisions such as fertilization and irrigation. Mainstream soil moisture detection methods include manual sampling and embedded electrode sensors. While these methods offer some accuracy, they suffer from cumbersome deployment, high costs, difficult maintenance, limited spatial coverage, and poor real-time performance. They are particularly unsuitable for scenarios requiring large-scale dynamic monitoring or deployment on mobile agricultural machinery for real-time operational monitoring.

[0003] With the development of near-infrared spectroscopy, its application in non-contact soil moisture detection has gradually attracted attention. Near-infrared spectroscopy has a strong absorption capacity for water in specific wavelengths. Research has shown that by analyzing soil reflectance spectral data, it is possible to quickly estimate the moisture content of the soil surface.

[0004] However, existing technologies mostly focus on estimating surface soil moisture, while effective methods for real-time measurement of moisture in deeper layers (e.g., 15 cm) remain lacking. In actual agricultural operations, crop seeds are often buried between 5 and 15 cm below the surface. Moisture levels at this depth directly impact sowing success and crop rooting. Surface moisture alone cannot fully reflect moisture conditions at depth.

[0005] A search revealed Chinese patent application number 202411536560.9, titled "A Soil Moisture Prediction System and Method Based on Multispectral Imaging," which proposes using multispectral imagery to analyze soil surface moisture and estimate moisture distribution through a model. However, this method focuses on extracting visual texture features and is primarily applied to surface moisture estimation. Furthermore, the model training process relies on centralized data collection, making it difficult to meet the requirements of privacy protection and multi-region generalization modeling.

[0006] With the rise of intelligent agricultural machinery and edge computing, data distribution in agricultural scenarios exhibits both regional heterogeneity and privacy sensitivity. Traditional centralized training models face the risks of data silos and privacy leaks, making it difficult to support multi-region, multi-device collaborative modeling. In this context, federated learning, as a distributed collaborative training framework, provides a new solution for data privacy protection and cross-device modeling. It allows agricultural machinery terminals to complete model training locally and only upload model parameters for cloud aggregation, effectively avoiding the risks and costs associated with uploading raw data.

[0007] In summary, there is an urgent need for an innovative system that combines near-infrared spectroscopy non-contact detection, deep moisture prediction modeling, and federated learning collaborative mechanisms. This system can ensure soil data privacy while achieving dynamic deep moisture perception and intelligent prediction in a wide range of scenarios.

[0008] Therefore, we propose a soil deep moisture prediction system and method based on near-infrared spectroscopy and federated learning to solve the above problems. SUMMARY

[0009] This section aims to summarize some aspects of the embodiments of the present application and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract and title of the specification to avoid obscuring the purpose of this section, abstract and title, and such simplifications or omissions cannot be used to limit the scope of the present application.

[0010] To solve the above technical problems, according to one aspect of the present application, the present application provides the following technical solutions:

[0011] A soil deep moisture prediction system based on near-infrared spectroscopy and federated learning, comprising:

[0012] A near-infrared spectroscopy acquisition device integrated into a mobile agricultural platform for dynamically acquiring reflectance spectroscopy data of the soil surface in the near-infrared band during operation;

[0013] An auxiliary information acquisition module integrated into the mobile agricultural platform for synchronously recording soil sampling location, time, soil type, and surface temperature data to provide context support for subsequent modeling;

[0014] An edge data preprocessing module deployed on the mobile agricultural platform and connected to the signal output end of the near-infrared spectroscopy acquisition device for receiving and processing the reflectance spectroscopy data acquired by the near-infrared spectroscopy acquisition device;

[0015] A federated learning client module embedded in each agricultural terminal for independently training a surface soil moisture estimation model and a deep prediction sub-model based on locally acquired and processed spectroscopy data and auxiliary information, and coordinating and aggregating with the cloud through a federated learning protocol to achieve distributed modeling without uploading raw data;

[0016] A cloud federated aggregation server receiving local model parameters uploaded by each federated learning client module, performing FedAvg model aggregation, outputting a unified prediction model, and periodically returning to the federated learning client module;

[0017] A human-computer interaction and device control module for displaying the predicted soil deep moisture value and automatically linking the irrigation system and seeding control block to achieve closed-loop control of intelligent agricultural operation.

[0018] As a preferred solution of the deep soil moisture prediction system based on near infrared spectroscopy and federated learning described in the present invention, the data preprocessing method of the edge data preprocessing module is:

[0019] Smoothing and filtering the collected raw near-infrared reflectance data using Savitzky–Golay smoothing, wavelet transform, or moving average algorithm;

[0020] In order to eliminate systematic errors such as illumination conditions, standard normal variable transformation and multivariate scatter correction method were used to perform spectral standardization and enhance the comparability between different samples;

[0021] Principal component analysis and variable selection algorithm are used to extract key band features related to moisture and construct feature vectors.

[0022] As a preferred embodiment of the deep soil moisture prediction system based on near-infrared spectroscopy and federated learning described in the present invention, the surface soil moisture estimation method in the federated learning client module is as follows: different types of soil samples are collected, and while obtaining near-infrared spectral data, the actual moisture value of the soil at the corresponding depth is measured using a standard method to construct paired feature-label samples;

[0023] The partial least squares regression (PLSR) algorithm was used to build the model, and the training sample set was input into the partial least squares regression (PLSR) model. K-fold cross validation was used for training evaluation, and the model parameters were adjusted to obtain the optimal performance.

[0024] The newly collected near-infrared spectral data were preprocessed in the same way and then input into the trained partial least squares regression (PLSR) model to output the corresponding surface soil moisture estimation.

[0025] As a preferred solution of the deep soil moisture prediction system based on near infrared spectroscopy and federated learning described in the present invention, the specific method of training the deep prediction sub-model is:

[0026] Based on the one-dimensional unsteady-state heat conduction equation, the initial shallow humidity is introduced as a boundary condition. Combined with input parameters such as soil thermal diffusivity, water-heat coupling parameters, and surface temperature changes, a spatiotemporal evolution model of deep humidity is derived. The model is described as follows:

[0027]

[0028] in, Indicates time Moment Depth The moisture content at is the soil moisture diffusion coefficient, Represents the disturbance term, and the preliminary prediction value is obtained by numerically solving the equation .

[0029] As a preferred solution of the deep soil moisture prediction system based on near-infrared spectroscopy and federated learning described in the present invention, multiple soil samples are collected, and while simultaneously collecting the surface near-infrared spectrum, the true moisture value of the corresponding 5-15 cm deep soil is obtained through profile sampling to construct a surface-deep moisture paired sample set;

[0030] The surface moisture value estimated by near-infrared spectroscopy is combined with auxiliary information such as location, time, soil type, and surface temperature as model input features to enhance the model's adaptability and prediction accuracy under various environmental conditions.

[0031] The gradient boosting tree (GBRT) algorithm was used to establish a regression model of surface-deep humidity.

[0032] The constructed sample set is used to train the model, and the cross-validation method is used to evaluate the model's prediction performance and optimize the parameter configuration;

[0033] The newly observed surface humidity and its related environmental parameters are input into the trained model to obtain the machine learning prediction value .

[0034] As a preferred solution of the deep soil moisture prediction system based on near infrared spectroscopy and federated learning described in the present invention, the prediction results of the physical model are Predicting results with machine learning Perform weighted fusion, the specific formula is as follows:

[0035]

[0036] in, is the weight coefficient, which is dynamically adjusted based on the average error of the model in historical samples, the uncertainty of the two models under current observation conditions, and the local model performance evaluation results fed back during the federated learning process to improve the overall prediction performance;

[0037] Each agricultural machine completes the model deployment after training locally, and regularly receives the model parameters after global optimization from federated learning to continuously optimize the prediction effect. After the new round of observed surface moisture and environmental parameters are input into the fusion model, the corresponding deep soil moisture estimation is output in real time, realizing the intelligent function of prediction while working.

[0038] As a preferred solution of the deep soil moisture prediction system based on near-infrared spectroscopy and federated learning described in the present invention, the specific manner in which the cloud-based federated aggregation server issues a unified prediction model after executing the aggregation strategy is as follows:

[0039] After each agricultural machine completes a certain round of local training or accumulates enough samples, it triggers a federated communication round and uploads its current local model parameters, namely the PLSR regression coefficients, without uploading any original spectral data or label information.

[0040] The uploaded model parameters are perturbed by differential privacy and encrypted communication is carried out through the TLS protocol;

[0041] The cloud-based federated aggregation server uses the standard FedAvg strategy to average the model weights uploaded by all federated learning client modules according to the number of samples and updates them as global model parameters.

[0042] The aggregated model is distributed to each federated learning client module using a unified global model strategy;

[0043] The system supports asynchronous federated training, allowing some agricultural machinery to skip a round of aggregation due to network disconnection or operation interruption, and the aggregation server automatically adjusts the weight distribution.

[0044] A method for predicting deep soil moisture based on near-infrared spectroscopy and federated learning, comprising the following steps:

[0045] Step 1: Scan the target area using a near-infrared spectrum acquisition device to obtain the soil surface reflectance spectrum;

[0046] Step 2: Use sensors deployed on the mobile agricultural machinery platform to synchronously record soil sampling location, time, soil type, and surface temperature data;

[0047] Step 3: Receive and process the near-infrared spectral image data output by the near-infrared spectral imaging device through the data preprocessing module on each agricultural machinery terminal;

[0048] Step 4: Each client builds a regression model based on its own sample data to estimate surface soil moisture;

[0049] Step 5: While keeping the data locally stored, each agricultural machine periodically uploads local model parameters to the cloud federation aggregation server. The server then executes the aggregation strategy and issues the global model.

[0050] Step 6: Each agricultural machine trains a deep moisture prediction model based on the surface moisture estimation obtained through federated training, auxiliary information, and historical deep moisture data;

[0051] Step 7: The human-computer interaction device outputs the prediction results to the display terminal or uploads them to the smart agriculture platform, and interacts with external equipment such as planters and irrigation systems.

[0052] Compared with the existing technology, the present invention has the following beneficial effects: the present invention obtains soil surface reflectance information through a near-infrared spectral camera, realizes non-contact rapid detection of soil moisture, avoids the tedious installation and maintenance operations of traditional buried sensors, has the advantages of convenient deployment, fast response speed and non-destructiveness, and is particularly suitable for dynamic monitoring of large-scale fields.

[0053] This invention innovatively introduces a federated learning framework, enabling each agricultural machine terminal to independently complete model training locally, uploading only the model parameters to the cloud for aggregation. This effectively avoids the centralized transmission and storage of raw near-infrared spectral data, significantly enhancing the system's ability to protect farm privacy data. Furthermore, the federated aggregation mechanism enables cross-regional and cross-soil type modeling collaboration, improving the model's generalization and environmental adaptability.

[0054] Leveraging spectral modeling and auxiliary features (such as surface moisture estimates, geographic location, time of year, soil type, and surface temperature), a multi-source fusion deep moisture prediction model was constructed. This model enables highly accurate indirect inference of soil moisture at specific depths (e.g., 15 cm) without deep sampling, expanding the application depth of near-infrared technology and improving the efficiency and accuracy of acquiring deep moisture information. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be derived from these drawings without inventive effort. Among them:

[0056] Figure 1 This is a schematic diagram of the structure of a deep soil moisture prediction system based on near-infrared spectroscopy and federated learning in the present invention;

[0057] Figure 2 This is a flow chart of a method for predicting deep soil moisture based on near-infrared spectroscopy and federated learning according to the present invention. DETAILED DESCRIPTION

[0058] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0059] Next, the present invention is described in detail with reference to schematic diagrams. For ease of illustration, cross-sectional views of device structures may be partially enlarged and not to scale when describing the embodiments of the present invention. Furthermore, the schematic diagrams are merely illustrative and should not limit the scope of protection of the present invention. Furthermore, in actual production, three-dimensional dimensions, including length, width, and depth, should be included.

[0060] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0061] Example 1

[0062] like Figure 1 As shown, an embodiment of the present invention proposes a deep soil moisture prediction system based on near-infrared spectroscopy and federated learning, including:

[0063] A near-infrared spectral acquisition device, mounted on the front of a mobile agricultural machine, faces the ground and continuously acquires near-infrared reflectance spectral data from the soil surface during operation. The acquired spectral information covers specific moisture-sensitive bands and can reflect the changing characteristics of soil moisture content and material composition, serving as a key input for subsequent moisture prediction.

[0064] The auxiliary information collection module is used to synchronously collect auxiliary parameters such as the geographic location information of the operation site, sampling timestamp, soil type, and surface temperature. These auxiliary variables are used as covariate inputs to enhance the generalization ability and environmental adaptability of the humidity prediction model.

[0065] The edge data preprocessing module is deployed on a mobile agricultural machinery platform and connected to the near-infrared spectral acquisition device. It receives the raw reflectance spectral data output by the device in real time and performs preprocessing operations such as spectral smoothing, noise filtering, and standard normal transformation to improve data quality and modeling stability.

[0066] The federated learning client module is embedded in each agricultural machinery terminal. Based on locally collected and processed spectral data and auxiliary information, it independently trains the surface soil moisture estimation model and the deep prediction sub-model. It then coordinates and aggregates with the cloud through the federated learning protocol, enabling distributed modeling of multiple agricultural machinery without uploading raw data, protecting farm privacy, and enhancing model generalization performance.

[0067] The cloud-based federated aggregation server receives local model parameters uploaded by each federated learning client module, performs FedAvg model aggregation, outputs a unified prediction model, and periodically transmits it back to each federated learning client module;

[0068] The human-computer interaction and equipment control module provides a graphical display interface and communication interface, which can present the prediction results to agricultural machinery operators in real time through the screen, or upload them to the agricultural Internet of Things platform to achieve automatic linkage and operation instruction generation with downstream control systems such as seeders and irrigation equipment.

[0069] Example 2

[0070] like Figure 2As shown, the present embodiment is basically the same as the previous embodiment, the difference lies in a soil deep humidity prediction method based on near-infrared spectroscopy and federated learning, comprising:

[0071] Step one: use the near-infrared spectrum acquisition device deployed on the mobile agricultural machine to scan the target farmland area, and obtain the reflectance spectrum of the soil surface. The collected near-infrared spectrum covers the water absorption sensitive band and can reflect the spectral response characteristics of the water content in the soil;

[0072] Step two: use the sensor system carried by the mobile agricultural machine platform to record the soil sampling position, collection time, soil type, ground temperature and other environmental auxiliary information in real time and synchronously, and provide feature supplement and sample label for subsequent modeling;

[0073] Step three: input the near-infrared spectrum image data into the edge data preprocessing module, and perform denoising, smoothing, baseline correction, wavelet transform, principal component analysis (PCA) and other processing on the original spectrum data to extract stable and representative spectrum features;

[0074] The specific way of data preprocessing is:

[0075] Smooth and filter the collected original near-infrared reflectivity data, use Savitzky-Golay smoothing, wavelet transform or moving average algorithm to reduce high-frequency noise interference and improve data stability;

[0076] In order to eliminate system errors such as light conditions, the standard normal variable transformation and multivariate scatter correction method are preferably used for spectral standardization processing to enhance the comparability between different samples;

[0077] Use principal component analysis and variable selection algorithm to extract key band features related to water and construct feature vectors;

[0078] Step four: each client builds a regression model based on its own sample data to estimate the 0~5cm surface soil humidity;

[0079] The specific way of each client building a regression model based on its own sample data to estimate the 0~5cm surface soil humidity is:

[0080] Collect different types of soil samples, obtain near-infrared spectrum data at the same time, and use standard methods to measure the true humidity value of the corresponding depth (0~5cm) soil to build paired feature-label samples;

[0081] Use partial least squares regression PLSR algorithm to model the model;

[0082] Input the training sample set into the partial least squares regression PLSR model, use K-fold cross-validation for training evaluation, and adjust the model parameters to obtain the optimal performance;

[0083] The newly collected near-infrared spectrum data is input into the trained partial least squares regression (PLSR) model after the same pretreatment, and the corresponding 0-5cm surface soil moisture estimation value is output;

[0084] Step five: Under the premise of keeping the data local and not uploading, each agricultural machine periodically uploads the local model parameters (such as PLSR regression coefficients or neural network weights) to the cloud federated aggregation server, and the server issues a global model after executing the aggregation strategy;

[0085] The specific way of periodically uploading local model parameters to the cloud federated aggregation server and issuing a global model after the server executes the aggregation strategy is as follows:

[0086] After each agricultural machine completes a certain number of training rounds or accumulates enough samples locally, it triggers a federated communication round and uploads its current local model parameters, i.e., PLSR regression coefficients, without uploading any original spectrum data or label information, thereby ensuring data privacy and communication efficiency;

[0087] The uploaded model parameters are subjected to differential privacy perturbation and encrypted communication through the TLS protocol to prevent man-in-the-middle attacks or model inversion and improve the federated security of the system;

[0088] The cloud federated aggregation server uses the standard FedAvg strategy to weight-average all client-uploaded model weights according to the sample size and updates them to global model parameters;

[0089] The aggregated model is issued to each client using a unified global model strategy;

[0090] The system supports asynchronous federated training, allowing some agricultural machines to skip a round of aggregation due to network disconnection or work interruption, and the aggregation server automatically adjusts the weight distribution to improve system robustness and adaptability;

[0091] Step six: Each agricultural machine trains a deep layer moisture prediction model based on the surface moisture estimation value obtained through federated training, auxiliary information, and historical deep layer moisture data;

[0092] The specific way of training the deep layer moisture prediction model is as follows:

[0093] Based on the one-dimensional non-steady-state heat conduction equation, the initial shallow layer moisture is introduced as a boundary condition, and input parameters such as soil thermal diffusivity, water-heat coupling parameters, and surface temperature variation are combined to derive a spatio-temporal evolution model of deep layer moisture. The model is described as follows:

[0094]

[0095] where, t represents time depth The moisture content at is the soil moisture diffusion coefficient, Denotes the disturbance term. By numerically solving this equation, we obtain the preliminary prediction value .

[0096] Collect multiple soil samples, and simultaneously collect near-infrared spectra of the surface layer. Use profile sampling to obtain the true moisture value of the corresponding 5-15 cm deep soil layer, and construct a surface moisture-deep moisture paired sample set.

[0097] The surface moisture value estimated by near-infrared spectroscopy is combined with auxiliary information such as location, time, soil type, and surface temperature as model input features to enhance the model's adaptability and prediction accuracy under various environmental conditions.

[0098] The gradient boosting tree (GBRT) algorithm was used to establish a regression model of surface-deep humidity.

[0099] The constructed sample set is used to train the model, and the cross-validation method is used to evaluate the model's prediction performance and optimize the parameter configuration;

[0100] The newly observed surface humidity and its related environmental parameters are input into the trained model to obtain the machine learning prediction value .

[0101] Prediction results of physical models Predicting results with machine learning Perform weighted fusion, the specific formula is as follows:

[0102]

[0103] in, The weight coefficient is dynamically adjusted according to the average error of the model in historical samples, the uncertainty of the two models under current observation conditions, and the local model performance evaluation results fed back during the federated learning process to improve the overall prediction performance.

[0104] Each agricultural machine deploys the trained model locally and regularly receives globally optimized model parameters from federated learning to continuously optimize prediction results. Newly observed surface moisture and environmental parameters are fed into the fusion model, which outputs a corresponding deep soil moisture estimate in real time, enabling intelligent prediction while operating.

[0105] Step 7: The prediction results are displayed on the agricultural machinery terminal interface via a human-computer interaction device or uploaded to a remote agricultural management platform. The system can then link smart devices such as planters and irrigation controllers based on current soil moisture and crop needs, enabling intelligent adjustment of agricultural machinery operating parameters and optimization of operation plans.

[0106] Although the present application has been described with reference to the embodiments above, various changes and modifications can be suggested to one skilled in the art, and it is intended that the present application encompass such changes and modifications as fall within the scope of the appended claims. Particularly, each feature disclosed in the description and / or the claims can be used in the combination with each of the features disclosed in the description and / or the claims, unless specifically stated otherwise. Therefore, the present application is not intended to be limited to the particular embodiments disclosed in the description and / or the claims.

Claims

1. A deep soil moisture prediction system based on near infrared spectroscopy and federated learning, characterized in that: include: A near-infrared spectrum acquisition device, integrated into a mobile agricultural machinery platform, is used to dynamically collect reflectance spectrum data of the soil surface in the near-infrared band during operation; An auxiliary information collection module, integrated into the mobile agricultural machinery platform, is used to synchronously record soil sampling location, time, soil type, and surface temperature data, providing contextual support for subsequent modeling; An edge data preprocessing module, deployed on a mobile agricultural machinery platform and connected to a signal output terminal of the near-infrared spectrum acquisition device, for receiving and processing the reflectance spectrum data collected by the near-infrared spectrum acquisition device; The federated learning client module is embedded in each agricultural machinery terminal. Based on locally collected and processed spectral data and auxiliary information, it independently trains the surface soil moisture estimation model and the deep soil moisture prediction sub-model. It coordinates and aggregates with the cloud through the federated learning protocol, enabling distributed modeling of multiple agricultural machinery without uploading raw data. The cloud-based federated aggregation server receives local model parameters uploaded by each federated learning client module, performs FedAvg model aggregation, outputs a unified prediction model, and periodically transmits it back to the federated learning client module; The human-computer interaction and equipment control module is used to display the deep soil moisture value predicted by the model, and automatically link the irrigation system and sowing control block to achieve closed-loop control of intelligent agricultural machinery operations.

2. The deep soil moisture prediction system based on near infrared spectroscopy and federated learning according to claim 1 is characterized in that: The data preprocessing method of the edge data preprocessing module is: Smoothing and filtering the collected raw near-infrared reflectance data using Savitzky–Golay smoothing, wavelet transform, or moving average algorithm; To eliminate systematic errors in illumination conditions, standard normal variable transformation and multivariate scatter correction were used to perform spectral standardization and enhance comparability among different samples. Principal component analysis and variable selection algorithm are used to extract key band features related to moisture and construct feature vectors.

3. The deep soil moisture prediction system based on near infrared spectroscopy and federated learning according to claim 1 is characterized in that: In the federated learning client module, the surface soil moisture estimation method is as follows: different types of soil samples are collected, and while obtaining near-infrared spectral data, the actual soil moisture values ​​at the corresponding depths are measured using standard methods to construct paired feature-label samples; The partial least squares regression (PLSR) algorithm was used to build the model, and the training sample set was input into the partial least squares regression (PLSR) model. K-fold cross validation was used for training evaluation, and the model parameters were adjusted to obtain the optimal performance. The newly collected near-infrared spectral data were preprocessed in the same way and then input into the trained partial least squares regression (PLSR) model to output the corresponding surface soil moisture estimation.

4. The deep soil moisture prediction system based on near infrared spectroscopy and federated learning according to claim 1 is characterized in that: The specific method of training the deep prediction sub-model is: Based on the one-dimensional unsteady-state heat conduction equation, the initial shallow humidity is introduced as the boundary condition. Combined with the soil thermal diffusivity, water-heat coupling parameters, and surface temperature change input parameters, a spatiotemporal evolution model of deep humidity is derived. The model is described as follows: ； in, Indicates time Moment Depth The moisture content at is the soil moisture diffusion coefficient, Represents the disturbance term, and the preliminary prediction value is obtained by numerically solving the equation .

5. The deep soil moisture prediction system based on near infrared spectroscopy and federated learning according to claim 4 is characterized in that: Collect multiple soil samples, and simultaneously collect near-infrared spectra of the surface layer. Use profile sampling to obtain the true moisture values ​​of the corresponding 5-15 cm deep soil layer, and construct a surface-deep moisture paired sample set. The surface moisture value estimated by near-infrared spectroscopy is combined with auxiliary information such as location, time, soil type, and surface temperature as model input features to enhance the model's adaptability and prediction accuracy under various environmental conditions. The gradient boosting tree (GBRT) algorithm was used to establish a regression model of surface-deep humidity. The constructed sample set is used to train the model, and the cross-validation method is used to evaluate the model's prediction performance and optimize the parameter configuration; The newly observed surface humidity and its related environmental parameters are input into the trained model to obtain the machine learning prediction value .

6. The deep soil moisture prediction system based on near infrared spectroscopy and federated learning according to claim 5 is characterized in that: Prediction results of physical models Predicting results with machine learning Perform weighted fusion, the specific formula is as follows: ； in, is the weight coefficient, which is dynamically adjusted based on the average error of the model in historical samples, the uncertainty of the two models under current observation conditions, and the local model performance evaluation results fed back during the federated learning process to improve the overall prediction performance; Each agricultural machine completes the model deployment after training locally, and regularly receives the model parameters after global optimization from federated learning to continuously optimize the prediction effect. After the new round of observed surface moisture and environmental parameters are input into the fusion model, the corresponding deep soil moisture estimation is output in real time, realizing the intelligent function of prediction while working.

7. The deep soil moisture prediction system based on near infrared spectroscopy and federated learning according to claim 1 is characterized in that: The specific method in which the cloud-based federated aggregation server issues a unified prediction model after executing the aggregation strategy is as follows: After each agricultural machine completes a certain round of local training or accumulates enough samples, it triggers a federated communication round and uploads its current local model parameters, namely the PLSR regression coefficients, without uploading any original spectral data or label information. The uploaded model parameters are perturbed by differential privacy and encrypted communication is carried out through the TLS protocol; The cloud-based federated aggregation server uses the standard FedAvg strategy to average the model weights uploaded by all federated learning client modules according to the number of samples and updates them as global model parameters. The aggregated model is distributed to each federated learning client module using a unified global model strategy; The system supports asynchronous federated training, allowing some agricultural machinery to skip a round of aggregation due to network disconnection or operation interruption, and the aggregation server automatically adjusts the weight distribution.

8. A method for predicting deep soil moisture based on near infrared spectroscopy and federated learning, used to implement a system for predicting deep soil moisture based on near infrared spectroscopy and federated learning as described in any one of claims 1 to 7, characterized in that: The following steps are involved: Step 1: Scan the target area using a near-infrared spectrum acquisition device to obtain the soil surface reflectance spectrum; Step 2: Use sensors deployed on the mobile agricultural machinery platform to synchronously record soil sampling location, time, soil type, and surface temperature data; Step 3: Receive and process the near-infrared spectral image data output by the near-infrared spectral imaging device through the data preprocessing module on each agricultural machinery terminal; Step 4: Each client builds a regression model based on its own sample data to estimate surface soil moisture; Step 5: While keeping the data locally stored, each agricultural machine periodically uploads local model parameters to the cloud federation aggregation server. The server then executes the aggregation strategy and issues the global model. Step 6: Each agricultural machine trains a deep moisture prediction model based on the surface moisture estimation obtained through federated training, auxiliary information, and historical deep moisture data; Step 7: The human-computer interaction device outputs the prediction results to the display terminal or uploads them to the smart agriculture platform, and interacts with the planter and irrigation system external equipment.

Images