Intelligent monitoring and early warning system for farmland soil moisture based on multispectral remote sensing

Abstract

This invention relates to the field of agricultural information technology, specifically disclosing an intelligent monitoring and early warning system for farmland soil moisture based on multispectral remote sensing. The system includes a data fusion and reconstruction module, an adaptive feature decoupling and transfer learning module, a multi-scale spatiotemporal collaborative inference module, and a dynamic threshold early warning and source tracing module. It generates continuous feature sequences through multi-source data fusion, decouples universal and regional features to adapt to new scenarios, performs high-precision inversion using multi-scale information, and provides anomaly early warning and source tracing based on dynamic thresholds. This invention achieves continuous, all-weather soil moisture monitoring, improves the model's generalization ability and monitoring accuracy in different farmland areas, and provides intelligent decision support for agricultural irrigation management.

Description

Technical Field

[0001] This invention belongs to the field of agricultural information technology, specifically relating to an intelligent monitoring and early warning system for farmland soil moisture based on multispectral remote sensing. Background Technology

[0002] In the fields of precision agriculture and smart agriculture, real-time and precise monitoring and management of the farmland environment using modern information technology is key to improving agricultural production efficiency and resource utilization efficiency. Among these, soil moisture, as a core parameter affecting crop growth, irrigation decisions, and yield prediction, is crucial for achieving refined management of agricultural production.

[0003] Multispectral remote sensing-based farmland soil moisture monitoring technology has become an important research direction in this field due to its advantages such as wide coverage, short acquisition cycle, and non-contact measurement. This technology mainly analyzes the surface reflectance spectral data acquired by multispectral sensors carried by satellites or UAVs, and uses the physical or statistical relationship between soil moisture and reflectance in specific bands to retrieve large-scale soil moisture information.

[0004] Existing technologies typically rely on continuous, high-quality remote sensing imagery data to build and operate monitoring models. However, in practical agricultural applications, these technologies face significant challenges: under common weather conditions such as cloudy and rainy weather, optical remote sensing data is easily obscured by clouds, leading to severe data gaps at key time points and creating numerous gaps in the monitoring time series, thus compromising the continuity and timeliness of monitoring. Simultaneously, agricultural scenarios are highly regional and diverse, with complex and variable soil moisture spectral response characteristics under different crop types, soil textures, and farming methods. Meanwhile, labeled data with high-precision ground-measured value matching suitable for model training and validation is extremely scarce. This directly results in data-driven intelligent monitoring models generally suffering from insufficient training and weak generalization ability, making it difficult to adapt to the complex and diverse actual farmland environments. These data-level bottlenecks limit the reliability and universality of multispectral remote sensing-based soil moisture monitoring systems in practical operations. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent monitoring and early warning system for farmland soil moisture based on multispectral remote sensing, in order to solve the problems in the existing technology of discontinuous remote sensing data sequences due to weather interference and insufficient generalization ability of intelligent models due to scarcity of labeled data.

[0006] This invention provides an intelligent monitoring and early warning system for farmland soil moisture based on multispectral remote sensing. The system includes a data fusion and reconstruction module, an adaptive feature decoupling and transfer learning module, a multi-scale spatiotemporal collaborative inference module, and a dynamic threshold early warning and source tracing module. The input of the data fusion and reconstruction module is connected to the remote sensing data source and the multi-source auxiliary data source, and its output is connected to the input of the adaptive feature decoupling and transfer learning module. The output of the adaptive feature decoupling and transfer learning module is connected to the input of the multi-scale spatiotemporal collaborative inference module. The output of the multi-scale spatiotemporal collaborative inference module is connected to the input of the dynamic threshold early warning and source tracing module.

[0007] The data fusion and reconstruction module is used to fuse and spatiotemporally reconstruct multi-source heterogeneous data when optical remote sensing data is missing due to cloud cover, in order to generate a continuous pseudo-optical feature sequence. This module specifically includes a multi-source data alignment submodule, a physical model-driven reconstruction submodule, and a sequence completion and smoothing submodule. The multi-source data alignment submodule first receives multispectral remote sensing image data streams from satellite or UAV platforms, and simultaneously receives passive microwave remote sensing brightness and temperature data from meteorological satellites, precipitation and temperature data from ground meteorological stations, and terrain data from a digital elevation model. This submodule uses a unified geographic coordinate system and timestamps to perform spatiotemporal registration on all input data, ensuring that all data layers are aligned on the same spatial grid and time point. The physical model-driven reconstruction submodule performs the core reconstruction calculations based on the aligned multi-source data. This submodule incorporates a coupled physical model based on the radiative transfer equation and the surface energy balance equation. This model uses passive microwave brightness and temperature, surface temperature, and precipitation data as the main driving inputs, and solves the equations to deduce the physical relationship between surface soil moisture content and surface reflectivity. Specifically, the model couples the microwave's sensitivity to soil moisture penetration with the optical reflectance's characterization of surface dryness and wetness. Under known initial values ​​of microwave-derived soil moisture and meteorological conditions, it iteratively calculates the theoretical surface reflectance values ​​for the corresponding optical band at the missing time, thereby generating pseudo-optical reflectance characteristics.

[0008] The sequence completion and smoothing submodule receives pseudo-optical feature points output by the physical model-driven reconstruction submodule and inserts them into the original incomplete multispectral remote sensing time series. This submodule employs a sequence modeling method based on temporal convolutional networks to smooth the completed mixed sequence, eliminating noise and abrupt changes that may be introduced by the physical model calculations, and finally outputting a spatiotemporally continuous and smooth pseudo-optical feature data cube.

[0009] The adaptive feature decoupling and transfer learning module is used to decouple transferable universal spectral features from limited labeled data and drive the model to quickly adapt to new, unlabeled scenarios. This module includes a feature decoupling encoding network, a domain discrimination and alignment submodule, and a few-sample fine-tuning submodule. The feature decoupling encoding network receives a continuous feature data cube as input from the data fusion and reconstruction module. This network employs a dual-branch encoder structure. One branch is a shared feature encoder, used to extract universal spectral features from the input data that are strongly correlated with soil moisture physical mechanisms and are insensitive to changes in crop type and soil texture. These features are mainly concentrated in the spectral variation patterns of the near-infrared and short-wave infrared bands, which are sensitive to water absorption. The other branch is a specific feature encoder, used to extract specific features related to local crop canopy structure, soil background reflectance, and other regional factors. The outputs of the two encoders are separated in the feature space.

[0010] The domain discrimination and alignment submodule connects to the output of the shared feature encoder. This submodule contains a domain classifier, trained to classify shared features from different regions and crop types as belonging to the same distribution domain. Through adversarial training, the shared feature encoder is forced to learn feature representations that can confuse the domain classifier, thereby maximizing the removal of region-specific information from features and ensuring the domain invariance of shared features. The few-sample fine-tuning submodule is activated when the model is deployed to a new farmland area. This submodule receives a very small amount of sample data from the target area, labeled with high-precision ground-measured soil moisture values. This submodule fixes the parameters of the shared feature encoder and performs rapid fine-tuning only on specific feature encoders and subsequent prediction network layers, using a small amount of labeled data to allow the model to quickly adapt to the spectral response characteristics of the new area.

[0011] The multi-scale spatiotemporal collaborative inference module integrates remote sensing information with different spatial resolutions and temporal frequencies to perform collaborative inference, generating a high spatial resolution and temporally continuous soil moisture distribution map. This module includes a multi-scale feature pyramid submodule, a spatiotemporal attention fusion submodule, and a soil moisture inversion submodule. The multi-scale feature pyramid submodule simultaneously receives universal features from the adaptive feature decoupling and transfer learning module, as well as native multispectral features from remote sensing data sources of different resolutions. This submodule constructs a feature pyramid containing feature maps at multiple spatial scales (high, medium, and low) through a series of convolution and upsampling operations. High-level features contain rich semantic information, while low-level features retain fine spatial details. The spatiotemporal attention fusion submodule performs weighted fusion of features at each scale in the feature pyramid. This submodule includes a spatiotemporal attention mechanism that automatically calculates the importance weights of each spatial location in the temporal series and each feature scale in the spatial context. For the temporal dimension, this mechanism focuses on key time points where soil moisture changes drastically; for the spatial dimension, it assigns higher weights to high-level semantic features in homogeneous regions and higher weights to low-level detailed features in detailed areas such as field boundaries and irrigation ditches. The soil moisture inversion submodule receives a unified feature representation after spatiotemporal attention-weighted fusion. This submodule consists of a fully connected neural network or a lightweight convolutional neural network, which maps the fused features to the final soil moisture value. The training objective of this network is to minimize the root mean square error between its predicted values ​​and the measured soil moisture values.

[0012] The dynamic threshold early warning and source tracing module is used to dynamically set soil moisture early warning thresholds based on historical data and real-time inference results, and to locate abnormal areas and analyze potential causes when an early warning is triggered. This module includes a dynamic threshold calculation submodule, an anomaly detection and spatial clustering submodule, and a multivariate correlation source tracing submodule. The dynamic threshold calculation submodule receives soil moisture inversion results for the current period and a historical period, as well as crop growth stage information and irrigation records for the same period. For each field unit, this submodule calculates the historical soil moisture percentile distribution for the current crop growth stage based on the crop's water requirement pattern, setting the 10th percentile as the drought early warning threshold and the 90th percentile as the excessive moisture early warning threshold. The thresholds are dynamically updated as the crop growth stage progresses.

[0013] The anomaly detection and spatial clustering submodule compares the current soil moisture value with dynamic thresholds in real time. An early warning is triggered when the soil moisture value in a certain area remains below the drought threshold or above the excessive moisture threshold for more than two monitoring cycles. This submodule further employs a spatial clustering algorithm to cluster continuously triggered pixels, forming contiguous abnormal patches, and calculates the area, average moisture deviation, and spatial morphology of each abnormal patch. The multivariate correlation and source tracing submodule is activated after the abnormal patch is formed. This submodule synchronously retrieves multi-source data for the area during the abnormal period and preceding periods, including the original remote sensing reflectance, meteorological data, topographic slope, and irrigation facility distribution map before reconstruction. By comparing and analyzing the statistical differences in these variables between the abnormal patch and the surrounding normal area, a source tracing report is generated. The report clearly identifies the dominant factors causing the abnormal soil moisture, such as scarce rainfall, uneven irrigation, or poor drainage due to topography.

[0014] Furthermore, the physical model drives the coupled radiative transfer equation and surface energy balance equation in the reconstruction submodule, and the solution process employs a numerical iterative method. First, the initial value of surface soil volumetric water content is obtained by inverting the Dixon model using passive microwave data. This initial value, along with near-surface air temperature, humidity, wind speed, and solar radiation data obtained from meteorological stations, is input into the surface energy balance equation to calculate the surface sensible heat flux and latent heat flux, thereby iteratively updating the estimated surface temperature. The updated surface temperature and soil water content are then substituted into the radiative transfer model adjusted based on vegetation cover to calculate the equivalent reflectance in the visible and near-infrared bands. This iterative process converges on the condition of minimizing the residual between the calculated reflectance and the reflectance observed on a cloudless day.

[0015] Furthermore, the training of the shared feature encoder and the specific feature encoder in the feature decoupling coding network employs a joint optimization and separation loss function strategy. The total loss function consists of three parts: soil moisture prediction regression loss, domain adversarial loss for shared features, and reconstruction loss for specific features. The regression loss ensures that the decoupled feature combination can accurately predict soil moisture; the domain adversarial loss is implemented through a gradient inversion layer, forcing the shared features to lose the ability to distinguish between different data sources; the reconstruction loss requires that the shared features and specific features be re-decoded to recover the original input spectrum, ensuring the integrity of the feature decoupling process and avoiding information loss.

[0016] Furthermore, the attention weight generation network in the spatiotemporal attention fusion submodule is a lightweight convolutional neural network. This network takes the concatenated multi-scale features of the current time step as input and outputs an attention weight map with the same spatial size as the input feature map. This weight map is normalized, with each position having a value between 0 and 1, representing the importance of that spatial position in the current inference. This weight is then element-wise multiplied with the original features to achieve adaptive weighting.

[0017] Furthermore, the dynamic threshold calculation submodule integrates crop growth stage information by incorporating a built-in database of standard crop growth stage water requirement curves. This database stores the suitable soil moisture ranges for different major crops at each growth stage. Based on the crop type and sowing date input by the user, the submodule automatically matches the current growth stage and uses the corresponding water requirement curve as a reference for calculating historical percentiles, making the warning threshold more consistent with agronomic principles.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0019] 1. This invention creatively combines the cloud-penetrating ability of passive microwave remote sensing with the high spatial resolution advantage of optical remote sensing through a data fusion and reconstruction module. When optical data is missing, a pseudo-optical feature is generated using a microwave-optical coupling inversion method based on a physical model, and then smoothly completed using a sequence deep learning model. This fundamentally solves the problem of monitoring data sequence interruption caused by weather interference, ensuring the continuous and stable operation of soil moisture monitoring under any weather conditions, and significantly improving the reliability and timeliness of the system.

[0020] 2. This invention proposes a model architecture for separating universal and regional knowledge from limited labeled data through an adaptive feature decoupling and transfer learning module. By forcing the learning of domain-invariant features through domain adversarial training, the core model grasps the essential spectral response law of soil moisture. When facing new farmland areas, only a small amount of labeled data is needed to fine-tune the specific feature branches for rapid adaptation, greatly reducing the dependence on large amounts of labeled data. This effectively overcomes the bottleneck of scarce labeled data in agricultural scenarios and significantly enhances the generalization ability and practical deployment efficiency of the monitoring model in different regions and crop types.

[0021] 3. This invention achieves the organic fusion of multi-resolution remote sensing information through a multi-scale spatiotemporal collaborative reasoning module, employing a feature pyramid and spatiotemporal attention mechanism. The system can simultaneously utilize the detailed information of high-resolution data and the temporal consistency information of low-resolution data, adaptively balancing the contributions of different information sources in different scenarios through the attention mechanism, ultimately generating a soil moisture product that possesses both high spatial detail and maintains temporal continuity, thereby improving the accuracy and spatial representation of monitoring results.

[0022] 4. This invention upgrades simple threshold alarms to intelligent decision support through a dynamic threshold early warning and tracing module. The threshold is dynamically adjusted according to the crop growth stage, better aligning with actual agricultural production. After anomaly detection, automatic spatial clustering and multivariate correlation analysis are performed, accurately reporting the location and extent of the anomaly, and quickly diagnosing the meteorological, irrigation, or topographical causes. This provides agricultural technicians with direct action guidelines, realizing a closed-loop intelligent service from monitoring to early warning to diagnosis, greatly enhancing the system's decision support value. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention;

[0024] Figure 2 This is a schematic diagram of the core principle framework of the adaptive feature decoupling and transfer learning module in this invention;

[0025] Figure 3 This is a logical flowchart of the data fusion and reconstruction module in this invention;

[0026] Figure 4 This is a logical flow diagram of the multi-scale spatiotemporal collaborative reasoning module in this invention;

[0027] Figure 5 This is a logical flow diagram of the dynamic threshold early warning and tracing module in this invention. Detailed Implementation

[0028] The overall technical architecture of the intelligent monitoring and early warning system for farmland soil moisture based on multispectral remote sensing proposed in this invention is shown in the attached figure. Figure 1 As shown in the figure, the system consists of four core functional modules cascaded in sequence: a data fusion and reconstruction module, an adaptive feature decoupling and transfer learning module, a multi-scale spatiotemporal collaborative reasoning module, and a dynamic threshold early warning and source tracing module. Each module communicates with the others through strictly defined data interfaces, ensuring that the entire system logically forms a complete closed-loop process from raw remote sensing data input to intelligent decision output. The following will be combined with the attached... Figure 1 To be continued Figure 5 This section provides a detailed implementation description of each functional module of the system, expanding upon it layer by layer.

[0029] First, the data fusion and reconstruction module, as the system's front-end processing unit, undertakes the crucial task of ensuring the spatiotemporal continuity of the input data. Please refer to the appendix. Figure 3This module's input connects to multiple heterogeneous data streams, including multispectral remote sensing imagery from optical satellites or UAV platforms, passive microwave remote sensing brightness and temperature data from meteorological satellites, precipitation and temperature observation sequences from ground meteorological stations, and terrain elevation and slope information from digital elevation models. All this data, upon entering the module, first passes through a multi-source data alignment submodule. This submodule uses a unified geographic coordinate system (e.g., WGS84) and standard timestamps (accurate to the hour) as a reference to perform spatiotemporal registration on all input data. Specifically, for the spatial dimension, bilinear interpolation or nearest-neighbor resampling methods are used to reproject and resample all data layers to the same regular grid, typically with a grid resolution of 10 meters. For the temporal dimension, the acquisition time of the optical remote sensing image is used as the anchor point, and other auxiliary data are aligned to the same time series node through temporal interpolation (e.g., cubic spline interpolation). After this step, all data is organized into a four-dimensional tensor structure with dimensions of [time step, spatial height, spatial width, number of feature channels], where the number of feature channels is equal to the sum of the total number of channels in all input data sources.

[0030] After spatiotemporal alignment, the data is fed into the physical model-driven reconstruction submodule. The core of this submodule lies in executing a physical inversion process that couples the radiative transfer equation and the surface energy balance equation. Its goal is to generate pseudo-optical features that are physically consistent with the true optical reflectance at time points missed by optical remote sensing due to cloud cover. This process first uses passive microwave brightness and temperature data to invert an initial estimate of the surface soil volumetric water content (typically referring to a depth of 0 to 5 cm) using the Dixon model. This initial value is then input into the surface energy balance equation along with parameters such as near-surface air temperature, relative humidity, wind speed, and solar shortwave radiation flux density from meteorological stations. This equation describes how net radiative flux is distributed as sensible heat flux, latent heat flux, soil heat flux, and changes in vegetation canopy heat storage.

[0031] The estimated surface temperature can be updated by solving this equation numerically and iteratively. The updated surface temperature and soil moisture content are used as inputs and substituted into a radiative transfer model corrected for vegetation cover. This model considers soil background reflection, multiple scattering from the vegetation canopy, and atmospheric path effects, ultimately calculating the theoretical surface reflectance in the visible light (e.g., red band) and near-infrared bands. This iterative process continues until the sum of squared residuals between the calculated reflectance and the reflectance actually observed on a nearby cloudless day is less than a preset convergence threshold (e.g., 0.001). At this point, the output reflectance is considered the pseudo-optical reflectance feature for that missing moment.

[0032] The pseudo-optical feature points generated by the physical model are typically discrete and may contain computational noise. Therefore, these feature points are fed into a sequence completion and smoothing submodule. This submodule receives a mixed time series containing real optical observations and pseudo-observations generated by the physical model. Internally, this submodule deploys a temporal convolutional network with a receptive field covering at least seven consecutive time steps. This network is constructed by stacking one-dimensional causal convolutional layers, each containing a dilated convolution operation to expand the temporal receptive field while preventing information leakage. The training objective of the network is to minimize the L2 norm loss between the output sequence and the real observation sequence under cloudless conditions. During the inference phase, the network performs end-to-end smoothing on the mixed sequence containing pseudo-feature points, effectively suppressing high-frequency oscillations and discontinuities that may be introduced during the physical model iteration process. The final output is a spatiotemporally continuous, smooth, and physically consistent pseudo-optical feature data cube with a temporal resolution of once per day, a spatial resolution of 10 meters, and spectral bands including blue, green, red, near-infrared, and two short-wave infrared bands, totaling six channels.

[0033] This continuous feature data cube is then fed into the adaptive feature decoupling and transfer learning module. Please refer to the appendix. Figure 2 This module aims to address the severe scarcity of labeled data in agricultural remote sensing scenarios and improve the model's generalization ability in new regions. Its core is a feature decoupling encoding network, employing a dual-branch parallel encoder structure. The shared feature encoder branch consists of residual blocks composed of multiple convolutional-batch normalization-activation functions. Its design goal is to extract universal features from the input spectrum that are strongly correlated with the physical state of soil moisture but robust to changes in crop type, soil texture, and planting patterns. These features are mainly manifested as nonlinear response patterns of near-infrared and short-wave infrared reflectance with soil moisture variations, such as changes in the depth and width of water absorption valleys. The specific feature encoder branch adopts a similar network structure, but its learning objective is to capture specific information strongly correlated with the local environment, such as the spectral reflectance characteristics of specific crop canopies and the background reflectance baseline of the local soil. The outputs of the two encoders are forcibly separated in the feature space, forming a shared feature tensor and a specific feature tensor, respectively.

[0034] To ensure the domain invariance of shared features, the system introduces a domain discrimination and alignment submodule. This submodule contains an independent domain classifier, whose input is the shared feature tensor and output is the predicted data source domain label (e.g., "North China Plain wheat region" or "Yangtze River Basin rice region"). During training, this domain classifier and the shared feature encoder undergo adversarial joint optimization. Specifically, during backpropagation, the gradient sign of the domain classification loss is inverted through a gradient inversion layer before being fed back to the shared feature encoder. This mechanism forces the feature representation generated by the shared feature encoder to maximally confuse the domain classifier, making it unable to accurately determine the feature source, thereby stripping away regionally specific information and retaining only universal knowledge related to the essence of soil moisture.

[0035] When the system needs to be deployed to entirely new farmland areas, the small-sample fine-tuning submodule is activated. At this point, the user only needs to provide a very small number of sample points within the target area (e.g., no more than 5 per plot) with high-precision ground-measured soil moisture values ​​(obtained via time-domain reflectometry or oven-drying). During the fine-tuning phase, all parameters of the shared feature encoder are frozen, allowing only the specific feature encoder and the subsequent regression head used for moisture prediction to update their parameters. Because the shared features have strong universality, the fine-tuning process requires only a few iterations (typically less than 100 steps) to allow the model to quickly adapt to the spectral-moisture mapping of the new area, thus achieving high-precision predictions even with extremely limited labeled data.

[0036] After feature decoupling and transfer learning, the universal features output by the system are fed into the multi-scale spatiotemporal collaborative inference module. Please refer to the appendix. Figure 4 The goal of this module is to fuse remote sensing information at different spatial resolutions to generate a soil moisture distribution map that combines high spatial detail with temporal coherence. The multi-scale feature pyramid submodule first receives universal features (spatial resolution of 10 meters) from the adaptive feature decoupling and transfer learning module, and simultaneously receives native multispectral features from low-resolution but more temporally continuous remote sensing data sources (e.g., MODIS, spatial resolution of 250 meters). This submodule generates feature maps at multiple scales through downsampling operations: 10 meters, 30 meters, 90 meters, and 250 meters. For each scale, semantic information is extracted through several convolutional layers, forming a bottom-up, semantically progressively enhanced feature pyramid.

[0037] Subsequently, the spatiotemporal attention fusion submodule intelligently weights and fuses features at various scales within the pyramid. This submodule contains a lightweight convolutional neural network specifically designed to generate spatiotemporal attention weights. The network takes a tensor, spatially concatenated from all scale features at the current time step, as input, passes it through two convolutional layers and a sigmoid activation function, and outputs an attention weight map with spatial dimensions identical to the highest resolution (10 meters) feature map. Each pixel value in this weight map is between 0 and 1, representing the importance of that spatial location in humidity inversion at the current time step. Furthermore, this submodule includes a temporal attention mechanism, calculating the contribution weight of each historical time point to the current prediction by performing a one-dimensional convolution operation on the feature sequence from the past 7 days. Finally, the features at each scale are weighted and summed according to the spatial and temporal attention weights to obtain a unified fusion feature representation rich in multi-scale contextual information.

[0038] The fused features are input into the soil moisture inversion submodule. This submodule consists of a three-layer fully connected neural network, each layer containing 128 neurons, and employs the ReLU activation function. The network output is a single-channel predicted value of soil volumetric water content, in cubic meters per cubic meter. During the training phase, the network uses measured ground-level soil moisture as a supervision signal and employs root mean square error as the loss function for end-to-end optimization. Because the input features have been fused and enhanced to a high standard through the aforementioned modules, this inversion submodule can output a high-precision spatial distribution map of soil moisture while maintaining computational lightweightness, with a spatial resolution of 10 meters and a temporal resolution of once per day.

[0039] Finally, the dynamic threshold early warning and traceability module receives the aforementioned high-precision soil moisture products and provides intelligent decision support services. Please refer to the appendix. Figure 5 The module begins with a dynamic threshold calculation submodule. This submodule contains a database of standard crop growth stage water requirement curves, storing suitable soil moisture ranges for major crops such as rice, wheat, and corn at various growth stages, including sowing, tillering, jointing, heading, grain filling, and maturity. Based on the crop type and sowing date input by the user during initialization, the system automatically calculates the current growth stage. Building on this, the submodule retrieves the historical soil moisture sequence of the target field over the past 30 days and calculates the empirical cumulative distribution function of this sequence at the current growth stage. The 10th percentile of this distribution is set as the drought warning threshold, and the 90th percentile as the over-water warning threshold. These two thresholds are updated daily to ensure they remain synchronized with the actual water requirements of the crops.

[0040] The anomaly detection and spatial clustering submodule monitors the comparison between the current soil moisture value and a dynamic threshold in real time. If the soil moisture value of a pixel is below the drought threshold or above the excessive moisture threshold for two consecutive monitoring periods (i.e., 2 days), the pixel is marked as an anomaly candidate. Subsequently, this submodule uses the DBSCAN spatial clustering algorithm to perform cluster analysis on all anomaly candidate points. The algorithm's neighborhood radius is set to 30 meters (i.e., 3 pixels), and the minimum number of cluster points is set to 9 (i.e., 3×3 pixel blocks). Through this operation, discrete anomaly points are aggregated into contiguous anomaly patches with clear boundaries. The system further calculates the geometric attributes of each patch, including total area (in square meters), average moisture deviation (the absolute difference relative to the threshold), and shape index (the ratio of perimeter to area), to quantify the severity and spatial morphology of the anomaly.

[0041] Once an abnormal patch is identified, the multivariate correlation source tracing submodule is immediately activated. This submodule simultaneously retrieves multi-source auxiliary data for the period of the anomaly and the preceding 7 days, including the original multispectral reflectance, cumulative precipitation, average temperature, topographic slope map, and vector distribution map of irrigation facilities before reconstruction. Within the abnormal patch and its 500-meter buffer zone, the system calculates the statistical mean, variance, and trend of the aforementioned variables. Through comparative analysis, the system identifies variables that significantly deviate from the normal area within the abnormal patch. For example, if the cumulative precipitation within the abnormal patch is more than 50% lower than the surrounding area, and irrigation facilities are sparsely distributed, the dominant cause is determined to be scarce precipitation and insufficient irrigation; if the abnormal patch is located in a low-lying area with a slope of less than 1%, and the surrounding area has good drainage, the cause is determined to be waterlogging due to the terrain. Finally, the system generates a structured source tracing report, clearly listing the location, area, severity, and one or two most likely dominant causes of the anomaly, providing agricultural technicians with precise intervention guidelines.

[0042] In summary, this embodiment constructs a complete solution from data acquisition and intelligent modeling to decision-making and early warning through the close collaboration of four modules. This solution not only overcomes the data interruption problem caused by weather interference but also effectively alleviates the bottleneck of scarce labeled data. Furthermore, through multi-scale fusion and dynamic traceability mechanisms, it significantly improves the accuracy, continuity, and practical value of soil moisture monitoring.

[0043] Based on the aforementioned embodiments, this embodiment optimizes and adjusts the implementation details of some sub-modules for specific application scenarios to adapt to more complex or resource-constrained deployment environments.

[0044] In the data fusion and reconstruction module, the radiative transfer model used in the physical model-driven reconstruction submodule can be dynamically selected based on the regional vegetation cover. In bare soil or sparsely vegetated areas with vegetation cover below 30%, the system employs a simplified soil-atmosphere bidirectional reflectance distribution function model. This model considers only soil background reflectance and single atmospheric scattering, resulting in high computational efficiency and suitability for large-scale rapid processing. In dense canopy areas with vegetation cover above 70%, the system switches to the more complex PROSAIL model. This model couples the PROSPECT leaf optical model with the SAIL canopy radiative transfer model, accurately simulating the influence of leaf internal structure, biochemical components, and canopy three-dimensional structure on spectral reflectance. Vegetation cover data is derived from the normalized vegetation index calculated from the multispectral imagery itself, and regions are divided using preset thresholds (e.g., NDVI=0.3 and NDVI=0.7). This adaptive model selection strategy effectively balances computational overhead while ensuring reconstruction accuracy.

[0045] In the adaptive feature decoupling and transfer learning module, the fine-tuning strategy of the few-sample fine-tuning submodule can be adjusted hierarchically based on the number of available labeled samples. When the number of labeled samples available for the target region is greater than or equal to 10, the system adopts a standard fine-tuning strategy, i.e., fixing the shared feature encoder and only fine-tuning the specific feature encoder and regression head. When the number of labeled samples is between 3 and 9, the system enables a semi-supervised fine-tuning strategy: during the fine-tuning process, a consistency regularization loss is introduced, requiring the model to maintain consistent prediction results for the same input after applying slight data augmentation (such as random scaling of spectral channels ±5%), thereby utilizing a large amount of unlabeled data to improve model stability. When the number of labeled samples is less than 3, the system completely skips the fine-tuning step, directly using the general model trained on the source domain for prediction, and marking the output results with a "low confidence" warning for users to refer to with caution. This hierarchical strategy enables the system to still provide usable monitoring services under extreme data scarcity conditions.

[0046] In the multi-scale spatiotemporal collaborative inference module, the attention weight generation network of the spatiotemporal attention fusion submodule can be compressed according to the computing power of the deployment platform. On cloud servers, this network can adopt a full convolutional structure; while on edge computing devices (such as embedded gateways deployed in farmland), the network is replaced with a lightweight version composed of depthwise separable convolutions, reducing the number of parameters to 1 / 5 of the original and increasing the inference speed by more than 3 times. Despite the simplified network structure, the lightweight network can still generate effective attention weights, ensuring fusion quality, because its input features themselves already contain rich semantic information.

[0047] In the dynamic threshold early warning and source tracing module, the variable comparison analysis in the multivariate correlation source tracing submodule incorporates statistical significance testing. Specifically, the system performs a two-sample t-test on the differences in the means of each variable between the abnormal patch and the surrounding normal area. Only when the p-value is less than 0.05 is the variable considered a statistically significant potential cause. Furthermore, the system calculates the Pearson correlation coefficient between each variable and the degree of soil moisture anomaly, prioritizing variables with an absolute correlation coefficient greater than 0.6 as high-probability causes. This source tracing method based on statistical inference significantly reduces the false positive rate and improves the reliability of diagnostic conclusions.

[0048] Through the above optimizations, this embodiment enhances the system's adaptability and robustness under different environmental, data, and hardware conditions while maintaining the core functionality, further expanding its application boundaries.

Claims

1. A smart monitoring and early warning system for farmland soil moisture based on multispectral remote sensing, characterized in that, include: The data fusion and reconstruction module is used to fuse and spatiotemporally reconstruct multi-source heterogeneous data when optical remote sensing data is missing due to cloud cover, in order to generate a continuous pseudo-optical feature sequence. An adaptive feature decoupling and transfer learning module is used to decouple transferable universal spectral features and region-specific features from limited labeled data, and drive the model to quickly adapt in new, unlabeled scenarios; the input of the adaptive feature decoupling and transfer learning module is connected to the output of the data fusion and reconstruction module. A multi-scale spatiotemporal collaborative reasoning module is used to integrate remote sensing information with different spatial resolutions and temporal frequencies to perform collaborative reasoning in order to generate a soil moisture distribution map with high spatial resolution and temporal continuity; the input end of the multi-scale spatiotemporal collaborative reasoning module is connected to the output end of the adaptive feature decoupling and transfer learning module. The dynamic threshold early warning and tracing module is used to dynamically set the soil moisture early warning threshold based on historical data and real-time inference results, and to locate abnormal areas and analyze potential causes when an early warning is triggered; the input of the dynamic threshold early warning and tracing module is connected to the output of the multi-scale spatiotemporal collaborative inference module. The data fusion and reconstruction module includes a multi-source data alignment sub-module, which is used to receive multispectral remote sensing image data streams from satellite or UAV platforms, passive microwave remote sensing brightness and temperature data from meteorological satellites, precipitation and temperature data from ground meteorological stations, and terrain data from digital elevation models, and to perform spatiotemporal registration of all input data using a unified geographic coordinate system and timestamp. The data fusion and reconstruction module also includes a physical model-driven reconstruction submodule. This physical model-driven reconstruction submodule is used to perform coupled physical model calculations based on the radiative transfer equation and the surface energy balance equation based on the aligned multi-source data. It uses passive microwave brightness temperature, surface temperature and precipitation data as the main driving inputs, solves the equations to inversely derive the physical relationship between surface soil moisture content and surface reflectivity, and iteratively calculates the theoretical surface reflectivity value of the corresponding optical band at the missing time, thereby generating pseudo optical reflectivity features. The data fusion and reconstruction module also includes a sequence completion and smoothing submodule. The sequence completion and smoothing submodule is used to receive pseudo-optical feature points output by the physical model-driven reconstruction submodule, insert them into the original incomplete multispectral remote sensing time series, and use a sequence modeling method based on temporal convolutional networks to smooth the completed mixed sequence, and finally output a spatiotemporally continuous and smooth pseudo-optical feature data cube.

2. The intelligent monitoring and early warning system for farmland soil moisture based on multispectral remote sensing according to claim 1, characterized in that, The adaptive feature decoupling and transfer learning module includes a feature decoupling encoding network, a domain discrimination and alignment submodule, and a few-sample fine-tuning submodule. The feature decoupling encoding network receives a continuous feature data cube from the data fusion and reconstruction module as input and adopts a dual-branch encoder structure. One branch is a shared feature encoder, which is used to extract universal spectral features that are strongly correlated with the physical mechanism of soil moisture and are not sensitive to changes in crop type and soil texture. The other branch is a specific feature encoder, which is used to extract specific features that are related to the local crop canopy structure and soil background reflectance. The domain discrimination and alignment submodule is connected to the output of the shared feature encoder and includes a domain classifier. It forces the shared feature encoder to learn feature representations that can confuse the domain classifier through adversarial training to ensure the domain invariance of the shared features. The small sample fine-tuning submodule is used to receive a small amount of sample data with high-precision ground-measured soil moisture values ​​labeled in the target area when the model is deployed to a new farmland area. It fixes the parameters of the shared feature encoder and performs rapid fine-tuning only on specific feature encoders and subsequent prediction network layers.

3. The intelligent monitoring and early warning system for farmland soil moisture based on multispectral remote sensing according to claim 2, characterized in that, The multi-scale spatiotemporal collaborative reasoning module includes a multi-scale feature pyramid sub-module, a spatiotemporal attention fusion sub-module, and a soil moisture inversion sub-module. The multi-scale feature pyramid sub-module is used to simultaneously receive universal features from the adaptive feature decoupling and transfer learning module and native multispectral features from remote sensing data sources of different resolutions, and construct a feature pyramid containing feature maps of multiple spatial scales through a series of convolution and upsampling operations. The spatiotemporal attention fusion submodule is used to perform weighted fusion of features at each scale in the feature pyramid. It includes a spatiotemporal attention mechanism, which automatically calculates the importance weight of each spatial location in the time series and each feature scale in the space. The soil moisture inversion submodule is used to receive the unified feature representation after spatiotemporal attention weighted fusion, and to map the fused features into the final soil moisture value through a fully connected neural network or a lightweight convolutional neural network.

4. The intelligent monitoring and early warning system for farmland soil moisture based on multispectral remote sensing according to claim 3, characterized in that, The dynamic threshold early warning and tracing module includes a dynamic threshold calculation submodule, an anomaly detection and spatial clustering submodule, and a multivariate correlation tracing submodule; The dynamic threshold calculation submodule is used to receive the soil moisture inversion results for the current and historical periods, as well as the crop growth stage information and irrigation records for the same period. For each field unit, it calculates the historical soil moisture percentile distribution for that period based on the water requirement pattern of the current crop growth stage, sets the 10th percentile as the drought warning threshold, and sets the 90th percentile as the over-wet warning threshold. The anomaly detection and spatial clustering submodule is used to compare the current soil moisture value with the dynamic threshold in real time. When the soil moisture value of a certain area is continuously lower than the drought threshold or higher than the excessive moisture threshold for more than two monitoring cycles, an early warning is triggered. The spatial clustering algorithm is used to cluster the pixels that trigger the warning continuously to form a continuous abnormal patch. The multivariate correlation tracing submodule is used to synchronously retrieve multi-source data of the area during the abnormal period and the preceding period after the abnormal patch is formed. By comparing and analyzing the statistical differences of these variables between the abnormal patch and the surrounding normal area, a tracing report is generated.

5. The intelligent monitoring and early warning system for farmland soil moisture based on multispectral remote sensing according to claim 4, characterized in that, The solution process for the coupled radiative transfer equation and surface energy balance equation in the physical model-driven reconstruction submodule employs a numerical iterative method. First, the initial value of surface soil volumetric water content is obtained by inverting passive microwave data using the Dixon model. This initial value, along with near-surface air temperature, humidity, wind speed, and solar radiation data obtained from meteorological stations, is input into the surface energy balance equation to calculate the surface sensible heat flux and latent heat flux, thereby iteratively updating the surface temperature estimate. The updated surface temperature and soil water content are then substituted into the radiative transfer model adjusted based on vegetation cover to calculate the equivalent reflectance in the visible and near-infrared bands. This iterative process converges on the condition of minimizing the residual between the calculated reflectance and the reflectance observed on a cloudless day.

6. The intelligent monitoring and early warning system for farmland soil moisture based on multispectral remote sensing according to claim 5, characterized in that, The training of the shared feature encoder and the specific feature encoder in the feature decoupling coding network adopts a joint optimization and separation loss function strategy. The total loss function consists of three parts: soil moisture prediction regression loss, domain adversarial loss of shared features, and reconstruction loss of specific features. The regression loss ensures that the decoupled feature combination can accurately predict soil moisture. The domain adversarial loss is implemented through a gradient inversion layer, which forces the shared features to lose the ability to distinguish between different data sources. The reconstruction loss requires that the shared features and specific features be re-decoded to recover the original input spectrum.

7. The intelligent monitoring and early warning system for farmland soil moisture based on multispectral remote sensing according to claim 6, characterized in that, The attention weight generation network in the spatiotemporal attention fusion submodule is a lightweight convolutional neural network. The network takes the concatenation of multi-scale features at the current time as input and outputs an attention weight map with the same spatial size as the input feature map. The weight map is normalized, and the value of each position is between 0 and 1, which indicates the importance of the spatial position in the current inference. The weight is multiplied element-wise with the original features to achieve adaptive weighting.

Images