A method and system for monitoring the growth of crops in saline-alkali soil

Abstract

The application provides a kind of saline-alkali soil crop growth monitoring method and system, method includes: construct graph-grid hybrid topology, use set Kalman filter to satellite, unmanned aerial vehicle, sensor and so on multi-source observation assimilation to unified grid;Build physical information spatiotemporal graph convolution network, embed space variation physical constraint to predict water salt dynamic;Fusion sunlight induced chlorophyll fluorescence and morphological index, through transfer learning to establish variety self-adaptive photosynthesis-morphology combined stress diagnosis model;Grid level digital twin is constructed, coupled water salt-crop-stress process and introduces sodium ion stress memory factor, deduces yield and biomass;With digital twin as forward simulator, model predictive control rolling optimization is used, executable precision irrigation and spraying prescription map is generated, and monitoring-decision-execution closed loop is formed.The application uses the above-mentioned saline-alkali soil crop growth monitoring method and system, realizes stress early warning, variety rapid adaptation and water-saving stable yield collaborative optimization.

Description

Technical Field

[0001] This invention relates to the field of crop monitoring in saline-alkali land, and in particular to a method and system for monitoring crop growth in saline-alkali land. Background Technology

[0002] As an important reserve of arable land, saline-alkali land faces multiple challenges in monitoring crop growth. Soil water and salt transport processes are complex and spatially heterogeneous. Traditional monitoring methods relying on limited sampling points are insufficient to reflect the true water and salt dynamics at the field scale, resulting in inadequate diagnosis of crop salt stress. Often, it is only detected after significant morphological changes in plants, delaying optimal intervention. In existing technologies, patent CN120875186B proposes a method for water and salt prediction and crop growth assessment based on a 3D-CNN spatiotemporal cube and a segmented salt tolerance model, achieving a significant leap from static sampling to spatiotemporal dynamic monitoring.

[0003] However, the existing technology still has the following shortcomings in practice: (1) When fusing multi-source data such as sensors, drones, and satellites, there is a lack of an effective assimilation mechanism for irregular spatial distribution and asynchronous temporal sequence, making it difficult to generate a spatiotemporally continuous physical consistent dataset; (2) The use of fixed grid 3D-CNN to extract water and salt features has poor adaptability to sensor missing and topological changes, and the physical constraints rely on preset fixed diffusion coefficients, which cannot reflect the impact of soil texture spatial variation on water and salt transport; (3) Crop stress diagnosis mainly relies on morphological indicators such as leaf area index, and stress detection lags behind photosynthetic physiological damage, and the salt-tolerant model varieties have poor adaptability, requiring extensive recalibration when changing or introducing new varieties; (4) Monitoring and decision-making are disconnected, and the generated agricultural plans are open-loop suggestions that cannot be automatically executed and feedback corrected, making it difficult to achieve a complete closed loop of monitoring-decision-execution. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for monitoring crop growth in saline-alkali land, in order to solve the problems of insufficient assimilation of multi-source data, weak physical consistency and poor adaptability of water and salt prediction, delayed stress diagnosis and difficulty in variety migration, and open-loop disconnect between decision-making and execution in the existing technology.

[0005] To achieve the above objectives, the present invention provides a method for monitoring crop growth in saline-alkali land, comprising the following steps: Step S1: Acquire multi-source sensing data, construct a graph-grid hybrid topology, and use ensemble Kalman filtering to assimilate heterogeneous multi-scale observations into a unified grid, outputting a spatiotemporally continuous assimilation dataset. Step S2: Using the spatiotemporal continuous assimilation dataset of S1 as input, and following the graph-lattice hybrid topology of S1, construct a physical information spatiotemporal graph convolutional network model. By introducing a physical consistency loss term, extract water and salt transport features and predict the spatiotemporal dynamics of water and salt. This prediction field serves as the driving data for the water and salt sub-model of S4. Step S3: Using the spatiotemporal continuous assimilation dataset from S1, construct an adaptive salt stress diagnosis model for varieties through transfer learning, and output the real-time stress level and photosynthetic limitation / morphological atrophy label. Step S4: Construct a digital twin instance for the grid in S1. The water-salt sub-model updates the root zone state based on the prediction field in S2. The crop growth sub-model dynamically adjusts physiological parameters according to the stress level and type label in S3, and introduces the salt stress memory factor to deduce and output yield and biomass predictions. The prediction results provide optimization objectives and state feedback for S5. Step S5: Using the digital twin of S4 as a forward simulator, with the output prediction of S4 as the optimization objective and the salt tolerance threshold of S3 as the constraint, construct a model predictive control framework for rolling solution, generate a control prescription map and issue it for execution; after execution, it is assimilated again through multi-source perception of S1 to form a monitoring-decision-execution closed loop.

[0006] Preferably, the multi-source sensing data in step S1 includes at least: satellite synthetic aperture radar data, satellite multispectral remote sensing data, UAV multispectral data, UAV thermal infrared data, UAV sunlight-induced chlorophyll fluorescence hyperspectral data, time-series data of stratified soil parameters collected by a field soil profile sensor network, field micro-weather station data, and groundwater monitoring data; each sensor node of the field soil profile sensor network is equipped with a multi-parameter sensing unit including at least a conductivity probe, a moisture probe, a temperature probe, a pH probe, and a sodium ion selective electrode; the graph-grid hybrid topology uses each field sensor location and the center of each grid with a preset spatial resolution as graph nodes, and constructs a weighted adjacency matrix based on the spatial distance between nodes and the similarity of soil texture. When a sensor node goes offline, the corresponding node is removed from the graph topology and the adjacency matrix is ​​dynamically reconstructed, and the model still maintains its predictive ability through neighboring node propagation.

[0007] Preferably, the physical information spatiotemporal graph convolutional network model in step S2 adopts an encoder-decoder sequence-to-sequence architecture, where both the encoder and decoder contain stacked Chebyshev graph convolutional layers and gated recurrent unit layers; the physical consistency loss term calculates the residual between the physical expectation value and the model prediction value of salt flux between adjacent grids based on the spatially variable water diffusion coefficient and salt diffusion coefficient dynamically estimated based on soil texture attributes for each grid cell; the spatially variable water diffusion coefficient and salt diffusion coefficient are obtained by dynamically estimating the sand content, silt content, and clay content of each grid cell by inputting them into the soil transfer function model.

[0008] Preferably, the construction process of the variety adaptive salt stress diagnostic model in step S3 is as follows: First, a basic model of salt tolerance response curves for multiple crop varieties is established, and a piecewise exponential nonlinear function is used to characterize the multi-threshold piecewise response law of biomass or yield of different crop varieties to root layer salinity; for the target crop variety, a small amount of measured biomass-salt correspondence data obtained from field labeling experiments is used to perform transfer learning on the basic model through a model-independent meta-learning strategy or parameter fine-tuning strategy, freezing the shallow network parameters, and training only the piecewise threshold layer and the piecewise decay coefficient layer. The update automatically determines the segmented salt stress threshold and segmented function parameters for this variety. The output process of the stress level and type label is as follows: Based on the canopy fluorescence yield and actual quantum efficiency of photosystem II obtained from the UAV's sunlight-induced chlorophyll fluorescence hyperspectral data, a photosynthetic stress index is constructed; Based on the normalized vegetation index and leaf area index obtained from the UAV's multispectral data, a morphological stress index is constructed; Using root soil salinity, the photosynthetic stress index, and the morphological stress index as inputs, the real-time stress level and stress type label for each grid cell are output.

[0009] Preferably, the joint determination rule for the stress level is as follows: when the photosynthetic stress index exceeds the first threshold for two consecutive days or the morphological stress index exceeds the second threshold, it is determined to be mild stress; when the photosynthetic stress index and the morphological stress index exceed the third threshold and the fourth threshold at the same time, it is determined to be moderate stress; and when any index exceeds the corresponding severe threshold, it is determined to be severe stress.

[0010] Preferably, the mathematical form of the improved salt stress memory factor described in step S4 is: ; in, Let be the cumulative sodium ion equivalent toxicity memory at time t. The attenuation coefficient is... For a historic moment The equivalent electrical conductivity of sodium ions in the root zone is obtained by converting the soil electrical conductivity to the sodium adsorption ratio. The threshold value for sodium ion toxicity in crops is defined as follows: the sensitivity coefficient for the growth stage is dynamically assigned by the crop growth sub-model based on the accumulated temperature to determine the current phenological stage. The sensitivity coefficient for the seedling stage ranges from 0.6 to 0.8, for the heading stage from 0.4 to 0.6, and for the maturity stage from 0.2 to 0.3.

[0011] Preferably, the objective function of the model predictive control optimization framework in step S5 is: ; in, To predict the control window length, Let be the predicted yield influencing factor at time t. Let t be the total irrigation water used at time t. Let t be the energy consumption for irrigation and spraying operations. The total amount of deep seepage salt leaching predicted at time t. , , , These are configurable weighting coefficients.

[0012] Preferably, the yield prediction in step S4 is achieved using an ensemble learning model that combines temporal convolutional networks and gradient boosting trees. The model uses daily biomass throughout the entire growth period, cumulative stress at each stage, final leaf area index, and root salinity statistics as input features to output a grid-level predicted yield and its confidence interval. When the probability that the predicted yield is lower than the preset minimum yield line exceeds a preset probability threshold, a corresponding level of probability warning is triggered.

[0013] Preferably, the method further includes step S6, post-season adaptive evolution: After crop harvest, an experience pool is constructed by collecting all accumulated observation data, assimilation status data, digital twin simulation data, stress diagnosis records, and closed-loop management execution effect data throughout the entire growth period; an offline reinforcement learning algorithm is used, with the final measured yield and resource consumption as reward signals, to jointly optimize the hyperparameters of the physical information spatiotemporal graph convolutional network model, the segmented thresholds of the variety adaptive salt stress diagnosis model, the variety parameters of the crop growth sub-model in the digital twin, and the weight coefficients of the model prediction controller; the optimized parameters and model are deployed to the next growing season to achieve year-by-year adaptive increase in monitoring accuracy.

[0014] A crop growth monitoring system for saline-alkali land includes: The perception layer is used to acquire multi-source perception data, and includes at least a satellite data receiving module, a UAV multi-payload data acquisition module, a field soil profile sensor network, a micro weather station and a groundwater monitoring module. An edge computing layer, deployed at a field base station, includes a data preprocessing unit and an edge assimilation unit, used to execute the ensemble Kalman filter assimilation algorithm described in step S1 of claim 1, and output assimilated data stream to the cloud; The cloud-based digital twin platform includes: a water-salt inversion engine, which deploys the physical information spatiotemporal graph convolutional network model described in step S2 of claim 1; a stress diagnosis engine, which deploys the variety-adaptive salt-stress diagnosis model described in step S3 of claim 1; a digital twin inference engine, which runs the grid-level digital twin instance described in step S4 of claim 1; and a model predictive control decision engine, which executes the model predictive control optimization described in step S5 of claim 1 to generate a daily zoned precise control prescription map. The execution layer, including the variable irrigation control system and the UAV variable spraying platform, receives the precise control prescription map and executes it automatically; It also includes a data storage and visualization platform for archiving data throughout the reproductive period, 3D visualization, and early warning information dissemination.

[0015] Therefore, the present invention employs the above-mentioned method and system for monitoring crop growth in saline-alkali land, and the technical effects are as follows: Deep fusion of multi-source data: Graph-grid hybrid topology and EnKF assimilation solve the problem of fusion of heterogeneous, multi-scale, and irregular sensor data, outputting a spatiotemporally continuous and physically consistent grid dataset, laying a high-quality data foundation for subsequent analysis.

[0016] Improved accuracy and robustness of water and salt prediction: The PI-STGCN model is naturally adapted to the dynamic changes in sensor topology. After embedding spatial variation physical constraints, the prediction of water and salt transport is more in line with the actual situation in the field, and the capture of secondary salinization trends is more accurate.

[0017] Early stress diagnosis and strong variety adaptability: The integration of the SIF photosynthetic stress index can achieve early warning 3-5 days earlier than morphological indicators; the transfer learning mechanism allows new varieties to complete model adaptation with only 3-5 sets of measured data, greatly reducing the system expansion cost.

[0018] Dynamic simulation and scientific decision-making throughout the entire growth period: The digital twin couples the water-salt-crop-stress process hourly, and the improved sodium ion memory factor and autonomous phenological triggering are closer to the real physiological laws; MPC closed-loop control directly transforms the monitoring results into executable and optimizable precision operation prescriptions, achieving the synergistic optimization of water conservation, stable yield and salt reduction. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the "end-edge-cloud" collaborative architecture of the system of the present invention; Figure 3 This is a schematic diagram of a graph-lattice hybrid topology. Figure 4 This is a structural diagram of a spatiotemporal graph convolutional network model for physical information. Figure 5 Flowchart for the diagnosis of adaptive photosynthetic-morphological combined stress in varieties; Figure 6 This is a diagram of the digital twin derivation and MPC closed-loop control framework. Detailed Implementation

[0020] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0022] Example 1 The study focused on a 500-mu (approximately 33 hectares) saline-alkali land corn-growing area with meadow alkaline soil, pH 8.5-9.5, and significant spatial differences in salinity.

[0023] like Figure 1 As shown, a method for monitoring crop growth in saline-alkali land includes the following steps: Step S1: Adaptive Assimilation and Fusion of Multi-Source Heterogeneous Data Soil profile sensor nodes were deployed in a 100m grid across the target plot. Each node contained four layers of EC (ecological sensors) for moisture and temperature (10 / 30 / 50 / 70cm). Sodium-selective electrodes were installed at 20% of representative locations. Data was uploaded to the edge computing node every 10 minutes via a LoRa gateway. A micro weather station and groundwater monitoring wells were also deployed. Sentinel-1 GRDSAR data (10m resolution, 6-day revisit) and Sentinel-2 multispectral data (10m resolution, 5-day revisit) were acquired periodically. A drone equipped with a multispectral, thermal infrared, and hyperspectral SIF (Synthetic Infrared) pod flew once each during the seedling, jointing, and heading stages, increasing to twice a week during stress warning periods. The flight altitude was 80m, acquiring 0.1m resolution SIF data and 5cm resolution multispectral data.

[0024] At the edge computing node, the monitoring area is discretized into a 30m×30m grid, constructing a graph-grid hybrid topology, such as... Figure 3 As shown: The graph nodes are the true coordinates of each sensor and the center points of all grid cells. Sensor nodes are connected to grid nodes within a 45m radius, and grid nodes are connected by 8-neighborhoods. The edge weights are determined by the product of the spatial distance Gaussian kernel and the soil texture connectivity similarity. Soil moisture (0-20 / 20-40 / 40-60cm), salinity, and leaf area index (LAI) of each grid point are used as state vectors. Corresponding observation operators are constructed for different types of observations, including SAR-retrieved soil moisture, Sentinel-2-retrieved NDVI / LAI, UAV SIF-retrieved LAI and ΦPSII, and sensor-measured EC and moisture. All valid observations are collected daily at 00:00 UTC, and EnKF (set size 100) is run to update the state of each grid cell, generating a spatiotemporally continuous assimilation dataset.

[0025] Step S2: Dynamic monitoring of water and salt content via a spatiotemporal graph network based on physical information like Figure 4As shown, a PI-STGCN model is constructed, with the input being the daily stratified water-salt grid data of the past 8 days output by S1. The encoder contains 2 layers of Chebyshev graph convolution (K=3) + 2 layers of GRU (128 hidden units), and the decoder has a symmetric structure to output predictions for the next 14 days. Using the Rosetta soil transfer function, substituting the sand / silt / clay content of each grid, the saturated hydraulic conductivity and water characteristic curves are dynamically estimated, and the spatially variable water diffusion coefficient Dw and salt diffusion coefficient Ds are further derived. Based on the simplified form of the unsaturated convection diffusion equation, the physical expectation value of 24-hour salt flux of adjacent grids is calculated, and the MSE is obtained by comparing it with the model's predicted change as Lphys. The total loss L = L_MSE + 0.25·Lphys, and the Adam optimizer is used for training (lr=1e-3), with a batch size of 16 and 200 training rounds. The model automatically outputs three-layer water-salt distribution maps and a probability map of salt rise for the next 14 days daily.

[0026] Step S3: Variety-Adaptive Photosynthesis-Morphology Combined Stress Diagnosis like Figure 5 As shown, the target maize variety "Zhengdan 958" was selected. Based on a pre-trained segmented index model of a variety with similar salt tolerance in the basic model library, five micro-plots were set up in the field (with different salinity gradients) to obtain biomass-soil EC corresponding data. The MAML strategy was adopted, freezing the feature extraction layer of the basic model and updating only the segment thresholds (low / medium / high salinity boundary points) and the attenuation coefficient layer. After 50 inner loop iterations, variety adaptation was completed, and the three segment thresholds for this variety were determined to be EC≤2.5dS / m (safe), 2.5-4.5dS / m (moderate), and >4.5dS / m (severe).

[0027] The dual stress index was calculated daily: apparent fluorescence yield was obtained from SIF760 / PAR, and ΦPSII was inverted using a local calibration model, with PSI = (ΦPSII_ref - ΦPSII) / ΦPSII_ref; MSI was calculated from multispectral NDVI and LAI. The average EC of the root layer of each grid after S1 assimilation was taken and input into the variety adaptive model along with PSI and MSI. Mild stress was triggered when PSI > 0.15 or MSI > 0.2 for two consecutive days; moderate stress was triggered when PSI > 0.25 and MSI > 0.3; and severe stress was triggered when PSI > 0.4 or MSI > 0.5. Stress type labels were also output: PSI-dominated stress was "photosynthetically limited," and foliar fertilizer application was recommended; MSI-dominated stress was "morphologically stunted," and pulsed irrigation with salt leaching was recommended.

[0028] Step S4: Fully Coupled Dynamic Deduction Driven by Digital Twin Digital twin instances were generated for each 30m grid. The water-salt sub-model received the S2 prediction field as the background field and updated the root zone water and salt levels at 1-hour intervals. The crop growth sub-model used a modified APSIM-DualCrop to directly read the real-time stress coefficients output by S3 (mapped to stress factors between 0 and 1 based on stress levels), dynamically regulating the maximum photosynthetic rate, leaf expansion rate, and root water uptake. The salt memory factor was calculated using a formula, with NaEC converted in real-time from soil EC and SAR, and λ1 set to 0.05. The crop model automatically determined the growth stage based on accumulated temperature ≥10℃ and output the sensitivity coefficient γstage (0.7 for seedling stage, 0.5 for jointing stage, 0.4 for heading stage, and 0.25 for maturity stage).

[0029] The digital twin runs in 1-hour increments, updating water and salt levels, stress levels, photosynthetic growth, and root salt uptake feedback sequentially at each step. Daily output at UTC 00:00 includes: a heatmap showing the difference between grid-based growth rates and ideal rates without stress; a 14-day biomass accumulation prediction curve; and the mature-stage yield predicted by the integrated model (TCN+GBDT) with a 90% confidence interval. In this example, the heading stage prediction shows that approximately 15% of the area on the eastern side of the plot is expected to experience a yield reduction of 25%-35%, reaching the probability standard for an orange alert.

[0030] Step S5: Closed-loop precise control driven by model predictive control like Figure 6 As shown, an MPC optimizer was constructed using S4 digital twin as the forward simulator. The control variables were irrigation amount (0-40mm), salt leaching time (0 / 1), and foliar fertilizer application rate (0-5L / acre) for each grid. The optimization window Tp = 7 days, and the control step size was 1 day. The objective function weights were set to... =1.0 (output) =0.3 (water usage) =0.1 (energy consumption) =0.1 (salt leaching). Constraints: Soil moisture content after irrigation ≤ field capacity, no runoff; root zone EC ≤ crop salt tolerance threshold 4.5 dS / m; total daily irrigation ≤ 300 m³. 3 / hm 2 The particle swarm optimization algorithm (population 50, iterations 100) is used to solve the problem, generating a grid-based irrigation and spraying prescription map for the next 7 days every morning at dawn.

[0031] The prescription map is sent to the field via API to a pulse solenoid valve variable irrigation system, implementing precise zonal irrigation in the stress area (estimated irrigation volume 28mm). Simultaneously, a drone variable spraying platform is instructed to spray 0.3% potassium dihydrogen phosphate foliar fertilizer on photosynthetically limited stress areas. After execution, the S1 sensor network transmits real-time data on soil moisture, EC changes, and SIF and NDVI changes acquired by the drone the following day. The digital twin status is updated synchronously, and MPC is continuously corrected the following day, forming a complete closed loop. For drainage ditch repair work that cannot be automated, work orders are automatically generated and pushed to the administrator's app.

[0032] Step S6: Post-Season Adaptive Evolution After maize harvest, an experience pool was constructed using 450GB of accumulated observational data from the entire growth period, assimilation status, twin projection trajectories, stress diagnosis records, and MPC execution results (measured yield distribution map). An offline CQL reinforcement learning algorithm was employed, using the weighted sum of measured yield and irrigation energy consumption for each grid cell as the reward. This reward was applied to the hyperparameters of PI-STGCN (learning rate, loss weights), the segmentation thresholds of the stress diagnosis model, the crop model's variety parameters (specific leaf area, maximum root depth, radiation use efficiency), and the weights of MPC. , , , Joint optimization was performed. After optimization, the predicted RMSE decreased by 8% compared to the previous quarter, and the accuracy of the salt tolerance threshold for varieties improved by 12%. The optimized parameters were directly deployed the following year, resulting in a year-on-year increase in system performance.

[0033] The results of this embodiment show that, compared with traditional fixed-point monitoring, the RMSE of water and salt prediction decreased from 0.98 dS / m to 0.41 dS / m; stress warnings were issued an average of 4.2 days earlier; water was saved by 22%; and the final maize yield increased by 18% compared with the control area under conventional management, while the area of ​​secondary salinization was reduced by 30%. When the system was extended to adjacent soybean plots, variety migration and adaptation were completed using only 4 sets of data, and the model deployment time was shortened from the usual 1 month to 5 days.

[0034] A crop growth monitoring system for saline-alkali land, such as Figure 2 As shown, it includes: The perception layer is used to acquire multi-source perception data, and includes at least a satellite data receiving module, a UAV multi-payload data acquisition module, a field soil profile sensor network, a micro weather station and a groundwater monitoring module. An edge computing layer, deployed at a field base station, includes a data preprocessing unit and an edge assimilation unit, used to execute the ensemble Kalman filter assimilation algorithm described in step S1 of claim 1, and output assimilated data stream to the cloud; The cloud-based digital twin platform includes: a water-salt inversion engine, which deploys the physical information spatiotemporal graph convolutional network model described in step S2 of claim 1; a stress diagnosis engine, which deploys the variety-adaptive salt-stress diagnosis model described in step S3 of claim 1; a digital twin inference engine, which runs the grid-level digital twin instance described in step S4 of claim 1; and a model predictive control decision engine, which executes the model predictive control optimization described in step S5 of claim 1 to generate a daily zoned precise control prescription map. The execution layer, including the variable irrigation control system and the UAV variable spraying platform, receives the precise control prescription map and executes it automatically; It also includes a data storage and visualization platform for archiving data throughout the reproductive period, 3D visualization, and early warning information dissemination.

[0035] Therefore, the present invention adopts the above-mentioned method and system for monitoring crop growth in saline-alkali land, which realizes early warning of stress, rapid variety adaptation, and optimal synergy between water conservation and stable yield.

[0036] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for monitoring crop growth in saline-alkali land, characterized in that, Includes the following steps: Step S1: Acquire multi-source sensing data, construct a graph-grid hybrid topology, and use ensemble Kalman filtering to assimilate heterogeneous multi-scale observations into a unified grid, outputting a spatiotemporally continuous assimilation dataset. Step S2: Using the spatiotemporal continuous assimilation dataset of S1 as input, and following the graph-lattice hybrid topology of S1, construct a physical information spatiotemporal graph convolutional network model. By introducing a physical consistency loss term, extract water and salt transport features and predict the spatiotemporal dynamics of water and salt. This prediction field serves as the driving data for the water and salt sub-model of S4. Step S3: Using the spatiotemporal continuous assimilation dataset from S1, construct an adaptive salt stress diagnosis model for varieties through transfer learning, and output the real-time stress level and photosynthetic limitation / morphological atrophy label. Step S4: Construct a digital twin instance for the grid in S1. The water-salt sub-model updates the root zone state based on the prediction field in S2. The crop growth sub-model dynamically adjusts physiological parameters according to the stress level and type label in S3, and introduces the salt stress memory factor to deduce and output yield and biomass predictions. The prediction results provide optimization objectives and state feedback for S5. Step S5: Using the digital twin of S4 as a forward simulator, with the output prediction of S4 as the optimization objective and the salt tolerance threshold of S3 as the constraint, construct a model predictive control framework for rolling solution, generate a control prescription map and issue it for execution; after execution, it is assimilated again through multi-source perception of S1 to form a monitoring-decision-execution closed loop.

2. The method for monitoring crop growth in saline-alkali land according to claim 1, characterized in that, The multi-source sensing data mentioned in step S1 includes at least: satellite synthetic aperture radar data, satellite multispectral remote sensing data, UAV multispectral data, UAV thermal infrared data, UAV solar-induced chlorophyll fluorescence hyperspectral data, time-series data of stratified soil parameters collected by a field soil profile sensor network, field micro-weather station data, and groundwater monitoring data. Each sensor node of the field soil profile sensor network is equipped with a multi-parameter sensing unit including at least a conductivity probe, a moisture probe, a temperature probe, a pH probe, and a sodium ion selective electrode. The graph-grid hybrid topology uses each field sensor location and the center of each grid with a preset spatial resolution as graph nodes. A weighted adjacency matrix is ​​constructed based on the spatial distance between nodes and the similarity of soil texture. When a sensor node goes offline, the corresponding node is removed from the graph topology and the adjacency matrix is ​​dynamically reconstructed. The model still maintains its predictive ability through neighborhood node propagation.

3. The method for monitoring crop growth in saline-alkali land according to claim 1, characterized in that, The physical information spatiotemporal graph convolutional network model in step S2 adopts an encoder-decoder sequence-to-sequence architecture. Both the encoder and decoder contain stacked Chebyshev graph convolutional layers and gated recurrent unit layers. The physical consistency loss term calculates the residual between the physical expectation value and the model prediction value of the salt flux between adjacent grids based on the spatial variation water diffusion coefficient and salt diffusion coefficient dynamically estimated based on soil texture attributes for each grid. The spatial variation water diffusion coefficient and salt diffusion coefficient are obtained by dynamically estimating the sand content, silt content, and clay content of each grid unit by inputting them into the soil transfer function model.

4. The method for monitoring crop growth in saline-alkali land according to claim 1, characterized in that, The construction process of the adaptive salt stress diagnosis model for crop varieties described in step S3 is as follows: First, a basic model of salt tolerance response curves for multiple crop varieties is established. A piecewise exponential nonlinear function is used to characterize the multi-threshold piecewise response law of biomass or yield of different crop varieties to root zone salinity. For the target crop variety, a small amount of measured biomass-salt correspondence data obtained from field labeling experiments is used to perform transfer learning on the basic model through a model-independent meta-learning strategy or parameter fine-tuning strategy. The shallow network parameters are frozen, and only the piecewise threshold layer and the piecewise decay coefficient layer are trained and updated. The system automatically determines the segmented salt stress threshold and segmented function parameters for the variety. The output process for the stress level and type label is as follows: Based on the canopy fluorescence yield and actual quantum efficiency of photosystem II obtained from the UAV's sunlight-induced chlorophyll fluorescence hyperspectral data, a photosynthetic stress index is constructed; Based on the normalized vegetation index and leaf area index obtained from the UAV's multispectral data, a morphological stress index is constructed; Using root soil salinity, the photosynthetic stress index, and the morphological stress index as inputs, the system outputs the real-time stress level and stress type label for each grid cell.

5. The method for monitoring crop growth in saline-alkali land according to claim 1, characterized in that, The joint determination rule for the stress level is as follows: when the photosynthetic stress index exceeds the first threshold for two consecutive days or the morphological stress index exceeds the second threshold, it is determined to be mild stress; when the photosynthetic stress index and the morphological stress index exceed the third threshold and the fourth threshold at the same time, it is determined to be moderate stress; when any index exceeds the corresponding severe threshold, it is determined to be severe stress.

6. The method for monitoring crop growth in saline-alkali land according to claim 1, characterized in that, The mathematical form of the improved salt stress memory factor described in step S4 is: ； in, Let be the cumulative sodium ion equivalent toxicity memory at time t. The attenuation coefficient is... For a historic moment The equivalent electrical conductivity of sodium ions in the root zone is obtained by converting the soil electrical conductivity to the sodium adsorption ratio. The threshold value for sodium ion toxicity in crops is defined as follows: the sensitivity coefficient for the growth stage is dynamically assigned by the crop growth sub-model based on the accumulated temperature to determine the current phenological stage. The sensitivity coefficient for the seedling stage ranges from 0.6 to 0.8, for the heading stage from 0.4 to 0.6, and for the maturity stage from 0.2 to 0.

3.

7. The method for monitoring crop growth in saline-alkali land according to claim 1, characterized in that, The objective function of the model predictive control optimization framework described in step S5 is: ； in, To predict the control window length, Let be the predicted yield influencing factor at time t. Let t be the total irrigation water used at time t. Let t be the energy consumption for irrigation and spraying operations. The total amount of deep seepage salt leaching predicted at time t. , , , These are configurable weighting coefficients.

8. The method for monitoring crop growth in saline-alkali land according to claim 1, characterized in that, The yield prediction in step S4 is achieved using an ensemble learning model that combines temporal convolutional networks and gradient boosting trees. The model uses daily biomass throughout the entire growth period, cumulative stress at each stage, final leaf area index, and root salinity statistics as input features to output a grid-level predicted yield and its confidence interval. When the probability that the predicted yield is lower than the preset minimum yield line exceeds the preset probability threshold, a corresponding level of probability warning is triggered.

9. The method for monitoring crop growth in saline-alkali land according to claim 1, characterized in that, It also includes step S6, post-season adaptive evolution: After crop harvest, an experience pool is constructed by using all accumulated observation data, assimilation status data, digital twin simulation data, stress diagnosis records, and closed-loop management execution effect data from the entire growth period; an offline reinforcement learning algorithm is used, with the final measured yield and resource consumption as reward signals, to jointly optimize the hyperparameters of the physical information spatiotemporal graph convolutional network model, the segmented thresholds of the variety adaptive salt stress diagnosis model, the variety parameters of the crop growth sub-model in the digital twin, and the weight coefficients of the model prediction controller; the optimized parameters and model are deployed to the next growing season to achieve year-by-year adaptive increase in monitoring accuracy.

10. A crop growth monitoring system for saline-alkali land, characterized in that, include: The perception layer is used to acquire multi-source perception data, and includes at least a satellite data receiving module, a UAV multi-payload data acquisition module, a field soil profile sensor network, a micro weather station and a groundwater monitoring module. An edge computing layer, deployed at a field base station, includes a data preprocessing unit and an edge assimilation unit, used to execute the ensemble Kalman filter assimilation algorithm described in step S1 of claim 1, and output assimilated data stream to the cloud; The cloud-based digital twin platform includes: a water-salt inversion engine, which deploys the physical information spatiotemporal graph convolutional network model described in step S2 of claim 1; a stress diagnosis engine, which deploys the variety-adaptive salt-stress diagnosis model described in step S3 of claim 1; a digital twin inference engine, which runs the grid-level digital twin instance described in step S4 of claim 1; and a model predictive control decision engine, which executes the model predictive control optimization described in step S5 of claim 1 to generate a daily zoned precise control prescription map. The execution layer, including the variable irrigation control system and the UAV variable spraying platform, receives the precise control prescription map and executes it automatically; It also includes a data storage and visualization platform for archiving data throughout the reproductive period, 3D visualization, and early warning information dissemination.

Images