Adaptive learning system and method for arid zone crop growth prediction

Abstract

This invention relates to an adaptive learning-based crop growth prediction system and method for arid regions, encompassing the field of crop growth prediction. It generates environmental baseline features and physiological representation vectors by simultaneously acquiring ground environmental data, remote sensing physiological images, and historical yields, thereby constructing a multi-dimensional state-space model. The model is then input into a dynamic model containing physiological response sensitivity parameters, and a set of expected growth trajectories under various water supply conditions is generated through difference equations. Based on this, a marginal yield loss matrix is ​​constructed, and differential resource allocation vectors for each sub-region are output under total water resource constraints. Furthermore, the crop growth model is dynamically corrected and the marginal yield loss matrix adjusted using the residual discrepancy between real-time physiological feature values ​​and predicted trajectories. This invention achieves accurate simulation of crop growth status and optimized water resource allocation, possessing adaptive learning and disaster mitigation capabilities.

Description

Technical Field

[0001] This invention relates to the field of growth prediction, and more specifically to an adaptive learning-based crop growth prediction system and method for arid regions. Background Technology

[0002] Against the backdrop of climate change, precision management of agriculture in arid regions has become a key link in ensuring global food security. Modern agricultural production relies heavily on integrated sensor networks and multi-scale remote sensing technologies to achieve digital characterization of the dynamics of large-scale farmland environments and crop life indicators.

[0003] However, a macro-irrigation district is a complex system composed of multiple plots with vastly different soil textures, phenological characteristics, and planting habits. As drought stress intensifies, the physiological responses of crops at different developmental stages to water deficit exhibit significant nonlinear characteristics. This dual spatial and temporal heterogeneity means that a single irrigation index or isolated monitoring data cannot reflect the marginal contribution of a unit of water input to the overall regional output. Because current technologies generally limit themselves to static descriptions of current crop growth, it is difficult to quantify and assess the impact of different irrigation ratios on the expected returns of each sub-region when faced with a limited total water resource. Summary of the Invention

[0004] The purpose of this application is to provide an adaptive learning method and system for predicting crop growth in arid areas. This method effectively reduces the interference of extreme environmental fluctuations and multi-source sensing data drift on the stability of crop growth status representation, improves the accuracy of crop growth trajectory prediction and marginal yield loss measurement, and thus enhances the reliability of automated allocation decision-making results for agricultural resources in arid areas.

[0005] The objective of this application can be achieved through the following technical solution: Firstly, an adaptive learning method for predicting crop growth in arid regions, comprising the following steps: Simultaneously acquire ground point source environmental data, regional-scale remote sensing physiological image data, and historical yield data for each sub-region within the target area; The ground point source environmental data is calibrated and fault detected to generate environmental baseline features, and the regional scale remote sensing physiological image data is simultaneously subjected to multispectral feature extraction and pattern recognition to generate physiological representation vectors. The environmental baseline features and the physiological representation vectors are fused in a spatiotemporal scale to construct a multidimensional state space model. The multidimensional state space model is input into a preset crop growth dynamic model, which includes physiological response sensitivity parameters for characterizing crop drought resistance; the crop growth dynamic model is simulated by difference equations, and a set of expected growth trajectories of the sub-region under various preset water supply conditions is generated according to a preset execution cycle. Based on the set of expected growth trajectories, the expected output loss of the sub-region under the corresponding water supply conditions is determined; a marginal output loss matrix is ​​constructed based on the expected output loss to quantify the nonlinear mapping relationship between unit water variables and the expected output loss. Under the preset total water resource constraint, the marginal product loss matrix of each sub-region is processed to output the differential resource allocation vector for each sub-region; The crop physiological characteristic values ​​are acquired in real time, and the feedback deviation residual between the crop physiological characteristic values ​​and the matching trajectory in the expected growth trajectory set is calculated. The crop growth dynamic model is synchronously corrected based on the feedback deviation residual, and the marginal yield loss matrix is ​​dynamically adjusted.

[0006] Secondly, the adaptive learning-based crop growth prediction system for arid regions includes the following modules: The data acquisition module is used to simultaneously acquire ground point source environmental data, regional-scale remote sensing physiological image data, and historical yield data for each sub-region within the target area; The data processing module is used to calibrate and detect faults in the ground point source environmental data to generate environmental baseline features, and simultaneously perform multispectral feature extraction and pattern recognition on the regional scale remote sensing physiological image data to generate physiological representation vectors. The data analysis module is used to fuse the environmental baseline features with the physiological representation vectors at spatiotemporal scales to construct a multidimensional state space model. The crop growth simulation module is used to input the multidimensional state space model into a preset crop growth dynamic model, which includes physiological response sensitivity parameters for characterizing crop drought resistance; and to generate a set of expected growth trajectories of the sub-region under various preset water supply conditions by performing difference equation simulation on the crop growth dynamic model. The marginal loss assessment module is used to determine the expected output loss of the sub-region under the corresponding water supply conditions based on the expected growth trajectory set; and to construct a marginal output loss matrix based on the expected output loss to quantify the nonlinear mapping relationship between unit water variable and the expected output loss. The resource optimization and scheduling module is used to process the marginal output loss matrix of each sub-region under the preset total water resource constraint and output the differential resource allocation vector for each sub-region. The feedback deviation module is used to acquire crop physiological characteristic values ​​in real time and calculate the feedback deviation residual between the crop physiological characteristic values ​​and the matching trajectory in the expected growth trajectory set. An adaptive self-learning module is used to synchronously correct the crop growth dynamic model based on the feedback deviation residuals and dynamically adjust the marginal yield loss matrix.

[0007] Compared with the prior art, the beneficial effects of this application are: 1. In this invention, a multi-dimensional state space model is constructed by spatiotemporally fusing ground point source environmental data and regional-scale remote sensing physiological image data. A difference equation simulation is then performed using a crop growth dynamic model that includes physiological response sensitivity parameters. This solves the problem that existing technologies cannot digitally characterize the spatial heterogeneity of arid plots and the nonlinear physiological response of crop development stages, and achieves accurate characterization of multi-dimensional environmental indicators and dynamic growth in complex irrigation areas.

[0008] 2. In this invention, a marginal output loss matrix is ​​constructed based on the expected growth trajectory set under various preset water supply conditions, and a differential resource allocation vector for each sub-region is output under the constraint of total water resources. This overcomes the defect of existing decision-making schemes that cannot quantify the marginal contribution of unit water input to output value, and realizes a leap from isolated static monitoring to quantitative assessment and resource optimization allocation with the goal of maximizing the overall output value of the region.

[0009] 3. In this invention, the feedback deviation residual between the measured physiological data and the expected trajectory is calculated in real time, and the crop growth dynamic model and the marginal yield loss matrix are corrected synchronously using an adaptive learning algorithm. At the same time, the non-water stress risk area is identified by combining the spatial discrete residual value. This solves the problem that the model is prone to prediction drift and misjudgment of abnormal growth causes in complex dynamic environments, and improves the decision robustness of agricultural management in arid areas in response to climate change and the pertinence of plant protection inspection. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the steps of the adaptive learning-based crop growth prediction method for arid regions in this application. Figure 2 This is a schematic diagram of the modules of the adaptive learning-based crop growth prediction system for arid regions proposed in this application. Detailed Implementation

[0011] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0012] Example 1, like Figure 1 As shown, the adaptive learning-based crop growth prediction method for arid regions includes the following steps: Simultaneously acquire ground point source environmental data, regional-scale remote sensing physiological image data, and historical yield data for each sub-region within the target area; ground point source environmental data refers to point-level environmental parameters collected in real time by sensor nodes distributed in each sub-region of farmland, which can be achieved by combining multi-parameter soil sensors and weather stations. Its role is to provide a high temporal resolution local environmental benchmark, providing accurate ground truth values ​​for subsequent model input.

[0013] The ground point source environmental data is calibrated and fault detected to generate environmental baseline features. Simultaneously, multispectral feature extraction and pattern recognition are performed on the regional-scale remote sensing physiological image data to generate physiological representation vectors. Environmental baseline features refer to a standardized set of environmental parameters after outlier removal and sensor drift correction. Specifically, this can be achieved through adaptive Kalman filtering based on historical data and a neighboring node cross-validation algorithm. Its function is to ensure the reliability of the input data and eliminate the interference of sensor faults on model accuracy. Physiological representation vectors refer to a set of multidimensional features extracted from multispectral remote sensing images that characterize the physiological state of crops. Specifically, this can be achieved by extracting band combination features such as the normalized difference vegetation index and chlorophyll fluorescence ratio index, combined with a convolutional feature extraction network. Its function is to capture the spatial distribution pattern of crop growth from a surface-domain perspective.

[0014] The environmental baseline features and the physiological representation vectors are fused at spatiotemporal scales to construct a multidimensional state space model. The multidimensional state space model refers to a high-dimensional feature matrix that integrates point-level environmental information and area-domain physiological information and has clear spatial coordinates and timestamp indexes. Its role is to provide a complete state description for crop growth dynamic simulation that simultaneously includes micro-environmental driving forces and macro-physiological responses.

[0015] The multidimensional state-space model is input into a preset crop growth dynamic model. The crop growth dynamic model includes physiological response sensitivity parameters used to characterize crop drought resistance. The crop growth dynamic model is simulated using difference equations, and a set of expected growth trajectories for the sub-region under various preset water supply conditions is generated according to a preset execution cycle. The crop growth dynamic model is a numerical model that describes the dynamic growth process of crops driven by water, light, and temperature in the form of a set of difference equations. The physiological response sensitivity parameters are adjustable coefficients in the model that characterize the stress response intensity of a specific crop variety to water deficit. Specifically, they can be initially calibrated using nonlinear least squares combined with historical yield data. Their function is to make the model simulation trajectory fit the real biological characteristics of local crops.

[0016] Based on the set of expected growth trajectories, the expected output loss of the sub-region under the corresponding water supply conditions is determined; a marginal output loss matrix is ​​constructed based on the expected output loss to quantify the nonlinear mapping relationship between unit water variables and the expected output loss; the marginal output loss matrix refers to the matrix structure formed by aggregating the partial derivatives of the expected output loss of each sub-region with respect to the water supply according to the spatial dimension, and its function is to quantify the impact of marginal water resource allocation on output as a mathematical object that can participate in global optimization.

[0017] Under the preset total water resource constraint, the marginal output loss matrix of each sub-region is processed to output the differential resource allocation vector for each sub-region. The differential resource allocation vector refers to the optimal water allocation sequence determined by competitive ranking of the marginal loss coefficients of each sub-region under the premise of satisfying the total water constraint. Its function is to achieve differentiated and precise scheduling of water resources with the goal of minimizing the overall output loss.

[0018] The crop physiological characteristic values ​​are acquired in real time, and the feedback deviation residual between the crop physiological characteristic values ​​and the matching trajectory in the expected growth trajectory set is calculated. The feedback deviation residual refers to the difference between the crop physiological characteristic values ​​extracted from the real-time remote sensing image and the predicted values ​​of the model for the corresponding time period. Its function is to serve as an error signal for model self-calibration and drive the adaptive learning process.

[0019] The crop growth dynamic model is synchronously corrected based on the feedback deviation residual, and the marginal yield loss matrix is ​​dynamically adjusted.

[0020] The core innovation of this application lies in constructing a dynamic water resource optimization scheduling framework with the marginal product loss matrix as its core. This framework transforms the nonlinear relationship between water and grain from a static description into a quantitative structure that can be updated in real time. Furthermore, it achieves closed-loop calibration of model parameters through an adaptive learning mechanism, thereby reducing the lag and low accuracy issues in water allocation decisions in heterogeneous irrigation areas using traditional methods.

[0021] This application further proposes specific steps for simultaneously acquiring ground point source environmental data, regional-scale remote sensing physiological image data, and historical yield data for each sub-region within the target area, including: According to the preset periodic scheduling strategy, synchronous acquisition commands are sent to the ground sensor network deployed in the target area to collect environmental parameters such as soil moisture content, air temperature, relative humidity, wind speed and evaporation from meteorological station nodes in each sub-area, generating ground point source environmental data frames with node numbers and acquisition timestamps. The synchronous scheduling of the ground sensor network can be achieved by sending unified trigger signals to each node through a distributed scheduling controller calibrated with Network Time Protocol (NTP). Its function is to ensure that the acquisition times of each dispersed ground node are aligned, eliminate data asynchronous errors caused by node clock drift, and ensure the effectiveness of timestamp matching between ground point source data frames and subsequent remote sensing images.

[0022] According to the preset remote sensing data acquisition plan, the latest multispectral image covering the target area is retrieved from the satellite remote sensing platform interface, the image imaging time is recorded, and a regional-scale remote sensing physiological image data package with geographic coordinate information is generated. The acquisition of remote sensing images can be achieved by combining a pre-configured satellite transit schedule with an image cloud quality screening mechanism. Its function is to ensure that the acquired remote sensing images are as close as possible to the ground acquisition cycle in time and that the image quality meets the requirements for subsequent feature extraction. At the same time, the cloud threshold is used to automatically skip unqualified images and trigger replacement image requests.

[0023] From the historical yield database of the agricultural management information system, the quarterly yield records of the corresponding plots for the past five years are retrieved according to the spatial code of the sub-region. These records are then linked with the spatial codes of the current ground point source data frames and remote sensing image data packets to generate a spatial alignment index table for the three sources of data. The retrieval and association of historical yield data can be achieved by performing a spatial join query with the plot centroid coordinates as the primary key. Its purpose is to establish the spatial relationship between the three types of data, providing a unified data view that can be directly indexed for subsequent data processing modules.

[0024] Consistency verification is performed on the timestamps and spatial codes of the three-source data: using the imaging time of the remote sensing image as the time reference, it is determined whether the deviation between the acquisition timestamp of each node in the ground point source data frame and the reference time is within the preset synchronization tolerance window (e.g., ±3 hours); node data exceeding the tolerance window are marked as timestamp anomalies, and interpolation is performed using the valid sampled values ​​of the adjacent timestamps to complete the data; the three-source data that pass the verification are packaged into a synchronization data snapshot for subsequent processing. The specific implementation of consistency verification can be achieved by combining a lightweight data quality control script with an anomaly flag mechanism. Its role is to intercept low-quality data combinations with inconsistent timestamps before the data enters the processing flow, ensuring the spatiotemporal consistency of the data required for the subsequent construction of the multidimensional state space model from the source.

[0025] This application's solution integrates three heterogeneous data streams—ground point source acquisition, remote sensing image acquisition, and historical yield retrieval—into a unified timestamp calibration and spatial coding association mechanism. This achieves synchronous locking and spatial alignment of multi-source agricultural sensing data, resolving the spatiotemporal misalignment problem between data sources caused by discrete acquisition times and inconsistent coordinate systems in traditional data stitching methods. By introducing consistency verification and anomaly interpolation completion mechanisms at the data entry point, it ensures that the data snapshots entering subsequent processing have a clear and unified spatiotemporal reference, improving the accuracy of multidimensional state-space model construction and the reliability of downstream predictive analysis.

[0026] This application further proposes specific steps for calibrating and fault detection processing of the ground point source environmental data to generate environmental baseline features, and simultaneously performing multispectral feature extraction and pattern recognition on the regional-scale remote sensing physiological image data to generate physiological representation vectors, including: Fault detection is performed on the sampled values ​​of each sensor node in the ground point source environmental data: the spatial difference of the same parameter between adjacent nodes at the same time is calculated. If the difference between the sampled value of a parameter of a node and the mean of its surrounding neighboring nodes exceeds a preset spatial consistency threshold, the parameter of that node is marked as a suspected fault value. At the same time, the jump amplitude of each parameter of the same node on the time axis is detected. If the difference between adjacent time points exceeds a preset temporal abrupt change threshold, it is additionally marked as a temporal anomaly value. Fault detection can be achieved by combining spatial interpolation residual analysis with temporal moving standard deviation monitoring. Its function is to distinguish between the real microclimate gradient and the reading anomaly caused by sensor hardware failure, and to avoid introducing equipment failure signals into subsequent feature calculations.

[0027] For sampling points marked as suspected fault values ​​or time-series anomalies, missing value repair is performed: spatial interpolation is preferentially used to replace the synchronous sampling mean of neighboring nodes in the same sub-region; if all neighboring nodes in the same sub-region are unavailable, the time-series backfill is performed using the historical average value of that node (same reproductive stage, similar date); after repair, zero-point drift deviation compensation is performed on each parameter of each node based on historical reference station data, the repaired environmental sampling values ​​are corrected to the physical reference range, and the environmental reference feature vector of each sub-region is output; this process can be implemented by a hierarchical missing value repair strategy combined with a reference station-based linear deviation compensation algorithm, which aims to eliminate the interference of sensor system errors and hardware drift on the accuracy of environmental characteristics while preserving the true spatial heterogeneity of microclimate.

[0028] Atmospheric correction is performed on remote sensing physiological image data to convert the radiance values ​​received by the satellite sensor into surface reflectance. Specifically, atmospheric correction can be achieved by using the 6S algorithm based on the radiative transfer model, with the imaging angle and atmospheric visibility parameters in the image header file as input to perform pixel-by-pixel correction. Its function is to eliminate the interference of atmospheric scattering and absorption on multispectral reflectance, so that the reflectance of images from different scenes and time phases is physically comparable.

[0029] Based on atmospherically corrected surface reflectance images, 12 vegetation physiological indices, including Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Normalized Difference Moisture Index (NDWI), and Red-edged Chlorophyll Index (CIre), were calculated. Subsequently, the images of each vegetation index were zoned and statistically analyzed according to the spatial boundaries of sub-regions. The mean and standard deviation of each vegetation index within each sub-region were calculated to generate a 24-dimensional (12 indices × mean + standard deviation) zoning statistical feature matrix. The calculation and zoning statistics of vegetation indices can be achieved by combining raster mask extraction and statistical operations with vector sub-region boundaries using a remote sensing image processing library (such as GDAL). Its role is to aggregate pixel-level remote sensing information to the spatial scale of sub-regions, aligning it with the spatial granularity of ground point source data, and providing a scale-consistent physiological feature expression for subsequent spatiotemporal fusion.

[0030] The zonal statistical feature matrix is ​​input into a pre-trained multispectral pattern recognition network to classify the current crop growth status of each sub-region into physiological patterns (such as normal growth, mild water deficit, severe water deficit, growth stagnation, etc.), and outputs the category label and confidence score of the corresponding physiological state for each sub-region. The zonal statistical feature matrix is ​​then concatenated with the physiological state category label and confidence score to generate the physiological representation vector of each sub-region. Specifically, the multispectral pattern recognition network can be implemented using a lightweight convolutional classifier based on transfer learning, which is trained in a supervised manner with a large number of historical image samples and ground-based verification data. Its role is to map continuous vegetation index features into discrete physiological state labels with biological interpretability, thereby enhancing the semantic expressive power of the physiological representation vector in subsequent state space modeling.

[0031] This application's solution achieves the synchronous, high-quality generation of ground physical environment features and remotely sensed crop physiological features by parallelizing and pipelined five processing steps: ground sensor fault detection and hierarchical missing data repair, baseline drift bias compensation, remote sensing atmospheric correction, vegetation physiological index zoning statistics, and multispectral pattern recognition. This overcomes the technical shortcomings of traditional processing workflows, such as the direct transmission of sensor fault values ​​to subsequent pollution modeling and the lack of scale alignment capabilities due to remote sensing features remaining at the pixel level. By introducing a spatial-temporal dual-dimensional fault detection and hierarchical repair strategy on the environmental feature side, it distinguishes between real microclimate gradients and equipment anomaly signals. Furthermore, by introducing pattern recognition semantic label expansion on the physiological feature side, it enables physiological representation vectors to possess both numerical continuity and state semantic interpretability, providing higher-quality and more expressive dual-path feature inputs for the spatiotemporal fusion of multidimensional state-space models.

[0032] This application further proposes that the specific steps for constructing a multidimensional state-space model include: Obtain the first spatial coordinates and first timestamp corresponding to the environmental baseline feature, and obtain the second spatial coordinates and second timestamp corresponding to the physiological representation vector; under the condition that the first timestamp and the second timestamp match, fuse and splice the environmental baseline feature and the physiological representation vector according to the first spatial coordinates to construct a multidimensional state space model.

[0033] The timestamp matching method is as follows: using the imaging time of the remote sensing image as a reference, the nearest whole-point sample value from the ground sensor data is used as the environmental reference feature for the corresponding time, with a maximum allowed time deviation of ±3 hours. For spatial coordinate alignment, the geographic centroid coordinates of the sub-region are used as the index key. The environmental reference feature vector (dimension E) and the physiological representation vector (dimension P) under the same centroid coordinates are directly concatenated to form a multidimensional state space model node with dimension E+P. The combination of nodes from all sub-regions constitutes a complete multidimensional state space model.

[0034] Through the above technical solution, this application achieves high-precision spatiotemporal alignment and fusion of ground point source data and remote sensing image data. This reduces the emissivity of data inconsistencies caused by differences in acquisition time and spatial resolution between different sensing modalities, ensures strict correspondence of features in each dimension of the multidimensional state space model in physical spatiotemporal space, and provides a high-quality initial state foundation for difference equation simulation.

[0035] This application further proposes that the specific steps for generating the expected set of growth trajectories include: A preset crop physiological characteristic parameter library is invoked to obtain the photosynthetic rate coefficient and transpiration efficiency coefficient at different growth stages; the photosynthetic rate coefficient and transpiration efficiency coefficient are weighted and corrected using the physiological response sensitivity parameter to obtain the corrected growth control parameters; the corrected growth control parameters and the data in the multidimensional state space model are substituted into the crop growth dynamic model, and the difference equation simulation is executed to generate the expected growth trajectory set of each sub-region under an increasing water supply gradient.

[0036] The crop physiological characteristic parameter database stores photosynthetic rate coefficients and transpiration efficiency coefficients covering four main growth stages: seedling stage, jointing stage, flowering stage, and grain-filling stage. These parameter values ​​were obtained through statistical analysis of years of measured data from regional agricultural experimental stations. The physiological response sensitivity parameter (range 0–1) is used for weighted correction of the photosynthetic rate coefficient and transpiration efficiency coefficient using multiplicative scaling, i.e.: The corrected growth control parameter = original coefficient × (1 - β × (1 - sensitivity parameter)); Where β is the drought stress amplification factor (default value 0.8).

[0037] The difference equation simulation uses a daily time step and recursively extrapolates for 30 days under 11 different water supply gradients, outputting the predicted sequences of daily LAI, daily biomass increment, and cumulative yield for each sub-region.

[0038] Through the above technical solution, this application realizes the generation of multi-scenario crop growth trajectories based on adaptive parameters of the growth period. By configuring photosynthetic and transpiration parameters at different growth stages and introducing physiological response sensitivity parameters to dynamically adjust the drought stress effect, the simulated trajectory can accurately reflect the nonlinear growth response law of crops under different water supply conditions, laying a refined simulation foundation for multi-scenario yield loss analysis.

[0039] This application further proposes that the specific steps for determining the expected loss of output value include: Extract the predicted fruit enlargement rate and leaf area index from the expected growth trajectory set; combine the historical yield data with the preset market unit price coefficient to calculate the predicted output value reduction under specific moisture conditions, and record it as the expected output value loss.

[0040] Fruit enlargement rate was extracted from the daily biomass increment sequence in the expected growth trajectory set, and leaf area index (LAI) predictions were directly read from the daily LAI output values. Historical yield data was used to calibrate the conversion coefficient between LAI and fruit enlargement rate to final yield (kg / acre), which was obtained by fitting historical measured data using multiple linear regression.

[0041] The output value is calculated as follows: Forecasted output (kg / mu) × market unit price coefficient (yuan / kg) = expected output value (yuan / mu). The expected output value loss = expected output value under full irrigation scenario - expected output value under specific water supply scenario. The larger the value, the greater the inhibition of output value by the water supply conditions.

[0042] Through the above technical solution, this application achieves an effective transformation of crop physiological growth prediction results into economic output loss indicators. By introducing historical yield data calibration conversion coefficients and market unit price coefficients for output value conversion, the growth differences under different sub-regions and different water supply scenarios can be compared using a unified monetized indicator, providing an economically comparable decision-making basis for subsequent marginal loss matrix construction and water resource scheduling optimization.

[0043] This application further proposes that the specific steps for constructing the marginal product loss matrix include: Obtain the dynamic functional relationship between the expected output loss and the water supply variable; perform differentiation on the dynamic functional relationship to determine the marginal loss change rate corresponding to the current reproductive stage of each sub-region; and aggregate the marginal loss change rates of each sub-region according to the spatial coordinate dimension through preset mapping weights to construct a marginal output loss matrix.

[0044] The dynamic functional relationship between expected output loss and water supply was obtained as a continuous function by fitting the predicted loss values ​​under 11 discrete water supply gradients using cubic spline interpolation. The derivative of this continuous function at the current actual water supply is the marginal loss change rate (unit: yuan / mm) for the current growth stage of the sub-region. Its physical meaning is the reduction in output loss that can be achieved by increasing irrigation by 1 mm based on the current water supply. The mapping weights were determined by comprehensively scoring the soil water retention capacity (determined by soil texture), plot slope, and historical loss sensitivity of each sub-region. After weight normalization, the matrix was filled into an M×N marginal output loss matrix according to the spatial coordinate dimensions (M and N are the number of sub-regions in the region, respectively).

[0045] Through the above technical solution, this application realizes a spatially structured expression of the marginal benefits of output loss in multiple sub-regions. By introducing continuous function fitting and numerical differentiation, the loss prediction results under discrete water supply scenarios are transformed into continuously differentiable marginal benefit curves. Combined with the aggregation processing of spatial mapping weights, the marginal output loss matrix can accurately characterize the differentiated marginal responses of different plots in the region to irrigation water volume, providing a precise optimization reference for differentiated water resource allocation.

[0046] This application further proposes that the specific steps for outputting the differential resource allocation vector include: The marginal weight coefficients of each sub-region are extracted from the marginal product loss matrix; within a preset range that satisfies the total water resource constraint, the sub-regions are competitively ranked according to the marginal weight coefficients; based on the results of the competitive ranking calculation, the optimal water allocation for each plot is determined, and a differential resource allocation vector is generated.

[0047] The competitive ranking calculation employs a greedy iterative strategy: First, all sub-regions are arranged in descending order of their marginal weight coefficients, forming a priority queue. Under the constraint of total available water, the highest-priority sub-regions are allocated their maximum single effective irrigation volume (determined based on the sub-region's soil field capacity limit). After each allocation, the corresponding allocation volume is deducted from the total available water, and the marginal weight coefficient of the sub-region under the current water supply is recalculated (due to diminishing marginal returns, the weight of sub-regions that have already allocated more water will decrease). The allocation continues after re-ranking until the total available water is exhausted or all sub-regions reach the full irrigation threshold. Finally, the allocated water volumes of each sub-region are summarized into a differential resource allocation vector, formatted as a mapping table of spatial coordinates of each sub-region and the corresponding allocated water volume (mm / day).

[0048] Through the above technical solution, this application achieves dynamic optimal differentiated allocation of limited water resources in heterogeneous farmland spaces. By incorporating the law of diminishing marginal returns into the competitive ranking iterative process, water allocation decisions can adaptively balance the marginal output suppression benefits of different sub-regions under total quantity constraints, avoiding resource misallocation caused by simple equal distribution or allocation by area, and improving the overall output guarantee efficiency of limited irrigation water resources.

[0049] This application further proposes that the specific steps for calculating the feedback bias residual include: Crop physiological characteristic values ​​are acquired in real time using the regional-scale remote sensing physiological image data; the difference between the crop physiological characteristic values ​​and the predicted physiological parameters for the corresponding time period in the expected growth trajectory set is calculated; and the output of the difference calculation is determined as the feedback bias residual.

[0050] Real-time extraction of crop physiological characteristics employs the same multispectral processing workflow as the generation of physiological characterization vectors. Using the latest remote sensing image as input, it outputs the measured LAI value and canopy water content index for each sub-region on that day. Predicted physiological parameters for the corresponding time period in the expected growth trajectory set are directly retrieved from the trajectory set using the date index. The difference calculation method is: Feedback bias residual = Measured physiological characteristic value - Predicted physiological parameter value. A negative difference indicates that the actual growth state is better than the prediction, while a positive difference indicates growth inhibition. The weighted root mean square of the bias residuals across multiple physiological characteristic dimensions is taken to obtain a comprehensive feedback bias residual scalar, which serves as the trigger for model correction.

[0051] Through the above technical solution, this application achieves online quantification of the deviation between crop growth status and model prediction based on real-time remote sensing monitoring. This enables the system to automatically evaluate the model prediction accuracy within each remote sensing data acquisition cycle, providing reliable error signal input for the adaptive learning algorithm, thereby supporting continuous online correction of the crop growth dynamic model.

[0052] This application further proposes that the specific steps for modifying the crop growth dynamic model and dynamically adjusting the marginal yield loss matrix include: Determine whether the feedback deviation residual exceeds a preset deviation tolerance threshold; if the feedback deviation residual is greater than the preset deviation tolerance threshold, invoke an adaptive learning algorithm to reverse-correct the physiological response sensitivity parameter in the crop growth dynamic model based on the feedback deviation residual; based on the corrected physiological response sensitivity parameter, synchronously update the mapping weight in the marginal yield loss matrix; if the feedback deviation residual is less than or equal to the preset deviation tolerance threshold, keep the physiological response sensitivity parameter and the mapping weight unchanged.

[0053] The bias tolerance threshold is set based on 1.5 times the remote sensing inversion error σ obtained from historical data statistics. This means that biases within the normal remote sensing inversion accuracy range do not trigger corrections, thus avoiding misidentification of sensor noise as model error. The adaptive learning algorithm uses gradient descent with momentum term, employing the sum of squared feedback bias residuals as the loss function to iteratively update the physiological response sensitivity parameters. The initial learning rate is set to 0.01, and the momentum coefficient is set to 0.9. After each iteration, it is necessary to verify whether the updated parameters are within the physiologically reasonable range (0.1–1.0). If they exceed the range, they are truncated to the boundary value. The synchronous update method for the mapping weights is as follows: the marginal loss change rate of each sub-region is recalculated using the corrected physiological response sensitivity parameters, and spatial aggregation is re-executed to update the mapping weight values ​​of the corresponding sub-regions in the marginal product loss matrix.

[0054] Through the above technical solution, this application achieves closed-loop collaborative online adaptive updating of the crop growth dynamic model and the resource scheduling decision matrix. By introducing a two-branch judgment mechanism with a deviation tolerance threshold and a gradient correction strategy for the driving force, timely response to actual field growth deviations is achieved while ensuring model stability. This enables the entire prediction and allocation system to continuously approximate the actual crop and water response patterns, improving the long-term effectiveness and robustness of water resource scheduling decisions in arid areas within a dynamic agricultural environment.

[0055] This application further proposes that it also includes: Extract the predicted physiological state value corresponding to each sub-region in the expected growth trajectory set, and extract the corresponding measured physiological state value from the physiological representation vector; calculate the spatial discrete residual value between the predicted physiological state value and the measured physiological state value; if the spatial discrete residual value exceeds a preset abnormal fluctuation threshold, perform spatial clustering analysis on the spatial discrete residual value to identify contiguous abnormal growth clusters within the sub-region; mark the contiguous abnormal growth clusters as non-water stress risk areas, and generate corresponding plant protection inspection priorities based on the magnitude of the spatial discrete residual value.

[0056] The spatial discrete residual value is calculated as follows: the Euclidean distance between the predicted physiological state value (LAI, canopy water content, etc.) and the measured physiological state value is calculated for each sub-region, and this distance is used as the spatial discrete residual scalar for that sub-region. The abnormal fluctuation threshold is set as the mean of the historical spatial discrete residuals plus twice the standard deviation (μ+2σ). Sub-regions exceeding this threshold are marked as candidate anomalies. DBSCAN spatial clustering analysis is performed on the spatial coordinates of the candidate anomalies, with a cluster neighborhood radius of 500 meters and a minimum cluster size of 3. The identified dense clusters are considered as contiguous abnormal growth clusters. The criteria for determining non-water stress risk areas are: the marginal yield loss matrix of the cluster area predicts a loss within the normal range (i.e., not caused by insufficient water supply), while the measured physiological state value is significantly lower than the prediction, suggesting possible interference from pests and diseases, soil compaction, or other non-water factors. Plant protection inspection priorities are classified by the mean of the spatial discrete residuals within the clustered area: those with a mean greater than 3σ are classified as Level 1 (urgent inspection), those with a mean between 2σ and 3σ are classified as Level 2 (recent inspection), and those with a mean less than 2σ but exceeding the threshold are classified as Level 3 (routine inspection).

[0057] Through the above technical solution, this application realizes the spatial identification of abnormal growth areas of non-water-stressed crops and the automatic generation of plant protection early warning priorities. By combining the clustering analysis results of spatial discrete residuals with water supply prediction, it distinguishes between crop growth abnormalities caused by insufficient irrigation and those caused by non-water factors such as pests and diseases. This provides differentiated and precise plant protection inspection guidance for agricultural managers in arid areas, making up for the shortcomings of traditional irrigation scheduling systems in perceiving non-irrigated agricultural risks.

[0058] As a preferred embodiment, the solution of this application is specifically implemented as follows: A network of ground-based meteorological stations was deployed in the target arid agricultural area, with sensor nodes spaced 500 meters apart and sampling once per hour to collect environmental parameters such as soil moisture content, air temperature, wind speed, and evaporation. Simultaneously, multispectral satellite remote sensing imagery with a resolution of 10 meters was used, acquired every 5 days, including red-edge (705nm), near-infrared (842nm), and shortwave infrared (1610nm) bands. Historical yield data was derived from actual harvest records of individual plots over the past five years, statistically summarized by sub-region.

[0059] Three methods are used for ground sensor data The criteria include outlier detection, with sampling points exceeding the threshold replaced by interpolation from neighboring nodes; simultaneous sensor zero-point drift correction is performed, and bias compensation for each node is based on regional meteorological reference station data. After atmospheric correction of the remote sensing imagery, 12 vegetation physiological indices, such as NDVI, NDWI, and red-edge chlorophyll index, are extracted. These indices are then used to identify different crop type regions through a pre-trained multispectral semantic segmentation network, generating physiological representation vectors for molecular regions with a dimension of 12×N (N being the number of sub-regions).

[0060] Using the geographic centroid coordinates of the sub-region as the alignment reference, the environmental baseline features (8 dimensions such as temperature, humidity, and soil moisture content) and the physiological representation vector (12 dimensions) are spliced ​​and fused under the condition of timestamp matching to construct a 20-dimensional multidimensional state space model. The update frequency is consistent with the remote sensing image acquisition cycle, that is, it is updated once every 5 days.

[0061] A multidimensional state-space model is input into a pre-defined crop growth dynamic model (a difference equation system based on the AquaCrop framework). The water supply gradient is set from 0 mm / day to 10 mm / day, with a step size of 1 mm / day, encompassing 11 water supply scenarios. The difference equations are recursively applied on a daily basis to generate time series predictions of biomass accumulation, leaf area index, and yield for each sub-region under the 11 water supply conditions over the next 30 days—essentially, a set of expected growth trajectories.

[0062] The predicted fruit enlargement rate (g / day) and leaf area index (LAI) are extracted from the expected growth trajectory set of each sub-region. Combined with the local historical crop yield coefficient and market price (yuan / kg), the output value under each water supply gradient is calculated. Using a fully irrigated scenario as a benchmark, the output value reduction under each water supply scenario is calculated and denoted as the expected output value loss (yuan / mu). The functional relationship between the expected output value loss and water supply is numerically differentiated to obtain the marginal loss change rate (yuan / mm) of each sub-region at the current growth stage. The marginal loss change rates of each sub-region are then aggregated into a marginal yield loss matrix using preset spatial mapping weights (configured according to soil type and slope factor).

[0063] Under the constraint of total available water (e.g., the daily dispatchable irrigation water volume is 500 cubic meters), the marginal weight coefficients of each sub-region are extracted from the marginal product loss matrix. The sub-regions are competitively sorted from high to low according to their marginal weights. Water allocation is prioritized for the sub-region with the greatest marginal product loss suppression benefit, and the allocation is gradually reduced until the total amount is exhausted. Finally, the resource allocation vector of each sub-region is output (unit: mm / day).

[0064] The measured LAI (Laminated Area Identification) and canopy water content index of each sub-region are extracted from real-time remote sensing images. These values ​​are then compared with the predicted physiological parameters for the corresponding time period in the expected growth trajectory set for that day to obtain the feedback bias residual. If the LAI feedback bias residual for a sub-region exceeds a preset bias tolerance threshold (e.g., 0.3), an adaptive learning algorithm is triggered. Stochastic gradient descent with momentum is used to reverse-correct the physiological response sensitivity parameters of that sub-region. After correction, the mapping weights of the corresponding sub-region in the marginal yield loss matrix are updated synchronously. If the bias residual is below the threshold, the parameters remain unchanged to maintain model stability.

[0065] like Figure 2 As shown, in this embodiment, an adaptive learning-based crop growth prediction system for arid regions is provided, comprising the following modules: The data acquisition module is used to synchronously acquire ground point source environmental data, regional-scale remote sensing physiological image data, and historical yield data of each sub-region within the target area. Specifically, this module can be composed of a ground meteorological sensor network control unit, a remote sensing image download and parsing interface, and an agricultural yield database connection adapter. Its function is to serve as the data entry point for the entire system, ensuring the synchronous access and storage management of multi-source sensing data.

[0066] The data processing module is used to calibrate and detect faults in the ground point source environmental data to generate environmental baseline features, and simultaneously perform multispectral feature extraction and pattern recognition on the regional-scale remote sensing physiological image data to generate physiological representation vectors. Specifically, this module can be composed of a sensor data quality control subunit and a multispectral feature extraction subunit in parallel. Its function is to transform the original heterogeneous data into high-quality usable feature representations and eliminate the interference of noise and errors on subsequent analysis.

[0067] The data analysis module is used to fuse the environmental baseline features with the physiological representation vectors at spatiotemporal scales to construct a multidimensional state space model. Specifically, this module can be composed of a timestamp alignment processing unit and a spatial coordinate index fusion unit. Its function is to establish a unified state description system across modalities and scales, providing structured input for crop growth dynamic simulation.

[0068] The crop growth simulation module is used to input the multidimensional state-space model into a preset crop growth dynamic model, and generate a set of expected growth trajectories for the sub-region under various preset water supply conditions by performing difference equation simulation on the crop growth dynamic model. Specifically, this module can be composed of a crop physiological characteristic parameter library management unit, a growth control parameter correction unit, and a multi-scenario difference equation parallel recursion unit. Its function is to extrapolate the crop growth process in parallel under multiple water supply scenarios based on the mechanism model, and output a growth trajectory matrix that can be used for loss analysis.

[0069] The marginal loss assessment module is used to determine the expected output loss of the sub-region under the corresponding water supply conditions based on the expected growth trajectory set, and to construct a marginal output loss matrix based on the expected output loss. Specifically, this module can be composed of an output conversion calculation unit, a function fitting and numerical differentiation unit, and a spatial aggregation weight calculation unit. Its function is to transform the differences in growth trajectories into a marginal benefit space matrix that can be directly used by the optimization algorithm.

[0070] The resource optimization scheduling module is used to process the marginal output loss matrix of each sub-region under the preset total water resource constraint, and output the differentiated resource allocation vector for each sub-region. Specifically, this module can be composed of a marginal weight extraction unit, a competitive sorting iterative solution unit, and a water allocation vector generation unit. Its function is to achieve the global optimal differentiated spatial allocation of water resources under the total constraint.

[0071] The feedback deviation module is used to acquire crop physiological characteristic values ​​in real time and calculate the feedback deviation residual between the crop physiological characteristic values ​​and the matching trajectory in the expected growth trajectory set. Specifically, this module can be composed of a real-time remote sensing feature extraction unit, a trajectory time period index retrieval unit, and a multi-dimensional deviation residual calculation unit. Its function is to automatically complete the model accuracy evaluation in each remote sensing data acquisition cycle and generate an error signal for use by the adaptive learning module.

[0072] An adaptive self-learning module is used to synchronously correct the crop growth dynamic model based on the feedback deviation residuals and dynamically adjust the marginal yield loss matrix. Specifically, this module can be composed of a deviation tolerance threshold judgment unit, an adaptive gradient descent parameter correction unit, and a marginal loss matrix weight synchronous update unit, so as to achieve dynamic self-improvement of prediction accuracy and scheduling effectiveness.

[0073] Through the above technical solutions, this application constructs a closed-loop system encompassing multi-source, multi-scale sensing, parallel deduction of growth mechanisms, and marginal benefit optimization scheduling. This achieves deep spatiotemporal fusion of ground point source environment and regional scale remote sensing features. By utilizing differential equation simulation, the complex physiological growth process of crops is transformed into a quantifiable marginal yield loss matrix. Thus, under the constraint of total water resources, differentiated and precise resource allocation at the sub-regional level is achieved through competitive ranking. Simultaneously, in conjunction with real-time feedback deviation residual calculation and adaptive gradient correction mechanisms, the system can dynamically optimize model accuracy and scheduling strategies based on the actual crop growth status. This overcomes the spatiotemporal mismatch and mechanism disconnect problems existing in traditional agricultural resource allocation, improving the scientific nature, predictability, and risk resistance of agricultural production management.

[0074] The above embodiments are only used to illustrate the technical methods of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of this application without departing from the spirit and scope of the technical methods of this application.

Claims

1. An adaptive learning-based crop growth prediction method for arid regions, characterized in that, Includes the following steps: Simultaneously acquire ground point source environmental data, regional-scale remote sensing physiological image data, and historical yield data for each sub-region within the target area; The ground point source environmental data is calibrated and fault detected to generate environmental baseline features, and the regional scale remote sensing physiological image data is simultaneously subjected to multispectral feature extraction and pattern recognition to generate physiological representation vectors. The environmental baseline features and the physiological representation vectors are fused in a spatiotemporal scale to construct a multidimensional state space model. The multidimensional state space model is input into a preset crop growth dynamic model, which includes physiological response sensitivity parameters for characterizing crop drought resistance; the crop growth dynamic model is simulated by difference equations, and a set of expected growth trajectories of the sub-region under various preset water supply conditions is generated according to a preset execution cycle. Based on the set of expected growth trajectories, the expected output loss of the sub-region under the corresponding water supply conditions is determined; a marginal output loss matrix is ​​constructed based on the expected output loss. Under the preset total water resource constraint, the marginal product loss matrix of each sub-region is processed to output the differential resource allocation vector for each sub-region; The crop physiological characteristic values ​​are acquired in real time, and the feedback deviation residual between the crop physiological characteristic values ​​and the matching trajectory in the expected growth trajectory set is calculated. The crop growth dynamic model is synchronously corrected based on the feedback deviation residual, and the marginal yield loss matrix is ​​dynamically adjusted.

2. The adaptive learning method for predicting crop growth in arid regions according to claim 1, characterized in that, The process of constructing a multidimensional state-space model includes: Obtain the first spatial coordinates and first timestamp corresponding to the environmental baseline features, and obtain the second spatial coordinates and second timestamp corresponding to the physiological representation vector; Under the condition that the first timestamp and the second timestamp match, the environmental baseline features and the physiological representation vector are fused and spliced ​​according to the first spatial coordinates to construct a multi-dimensional state space model.

3. The adaptive learning method for predicting crop growth in arid regions according to claim 1, characterized in that, The process of generating the expected set of growth trajectories includes: The system calls a pre-defined database of crop physiological parameters to obtain the photosynthetic rate coefficient and transpiration efficiency coefficient at different growth stages. The photosynthetic rate coefficient and the transpiration efficiency coefficient are weighted and corrected using the physiological response sensitivity parameters to obtain the corrected growth control parameters. The modified growth control parameters and the data in the multidimensional state-space model are substituted into the crop growth dynamic model, and the difference equation simulation is performed to generate the expected growth trajectory set of each sub-region under an increasing water supply gradient.

4. The adaptive learning method for predicting crop growth in arid regions according to claim 1, characterized in that, The process of determining the expected output loss is as follows: Extract the predicted fruit enlargement rate and leaf area index from the expected growth trajectory set; By combining the historical production data with the preset market unit price coefficient, the predicted value of output reduction under specific moisture conditions is calculated and recorded as the expected output loss.

5. The adaptive learning method for predicting crop growth in arid regions according to claim 1, characterized in that, The process of constructing the marginal product loss matrix includes: Obtain the dynamic functional relationship between the expected output loss and the water supply variable; The dynamic function relationship is differentiated to determine the marginal loss rate of each sub-region at the current reproductive stage; By using preset mapping weights, the marginal loss change rates of each sub-region are aggregated according to spatial coordinate dimensions to construct a marginal output loss matrix.

6. The adaptive learning method for predicting crop growth in arid regions according to claim 1, characterized in that, The process of outputting the differential resource allocation vector includes: Extract the marginal weight coefficients of each sub-region from the marginal product loss matrix; Within a preset range that satisfies the total water resources constraint, each sub-region is competitively ranked according to the marginal weight coefficient. Based on the results of the competitive ranking calculation, the optimal water allocation for each plot is determined, and a differential resource allocation vector is generated.

7. The adaptive learning method for predicting crop growth in arid regions according to claim 1, characterized in that, The process of calculating the feedback bias residual includes: Crop physiological characteristic values ​​are obtained in real time using the regional-scale remote sensing physiological image data; The difference between the crop physiological characteristic values ​​and the predicted physiological parameters for the corresponding time period in the expected growth trajectory set is calculated. The output of the difference operation is determined as the feedback deviation residual.

8. The adaptive learning method for predicting crop growth in arid regions according to claim 5, characterized in that, The process of revising the crop growth dynamic model and dynamically adjusting the marginal yield loss matrix includes: Determine whether the feedback deviation residual exceeds a preset deviation tolerance threshold; If the feedback deviation residual is greater than the preset deviation tolerance threshold, then the adaptive learning algorithm is invoked to reverse the physiological response sensitivity parameter in the crop growth dynamic model based on the feedback deviation residual. Based on the corrected physiological response sensitivity parameters, the mapping weights in the marginal product loss matrix are updated synchronously. If the feedback deviation residual is less than or equal to the preset deviation tolerance threshold, then the physiological response sensitivity parameter and the mapping weight remain unchanged.

9. The adaptive learning method for predicting crop growth in arid regions according to claim 1, characterized in that, Also includes: Extract the predicted physiological state value corresponding to each sub-region in the expected growth trajectory set, and extract the corresponding measured physiological state value from the physiological representation vector; Calculate the spatial discrete residual between the predicted physiological state value and the measured physiological state value; If the spatial discrete residual value exceeds the preset abnormal fluctuation threshold, then spatial cluster analysis is performed on the spatial discrete residual value to identify the continuous abnormal growth clusters in the sub-region. The contiguous abnormal growth clusters are marked as non-water stress risk areas, and corresponding plant protection inspection priorities are generated based on the magnitude of the spatial discrete residual values.

10. An adaptive learning-based crop growth prediction system for arid regions, characterized by: The application includes an adaptive learning method for predicting crop growth in arid regions as described in any one of claims 1 to 9, comprising: The data acquisition module is used to simultaneously acquire ground point source environmental data, regional-scale remote sensing physiological image data, and historical yield data for each sub-region within the target area; The data processing module is used to calibrate and detect faults in the ground point source environmental data to generate environmental baseline features, and simultaneously perform multispectral feature extraction and pattern recognition on the regional scale remote sensing physiological image data to generate physiological representation vectors. The data analysis module is used to fuse the environmental baseline features with the physiological representation vectors at spatiotemporal scales to construct a multidimensional state space model. The crop growth simulation module is used to input the multidimensional state space model into a preset crop growth dynamic model, which includes physiological response sensitivity parameters for characterizing crop drought resistance; and to generate a set of expected growth trajectories of the sub-region under various preset water supply conditions by performing difference equation simulation on the crop growth dynamic model. The marginal loss assessment module is used to determine the expected output loss of the sub-region under the corresponding water supply conditions based on the expected growth trajectory set; and to construct a marginal output loss matrix based on the expected output loss. The resource optimization and scheduling module is used to process the marginal output loss matrix of each sub-region under the preset total water resource constraint and output the differential resource allocation vector for each sub-region. The feedback deviation module is used to acquire crop physiological characteristic values ​​in real time and calculate the feedback deviation residual between the crop physiological characteristic values ​​and the matching trajectory in the expected growth trajectory set. An adaptive self-learning module is used to synchronously correct the crop growth dynamic model based on the feedback deviation residuals and dynamically adjust the marginal yield loss matrix.

Images