Machine learning based prediction method for evolution characteristics of water requirement of dryland farmland

Abstract

The application discloses a method for predicting the evolution characteristics of water demand in dry farmland based on machine learning, and particularly relates to the technical field of intelligent agricultural data processing. The method collects multi-source sensing signals, maps and constructs a root-soil-air continuum characteristic space covering multi-spectral reflection, deep water content and meteorological forcing sequence; extracts the radiation attenuation gradient characteristics under continuous rain-free stress to construct a surface crust degradation index sequence, and tracks the water consumption penetration characteristics to extract the crop root water-seeking memory penetration lower boundary; discretizes the farmland into spatial topological nodes, analyzes the spatial heterogeneity difference of crust degradation between nodes and converts it into a water transverse conduction resistance factor, fuses the spatial distance attribute to construct a micro-topographic water potential dynamic graph matrix; inputs the matrix and meteorological sequence into a spatio-temporal graph attention network, and embeds a balance residual loss term with the root penetration lower boundary as the penetration cutoff boundary, and finally outputs the water demand dynamic prediction trajectory.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture data processing technology, specifically a method for predicting the water demand evolution characteristics of dryland farmland based on machine learning. Background Technology

[0002] Dryland farmland is a crucial component of agricultural production, with crop growth highly dependent on natural rainfall and limited irrigation water resources. With the dynamic evolution of the climate environment, the evaporation process of soil moisture and the water consumption patterns of crops in dryland exhibit highly complex nonlinear changes. Accurately predicting the long-term water demand evolution characteristics of dryland farmland is of great significance for guiding agricultural water resource allocation, optimizing drought resistance strategies, and ensuring crop production. With the development of sensing technology and data science, analyzing crop water demand using multi-source monitoring data and algorithmic models has become an important direction in the development of smart agriculture.

[0003] However, in practical forecasting applications, dryland surfaces are highly susceptible to physical degradation such as crusting or cracking under prolonged drought, and crop roots undergo morphological changes to adapt to drought by moving deeper into water-rich areas. Existing forecasting methods largely rely on static meteorological parameters and surface soil moisture data for empirical formula extrapolation or fitting to conventional time-series models, neglecting the changes in water evapotranspiration resistance caused by soil structure degradation and the expansion of the available water boundary due to the dynamic downward movement of root depth. Furthermore, existing technologies typically treat farmland as isolated grid computing units, failing to consider micro-topographical differences and changes in lateral water conduction impedance caused by surface degradation. This leads to distorted predictions that deviate from objective water balance laws when facing long-term evolution, insufficient adaptability to dynamic changes during drought intensification, and difficulty in providing accurate water demand evolution trajectories. Therefore, it remains necessary to provide a machine learning-based method for predicting the water demand evolution characteristics of dryland farmland to improve the accuracy and rationality of predictions under extreme drought conditions. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a machine learning-based method for predicting the water demand evolution characteristics of dryland farmland to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the water demand evolution characteristics of dryland farmland based on machine learning, comprising the following steps: S1, collecting multi-source sensor signals from dryland farmland and mapping them to construct a root-soil-gas continuum feature space; the root-soil-gas continuum feature space is composed of a surface multispectral reflectance sequence, a deep profile water content dynamic sequence, and a meteorological forcing time series mapping; S2, extracting the radiation attenuation gradient characteristics of the red band and near-infrared band in the surface multispectral reflectance sequence under continuous rainless stress, constructing a surface crust degradation index sequence; tracking the water consumption penetration characteristics of the deep profile water content dynamic sequence in the vertical direction, identifying the water stress retardation interface, and extracting the lower boundary of the crop root water-seeking memory penetration; S3. Discretize dryland farmland into spatial topological nodes, analyze the spatial heterogeneous distribution differences of surface crust degradation index sequences between adjacent spatial topological nodes, transform the spatial heterogeneous distribution differences into surface water lateral conduction impedance factors, and integrate the surface water lateral conduction impedance factors into the spatial distance attributes of spatial topological nodes to construct a micro-topographic water potential dynamic map matrix; S4. Input the micro-topographic water potential dynamic map matrix and meteorological forcing time series into a spatiotemporal graph attention network to perform spatiotemporal feature aggregation operations, embed the balance residual loss term in the objective optimization function into the spatiotemporal graph attention network, configure the lower bound of crop root water-seeking memory penetration as the deep infiltration flux cutoff boundary of the objective optimization function, and output the dynamic prediction trajectory of water demand for dryland farmland.

[0006] In a preferred embodiment, the specific process of acquiring multi-source sensor signals from dryland farmland and mapping them to construct a root-soil-atmosphere continuum feature space is as follows: acquiring UAV multispectral remote sensing images, vertical soil moisture sensor network signals, and meteorological observation data; performing radiometric calibration and atmospheric correction on the UAV multispectral remote sensing images, extracting time-series pixel tensors, and generating a surface multispectral reflectance sequence; performing depth interpolation reconstruction and timestamp alignment mapping on the vertical soil moisture sensor network signals to generate a deep profile moisture content dynamic sequence; performing multidimensional forcing factor feature encoding on the meteorological observation data to generate a meteorological forcing time series; and jointly mapping the surface multispectral reflectance sequence, the deep profile moisture content dynamic sequence, and the meteorological forcing time series to construct a root-soil-atmosphere continuum feature space.

[0007] In a preferred embodiment, the specific process of extracting the radiation attenuation gradient features of the red and near-infrared bands in the surface multispectral reflectance sequence under continuous drought stress and constructing the surface crust degradation index sequence is as follows: Based on the meteorological forcing time series, a continuous drought stress time window is determined; within this window, the time-varying reflectance tensors of the red and near-infrared bands in the surface multispectral reflectance sequence are extracted; nonlinear divergence feature extraction is performed on the time-varying reflectance tensors to obtain the radiation attenuation gradient features of the surface soil as drought stress intensifies; the radiation attenuation gradient features are input into a preset surface structure degradation response function for mapping, outputting the surface soil reflectance attenuation rate parameter; the cumulative evolution trend of the surface soil reflectance attenuation rate parameter over time is tracked to generate the surface crust degradation index sequence.

[0008] In a preferred embodiment, the specific process of tracking the water consumption penetration characteristics of the deep profile water content dynamic sequence in the vertical direction, identifying the water stress retardation interface, and extracting the lower bound of crop root water-attracting memory penetration is as follows: Perform spatial partial derivative analysis on the deep profile water content dynamic sequence in the vertical depth dimension to obtain the water content vertical gradient distribution matrix; extract the spatial critical depth where the water consumption rate exhibits a step-like decay in the time dimension from the water content vertical gradient distribution matrix, and mark the spatial critical depth as the water stress retardation interface; track the vertical downward evolution trajectory of the water stress retardation interface within historical drought stress cycles, and extract the maximum penetration depth extremum in the vertical downward evolution trajectory; convert the maximum penetration depth extremum into a spatial limit for adaptive downward extension of root morphology, generating the lower bound of crop root water-attracting memory penetration.

[0009] In a preferred embodiment, the specific process of discretizing dryland farmland into spatial topological nodes and analyzing the spatial heterogeneous distribution differences of the surface crust degradation index sequence between adjacent spatial topological nodes is as follows: The dryland farmland is mapped into a two-dimensional grid space, and the coordinate system features of the center of the grid cell in the two-dimensional grid space are extracted to construct a set of spatial topological nodes; the specific logic of performing spatial coordinate alignment mapping between the surface crust degradation index sequence and the set of spatial topological nodes and assigning dynamic degradation attributes to the spatial topological nodes is as follows: Based on the center coordinate system features of each spatial topological node, under the set current time slice, the spatial dimension resampling of the surface crust degradation index sequence with spatial continuity is performed using a bilinear interpolation algorithm, and the transient feature data of the corresponding coordinates are accurately matched to the center of each two-dimensional grid cell, thereby assigning each spatial topological node a dynamic degradation attribute value that is updated synchronously with time.

[0010] In a preferred embodiment, the specific process of transforming spatial heterogeneous distribution differences into surface water lateral conduction impedance factors and integrating these factors into the spatial distance attributes of spatial topological nodes to construct a micro-topographic water potential dynamic map matrix is ​​as follows: A soil moisture movement hydraulic retardation response function is introduced to map spatial heterogeneous distribution differences into surface water lateral conduction impedance factors that restrict the lateral transport of surface water; Euclidean spatial coordinate distribution features between adjacent nodes are extracted based on the spatial topological node set to generate a spatial topological node spatial distance attribute matrix; a spatial penalty attenuation mapping is performed on the spatial topological node spatial distance attribute matrix, and surface water lateral conduction impedance factors are injected into the spatial penalty attenuation mapping mechanism to generate a dynamic spatial weighted connection matrix; the dynamic spatial weighted connection matrix is ​​combined with the micro-topographic water flow convergence topology extracted from the digital elevation model for cross-modal feature stitching to construct a micro-topographic water potential dynamic map matrix.

[0011] In a preferred embodiment, the specific process of inputting the micro-topographic water potential dynamic map matrix and the meteorological forcing time series into a spatiotemporal graph attention network to perform spatiotemporal feature aggregation operations, and embedding the balance residual loss term in the objective optimization function into the spatiotemporal graph attention network is as follows: The micro-topographic water potential dynamic map matrix is ​​input into the spatiotemporal graph attention network as a spatial topological prior structure, driving the spatial attention mechanism to perform neighborhood feature aggregation on the meteorological forcing time series, extracting spatial dependence features of water transport; a time-dimensional self-attention encoding operation is performed on the spatial dependence features of water transport to capture the cumulative effect of temporal drought and generate a high-order spatiotemporal water demand feature expression; a farmland water flux balance equation is constructed that couples precipitation variables, evapotranspiration variables, deep infiltration variables, and soil water storage variables; the high-order spatiotemporal water demand feature expression is mapped to a predicted water demand flux variable, the predicted water demand flux variable is substituted into the farmland water flux balance equation to generate a balance residual loss term, and the balance residual loss term is embedded into the objective optimization function of the spatiotemporal graph attention network.

[0012] In a preferred embodiment, the specific process of configuring the lower bound of crop root water-seeking memory penetration as the deep infiltration flux cutoff boundary of the objective optimization function and outputting the dynamic prediction trajectory of water demand for dryland farmland is as follows: The deep infiltration flux solution domain is extracted from the farmland water flux balance equation, and the lower bound of crop root water-seeking memory penetration is introduced into the deep infiltration flux solution domain as the vertical water interaction limit depth; a water flux cutoff condition is established at the vertical water interaction limit depth, forcibly terminating invalid deep water infiltration calculation components exceeding the lower bound of crop root water-seeking memory penetration, generating the deep infiltration flux cutoff boundary; under the constraint of the deep infiltration flux cutoff boundary, the backpropagation of the objective optimization function of the spatiotemporal graph attention network is performed, updating the weight parameters of the spatiotemporal graph attention network until the objective optimization function converges, completing the adaptive optimization of network parameters; the spatiotemporal graph attention network with optimized parameters performs forward evolution extrapolation on the meteorological forced time series, generating the dynamic prediction trajectory of water demand for dryland farmland that characterizes the evolution of water demand within the future time window.

[0013] The technical effects and advantages of this invention are as follows: (1) A machine learning-based method for predicting the water demand evolution characteristics of dryland farmland, after collecting multi-source sensor signals and mapping to construct a root-soil-air continuum feature space, extracts the radiation attenuation gradient features of the surface multispectral reflectance sequence to construct the surface crust degradation index sequence, and tracks the water consumption penetration features of the deep profile water content dynamic sequence in the vertical direction to extract the lower boundary of crop root water-seeking memory penetration. Thus, it can overcome the shortcomings of existing methods that rely on static environmental parameters and ignore the special physicochemical feedback of dryland, and transform the drought-induced surface soil structure degradation and root adaptive morphological changes into quantifiable dynamic feature boundaries, thereby improving the perception ability and feature extraction accuracy of the prediction process for the complex evolution phenomena of the dryland subsurface.

[0014] (2) A machine learning-based method for predicting the water demand evolution characteristics of dryland farmland discretizes dryland farmland into spatial topological nodes. The spatial heterogeneous distribution differences of the surface crust degradation index sequence between adjacent nodes are transformed into the surface water lateral conduction impedance factor and fused into the spatial distance attribute to construct a micro-topographic water potential dynamic map matrix. Then, it is input into a spatiotemporal map attention network along with a meteorological forcing time series, and a balance residual loss term with the root water-seeking memory penetration lower boundary as the cutoff boundary is embedded in the network. This can improve the problem that independent grid computing in the existing technology does not consider the lateral redistribution of water, and constrain the iterative process of network parameters through the farmland water balance equation, avoiding the distortion results that deviate from objective physical norms in long-term predictions of pure data-driven models, and improving the prediction reliability and model generalization adaptability in drought evolution scenarios.

[0015] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0016] Figure 1 This is a flowchart of the machine learning-based method for predicting the water demand evolution characteristics of dryland farmland according to the present invention. Figure 2 This is a schematic diagram illustrating the response evolution of the lateral conduction impedance factor of surface moisture in an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating the identification of the step change rate of water consumption rate and the stress retardation interface in an embodiment of the present invention. Figure 4 This is a schematic diagram of the distribution of the dynamic spatial weight connection matrix in an embodiment of the present invention. Detailed Implementation

[0017] This application's embodiments address the problems of prediction distortion and poor generalization adaptability in existing technologies when facing continuous drought, due to neglecting soil physical degradation and not being constrained by farmland water conservation, by using a machine learning-based method for predicting the evolution characteristics of water demand in dryland farmland.

[0018] The overall approach of the scheme in this application is as follows: First, multi-source sensor signals from dryland farmland are collected and mapped to construct a root-soil-atmosphere continuum feature space encompassing surface multispectral reflectance sequences, deep profile moisture content dynamic sequences, and meteorological forcing time series. Then, the radiative attenuation gradient features of the multispectral reflectance sequences are extracted to generate a surface crust degradation index sequence, and the evolution of water consumption in the deep profile moisture content sequence in the vertical direction is tracked to obtain the lower bound of crop root water-seeking memory penetration. Next, the dryland farmland is discretized into spatial topological nodes, and the spatial differences in crust degradation indices between adjacent nodes are analyzed and transformed into a surface water transverse conduction impedance factor. This impedance factor is then fused into the spatial distance attribute of the nodes to establish a micro-topographic water potential dynamic map matrix. Finally, the micro-topographic water potential dynamic map matrix and the meteorological forcing time series are input into a spatiotemporal map attention network for spatiotemporal feature aggregation operations. Simultaneously, the lower bound of crop root water-seeking memory penetration is configured as a deep infiltration flux cutoff boundary, and a balance residual loss term is calculated in the network to constrain parameter updates. Finally, a dynamic prediction trajectory of water demand for dryland farmland conforming to the evolution mechanism is output.

[0019] Example 1; please refer to Figure 1This invention provides a technical solution: a method for predicting the water demand evolution characteristics of dryland farmland based on machine learning, comprising the following steps: S1, collecting multi-source sensor signals from dryland farmland and mapping them to construct a root-soil-gas continuum feature space; the root-soil-gas continuum feature space is composed of a surface multispectral reflectance sequence, a deep profile water content dynamic sequence, and a meteorological forcing time series mapping; S2, extracting the radiation attenuation gradient characteristics of the red band and near-infrared band in the surface multispectral reflectance sequence under continuous rainless stress, constructing a surface crust degradation index sequence; tracking the water consumption penetration characteristics of the deep profile water content dynamic sequence in the vertical direction, identifying the water stress retardation interface, and extracting the lower boundary of the crop root water-seeking memory penetration; S3 1. Discretize dryland farmland into spatial topological nodes, analyze the spatial heterogeneous distribution differences of surface crust degradation index sequences between adjacent spatial topological nodes, transform the spatial heterogeneous distribution differences into surface water lateral conduction impedance factors, and integrate the surface water lateral conduction impedance factors into the spatial distance attributes of spatial topological nodes to construct a micro-topographic water potential dynamic map matrix; S4. Input the micro-topographic water potential dynamic map matrix and meteorological forcing time series into a spatiotemporal graph attention network to perform spatiotemporal feature aggregation operations, embed the balance residual loss term in the objective optimization function into the spatiotemporal graph attention network, configure the lower bound of crop root water-seeking memory penetration as the deep infiltration flux cutoff boundary of the objective optimization function, and output the dynamic prediction trajectory of water demand for dryland farmland.

[0020] In this implementation plan, step S1 directly performs multi-dimensional automated collection and formatting of basic data on the arid land environment. Since a single sensor cannot comprehensively reflect the moisture status of farmland, this step acquires multispectral reflectance data reflecting the physical state of the surface, profile moisture content data reflecting groundwater storage, and external meteorological change data. By combining these data, a root-soil-atmosphere continuum feature space is constructed. This is a multi-dimensional data matrix that fully maps the actual physical location of water exchange between plant roots, soil medium, and the atmospheric environment at the data level. The surface multispectral reflectance sequence records the change in surface reflectance over time, the deep profile moisture content dynamic sequence reflects the actual moisture content changes at different underground depths, and the meteorological forcing time series includes temperature and external meteorological factors driving water evaporation. These three parts together provide historical samples and environmental boundary conditions for prediction.

[0021] Step S2 processes the collected raw sequence data to analyze the degree of physical damage to the dryland surface and the biological adaptation limit of crops. First, multispectral data is processed to calculate the rate of change of reflectance in the red and near-infrared bands during prolonged drought, i.e., the radiative attenuation gradient characteristics. This quantifies the degree of hardening and compaction of the surface soil due to water loss, forming a surface crust degradation index sequence reflecting the declining trend of soil permeability. Simultaneously, deep soil moisture data is analyzed to observe the vertical diffusion depth of water loss over time. When a certain depth is reached, water consumption decreases sharply; this layer is marked as the water stress retardation interface. Based on this, the lower limit of crop root hydrotropism memory penetration is determined. This lower limit represents the maximum effective depth that crops can reach under long-term drought stress to encourage root penetration downwards, objectively defining the vertical physical baseline of farmland water consumption.

[0022] Step S3 divides the continuous farmland area into multiple independent grid computing units, which constitute spatial topological nodes, thereby reconstructing the spatial connectivity between farmland plots. Due to differences in drought and crusting levels in each node area, the resistance encountered by surface water flowing between different grids varies, and this uneven physical degradation manifests as spatial heterogeneous distribution differences. This step transforms these differences into a quantified surface water lateral conduction resistance factor, representing the actual resistance of surface crusting to the lateral infiltration and flow of water. Based on this, and combining the actual physical distances between each grid node, a micro-topographic water potential dynamic map matrix is ​​generated. This matrix is ​​a network topology structure with dynamic connection weights; the weight values ​​between nodes are jointly determined by spatial distance and surface crusting resistance, reflecting the non-uniform transport path of water in the horizontal direction of farmland under complex surface degradation conditions.

[0023] Step S4 inputs the previously constructed micro-topographic water potential dynamic map matrix and meteorological forcing time series into the spatiotemporal graph attention network to extract the spatial dependencies of water and its cumulative evolution characteristics over time. A balanced residual loss term is added to the objective optimization function of the network training; this is a mathematical constraint penalty mechanism representing the need to maintain mass conservation between total farmland water input and consumption. Combined with the maximum root depth extracted in the previous step, it is used as the cutoff boundary for deep infiltration flux, forcing the network model to exclude water exceeding the root depth from effective consumption when calculating deep infiltration losses. Through forward computation of the network with clear physical boundary constraints, the final output is a dynamic prediction trajectory of water demand for dryland farmland that conforms to the objective laws of water consumption evolution, completing the quantitative calculation of future water evolution characteristics.

[0024] Specifically, the process of collecting multi-source sensor signals from dryland farmland and mapping them to construct a root-soil-atmosphere continuum feature space is as follows: acquiring UAV multispectral remote sensing images, vertical soil moisture sensor network signals, and meteorological observation data; performing radiometric calibration and atmospheric correction on the UAV multispectral remote sensing images, extracting time-series pixel tensors, and generating a surface multispectral reflectance sequence; performing depth interpolation reconstruction and timestamp alignment mapping on the vertical soil moisture sensor network signals to generate a deep profile moisture content dynamic sequence; performing multidimensional forcing factor feature encoding on the meteorological observation data to generate a meteorological forcing time series; and jointly mapping the surface multispectral reflectance sequence, the deep profile moisture content dynamic sequence, and the meteorological forcing time series to construct a root-soil-atmosphere continuum feature space.

[0025] In this implementation scheme, the acquisition of multi-source sensor signals relies on synchronous sampling in time and space by hardware devices. A UAV equipped with a multispectral camera flies along a set route to acquire surface reflection images, a weather station simultaneously records environmental meteorological data, and moisture sensors buried at different depths in the soil collect moisture signals from the undisturbed soil. For generating the dynamic sequence of deep profile moisture content, since physical moisture sensors can only be deployed at limited discrete depths and cannot directly reflect the moisture distribution of a continuous profile, depth interpolation reconstruction of the vertical soil moisture sensor network signals is required. Specifically, continuous moisture content is calculated using the following composite formula of spatial inverse distance and physical attenuation: ;in, In the formula, This represents the dynamic sequence value of water content in the deep profile at a horizontal coordinate of x and y, a vertical depth of z, and a time of t; S represents the total number of layers of soil moisture sensors buried. Indicates the sensor hierarchy number; v represents the iterative auxiliary number used for normalization summation. Indicates physical depth First Raw moisture content data collected by the layer sensor at time t; This represents the depth reconstruction weight coefficient. The method for determining this weight coefficient is based on the physical distance attenuation law of soil vertical profile. The sensor's original value that is closer to the target resolution depth z is assigned a greater weight. The soil bulk density decay constant is represented by a value obtained through pre-analysis of soil samples taken from the target farmland using a ring corer. After interpolation and reconstruction, the generated deep profile moisture content dynamic sequence, surface multispectral reflectance sequence, and coded meteorological forcing time series are spliced ​​together in a multidimensional matrix on timestamp t and spatial coordinates x and y to complete the mapping and construction of the root-soil-atmosphere continuum feature space.

[0026] Specifically, the process of extracting the radiation attenuation gradient features of the red and near-infrared bands in the surface multispectral reflectance sequence under continuous drought stress and constructing the surface crust degradation index sequence is as follows: Based on the meteorological forcing time series, a continuous drought stress time window is determined. Within this window, the time-varying reflectance tensors of the red and near-infrared bands in the surface multispectral reflectance sequence are extracted. Nonlinear divergence feature extraction is performed on the time-varying reflectance tensors to obtain the radiation attenuation gradient features of the surface soil as drought stress intensifies. The radiation attenuation gradient features are input into a pre-set surface structure degradation response function for mapping, outputting the surface soil reflectance attenuation rate parameter. The cumulative evolution trend of the surface soil reflectance attenuation rate parameter over time is tracked to generate the surface crust degradation index sequence.

[0027] In this implementation plan, the assessment of surface crust degradation focuses on the dynamic evolution of soil surface reflectance characteristics during prolonged drought. Based on meteorological forcing time series, continuous rainless stress time windows with zero continuous rainfall and solar radiation exceeding the baseline value are selected, and red and near-infrared band data for corresponding spatial coordinates are extracted within these time periods. The radiation attenuation gradient characteristics of surface soil as drought stress intensifies are obtained through the following nonlinear divergence calculation: In the formula, This represents the radiative attenuation gradient characteristics at coordinates x and y at time t; This represents the near-infrared reflectance intercepted at time t; This represents the reflectivity of the red light band intercepted at time t; and These represent the reflectivity of the corresponding band intercepted at the previous adjacent time step t-1; This represents a preset minimum positive constant to prevent the denominator from being zero. This formula characterizes the relative changes in red and near-infrared reflectance caused by soil water loss. Subsequently, this gradient characteristic is substituted into a preset surface structure degradation response function for cumulative mapping, generating a surface crust degradation index sequence: In the formula, This represents the sequence value of the surface crust degradation index generated at time t; This represents the degradation exponential ground state at the previous time step; This represents the crust sensitivity coefficient, which characterizes the rate of physical shrinkage and hardening of a specific soil texture during dehydration. The specific calibration process for its pre-assignment is as follows: extract the mass percentages of clay, silt, and sand from the historical soil particle size distribution survey data of the target farmland; input these mass percentage parameters into a pre-built soil transfer function to derive pore shrinkage characteristic parameters; perform normalization on these characteristic parameters; and then calibrate them as the crust sensitivity coefficient for this grid node. Through continuous cumulative calculations over time, a quantitative characterization of the nonlinear structural degradation process of the dryland surface layer caused by water loss is achieved.

[0028] Specifically, the process of tracking the water consumption penetration characteristics of the deep profile water content dynamic sequence in the vertical direction, identifying the water stress retardation interface, and extracting the lower bound of crop root water-seeking memory penetration is as follows: Perform spatial partial derivative analysis on the deep profile water content dynamic sequence in the vertical depth dimension to obtain the water content vertical gradient distribution matrix; extract the spatial critical depth where the water consumption rate exhibits a step-like decay in the time dimension from the water content vertical gradient distribution matrix, and mark the spatial critical depth as the water stress retardation interface; track the vertical downward evolution trajectory of the water stress retardation interface within historical drought stress cycles, extract the maximum penetration depth extremum and the corresponding water consumption downward exploration rate in the vertical downward evolution trajectory; combine the biological absolute root extreme value of the target crop to construct a dynamic downward extension compensation function driven by the meteorological forcing time series; fuse the maximum penetration depth extremum with the output of the dynamic downward extension compensation function to generate a crop root water-seeking memory penetration lower bound that is dynamically updated in real time with future drought evolution.

[0029] In this implementation scheme, for extreme drought stress conditions exceeding historical levels, a dynamic downward extension compensation mechanism is introduced, taking into account the adaptive downward extension characteristics of crop roots. When the predicted meteorological forcing exceeds historical extremes, the biological and physiological limit depth of the target crop is used as the constraint baseline, driving the crop root system's water-seeking memory penetration lower bound to dynamically extend the calculation according to the historical water consumption downward rate. This mechanism, while maintaining the conservation of physical water, achieves dynamic adaptive updating of the truncated boundary, reduces the calculation bias of the model under unknown extreme meteorological conditions, and improves the generalization and adaptability of the network model to drought data outside the boundary. Identifying the crop root depth essentially defines the physical lower bound of water consumption for subsequent model calculations, avoiding unrealistic water consumption predictions from the neural network algorithm in deep, rootless soil layers. Based on the generated deep profile water content dynamic sequence, the partial derivatives of water consumption rates in both the vertical and temporal dimensions are calculated to generate a vertical gradient distribution matrix of water content. In the formula, This represents the rate of water loss in the soil at depth z at time t. Indicates the time step of adjacent detections; This indicates the profile moisture content at the previous time step; The elements in the vertical gradient distribution matrix of water content represent the step rate of change of water consumption rate in the vertical direction. Indicates the analytical step size in depth space; This represents the rate of water consumption at a shallower, adjacent layer. During droughts, the rate of water consumption is extremely high in densely rooted areas, while water consumption approaches zero in very deep areas untouched by roots. When the matrix elements... The absolute value of the space step threshold is exceeded at a certain depth. When this occurs, it indicates that water consumption exhibits a nonlinear step-like decline at this depth, which is then identified as the water stress retardation interface. Regarding the critical threshold of spatial step jumps The method for determining the fully irrigated control group experiment is as follows: The same batch of crops is planted in an in-situ observation area with soil texture similar to the target farmland. Supplemental irrigation is used to maintain the field water holding capacity at over 85% to prevent water stress. During the stable root development period, a vertically deployed soil moisture sensor network is used to synchronously collect the water content of each layer, calculating the baseline water consumption rate at each depth under drought stress. The difference in consumption rates between adjacent physical depth layers is compared, and the absolute value of the maximum rate drop on the baseline evolution profile is extracted. Five percent of this maximum absolute value is used as the baseline. In a specific semi-arid experimental field on the Loess Plateau, the highest daily water consumption rate was observed in the surface to shallow root zone (0-40cm), reaching 3.2mm / d, while in the deep subsoil (below 80cm), this rate was less than 0.1mm / d, with a maximum absolute difference in rate of 3.1mm / d. Based on the above mechanism, the spatial step threshold under this working condition was calculated and calibrated. mm / d. Finally, the maximum extreme value of vertical downward displacement of the hindrance interface was extracted throughout the entire historical drought cycle: In the formula, This indicates that the extracted crop root hydrotaxis memory penetrates the lower boundary; This represents the time set of historical drought stress cycles. This lower bound characterizes the maximum spatial physical limit to which crop roots can extend downwards in search of deeper water under extreme drought stress. For example... Figure 3 As shown, this graph uses vertical absolute depth as the horizontal axis and the step rate of change of water consumption rate (matrix element) as the horizontal axis. The vertical axis represents the absolute value of water consumption. In shallow soil and areas with active root systems, the vertical gradient of water consumption remains within the normal scatter plot distribution range. However, at specific extreme depths, the water consumption rate experiences a precipitous drop, resulting in an extreme abrupt change in the step rate of change (as shown by the peak scatter plot in the figure). When this abrupt change scatter plot breaks upward through the horizontal baseline in the figure (i.e., the spatial step threshold), the vertical gradient becomes extremely significant. When this occurs, the system accurately locates the depth as the water stress retardation interface. By extracting the downward extreme values ​​from historical cycles, the system can objectively identify the lower boundary of the crop's root hydrotropism memory penetration.

[0030] Specifically, the process of discretizing dryland farmland into spatial topological nodes and analyzing the spatial heterogeneous distribution differences of the surface crust degradation index sequence between adjacent spatial topological nodes is as follows: The dryland farmland is mapped into a two-dimensional grid space, and the coordinate system features of the center of the grid cell in the two-dimensional grid space are extracted to construct a set of spatial topological nodes; the spatial coordinate alignment mapping of the surface crust degradation index sequence and the set of spatial topological nodes is performed, and the specific logic of assigning dynamic degradation attributes to the spatial topological nodes is as follows: Based on the center coordinate system features of each spatial topological node, under the set current time slice, the spatially continuous surface crust degradation index sequence is resampled in spatial dimension using a bilinear interpolation algorithm, and the transient feature data of the corresponding coordinates are accurately matched to the center of each two-dimensional grid cell, thereby assigning each spatial topological node a dynamic degradation attribute value that is updated synchronously with time.

[0031] In this implementation scheme, the two-dimensional geographic coordinate system of the dryland farmland is divided into a uniform two-dimensional grid space. The latitude and longitude coordinates of the geometric center of each grid cell are extracted as the reference position, generating a set containing several discretized spatial topological nodes. The surface crust degradation index sequence generated in the previous steps is obtained and registered to the corresponding spatial topological nodes according to their spatial location coordinates, so that each spatial topological node has a dynamically updated degradation attribute value over time. Subsequently, a connectivity distance search radius is set, and nodes whose spatial Euclidean distance is less than the connectivity distance search radius are established as adjacent connected relationships, constructing an adjacency connected network structure. Under the constraints of this network structure, the spatial feature gradient tensor between adjacent spatial topological nodes is calculated: In the formula, This represents the spatial feature gradient tensor between the core node p and its neighboring node q. This represents the value of the dynamic degradation attribute of the core node p; This represents the value of the dynamic degradation attribute of the adjacent node q; This represents the actual Euclidean distance between the core node p and its neighboring node q; Let represent the unit direction vector from core node p to neighboring node q; where represents the dynamic degradation attribute value of core node p. It is a sequence of surface crust degradation index. Based on the values ​​extracted from the coordinate mapping at time t, the degree of nonlinear deviation of the spatial feature gradient tensor in the evolution direction is analyzed to generate spatial heterogeneous distribution differences: In the formula, This represents the spatial heterogeneous distribution difference between the core node p and its neighboring node q; Indicates the scaling factor; Represents the nonlinear saturation parameter; This represents the degradation potential energy weighting constant; and The determination method is to obtain the water-holding characteristic curve of the historical soil sampling data of the target farmland by performing multiple regression fitting; The values ​​are directly assigned based on empirical values ​​from a database of physical properties of soils of the same type.

[0032] Specifically, the process of transforming spatial heterogeneous distribution differences into surface water lateral conduction impedance factors and integrating these factors into the spatial distance attributes of spatial topological nodes to construct a dynamic micro-topographic water potential map matrix is ​​as follows: A soil moisture movement hydraulic retardation response function is introduced to map spatial heterogeneous distribution differences into surface water lateral conduction impedance factors that restrict the lateral transport of surface water; Euclidean spatial coordinate distribution features between adjacent nodes are extracted based on the spatial topological node set to generate a spatial topological node spatial distance attribute matrix; a spatial penalty attenuation mapping is performed on the spatial topological node spatial distance attribute matrix, and the surface water lateral conduction impedance factor is injected into the spatial penalty attenuation mapping mechanism to generate a dynamic spatial weighted connection matrix; the dynamic spatial weighted connection matrix is ​​combined with the micro-topographic water flow convergence topology extracted from the digital elevation model for cross-modal feature stitching to construct the dynamic micro-topographic water potential map matrix.

[0033] In this implementation scheme, the hydraulic resistance response function of soil moisture movement is used to process the above-mentioned spatial heterogeneous distribution differences. This response function is calculated by mapping according to the Logistic curve law and outputs the surface water lateral conduction resistance factor: In the formula, This represents the lateral conduction impedance factor of surface water between the core node p and its adjacent node q. This represents the upper limit constant of the reference transverse impedance; This represents the rate parameter of the hindrance response; This represents the threshold for heterogeneous abrupt change. The method for determining this threshold is to calibrate the critical point of permeability abrupt change when the macropores in the soil close under extreme drought conditions in the target farmland. For example, in a permeability penetration experiment targeting the surface layer of loam, when spatial heterogeneous distribution differences are measured... As the soil approaches the critical point of 5.0, the macropore structure of the surface soil closes, and the lateral permeability coefficient of water drops sharply. The system uses this magnitude to determine the heterogeneity threshold. Calibrated as and the upper limit constant of the reference transverse impedance. Normalization converges to 1.0. For example... Figure 2 As shown in the figure, this graph illustrates the lateral conduction impedance factor. Differences in spatial heterogeneous distribution The evolution of the changing Logistic curve. As shown in the figure, when the spatial heterogeneous distribution differences between adjacent nodes are small and have not reached the threshold of heterogeneous abrupt change... At this time, the conduction impedance factor is at a low level and shows a gradual upward trend; when the spatial heterogeneous distribution difference reaches and exceeds the heterogeneous abrupt change threshold... When this occurs, it indicates a sudden change in the soil's physicochemical properties (such as the closure of macropores leading to the blockage of lateral transport), and the conduction impedance factor exhibits a nonlinear step increase, eventually gradually saturating and approaching the upper limit constant of the reference lateral impedance. This mapping curve effectively reflects the physical obstruction effect of surface crusting on the lateral transport of water. The Euclidean spatial coordinate distribution characteristics between adjacent nodes are extracted to form a spatial distance attribute matrix for spatial topology nodes. Based on this matrix, a spatial penalty attenuation mapping calculation is performed, and a surface water lateral conduction impedance factor is injected into the calculation to generate a dynamic spatial weighted connection matrix. In the formula, This represents the connection weight element between the core node p and its adjacent node q in the dynamic space weighted connection matrix; This represents the spatial attenuation weighting coefficient. The method for determining the spatial attenuation weighting coefficient is to extract the reciprocal of the range parameter by performing semivariance function calculation and analysis on historical soil moisture observation data. This indicates impedance penalty sensitivity. Elevation data for each node is extracted from the digital elevation model data, and the micro-topographical flow convergence topology is analyzed. In the formula, This represents the water flow potential energy driving term from the core node p to the adjacent node q in the micro-topographic water flow convergence topology. This represents the absolute elevation value of the core node p; This represents the absolute elevation value of adjacent node q. The dynamic spatial weight concatenation matrix elements are then combined with the micro-topographic water flow convergence topology elements through cross-modal feature concatenation to generate the edge feature elements of the micro-topographic water potential dynamic map matrix. In the formula, This represents the multi-channel feature vector matrix element between two nodes in the micro-topography water potential dynamic map matrix; This indicates the cross-modal feature concatenation operator; This represents the terrain elevation driving gain constant. For example... Figure 4 As shown, the horizontal base plane in the figure is composed of topological nodes X and Y, representing the two-dimensional grid coordinate space of the discretized farmland; the vertical Z-axis weights represent the dynamic connection weights of the water potential generated between adjacent topological nodes (i.e., matrix elements). This three-dimensional bar chart visually illustrates the obstruction of lateral water conduction caused by arid land physical degradation: the connection weights are not only controlled by the distribution characteristics of spatial Euclidean coordinates, exhibiting an exponential decay with distance (i.e., spatial penalty), but are also simultaneously subject to the dynamic constraints of the surface water lateral conduction resistance factor caused by surface crust degradation. Figure 4As can be seen, due to the intervention of micro-topographic differences and heterogeneous crust impedance, the weighted columns between different nodes exhibit a non-uniform, undulating distribution. This weight allocation based on physical mechanisms completely corrects the error of equal grid connectivity in traditional calculations, accurately reproducing the nonlinear interference process of damaged landform boundaries on the lateral transport of water in real arid land.

[0034] Specifically, the process of inputting the micro-topographic water potential dynamic map matrix and the meteorological forcing time series into the spatiotemporal graph attention network to perform spatiotemporal feature aggregation operations, and embedding the balance residual loss term in the objective optimization function into the spatiotemporal graph attention network is as follows: The micro-topographic water potential dynamic map matrix is ​​input into the spatiotemporal graph attention network as a spatial topological prior structure, driving the spatial attention mechanism to perform neighborhood feature aggregation on the meteorological forcing time series and extracting spatial dependence features of water transport; a time-dimensional self-attention encoding operation is performed on the spatial dependence features of water transport to capture the cumulative effect of time-series drought and generate a high-order spatiotemporal water demand feature expression; a farmland water flux balance equation is constructed that couples precipitation variables, evapotranspiration variables, deep infiltration variables, and soil water storage variables; the high-order spatiotemporal water demand feature expression is mapped to the predicted water demand flux variable, the predicted water demand flux variable is substituted into the farmland water flux balance equation to generate a balance residual loss term, and the balance residual loss term is embedded into the objective optimization function of the spatiotemporal graph attention network.

[0035] In this implementation scheme, the spatiotemporal map attention network relies on a multi-source sensor base data array. The feature elements of the micro-topographic water potential dynamic map matrix are used as the spatial topological prior structure, combined with meteorological forcing time series vectors, to drive the spatial attention mechanism to perform neighborhood feature aggregation. The computational logic for the spatially dependent features of water transport is as follows: In the formula, This represents the spatial dependence of water transport at topological node p at time t; Let A represent the set of adjacent connected nodes of node p; u and A represent the sets of nodes connected to each other. The neighbor node index variable in the data; This represents the spatial attention weight vector, which is determined by randomly initializing it using a standard normal distribution and dynamically updating it through gradient descent during backpropagation of the network layers. and These represent the meteorological forcing time series feature vectors of the core node and its neighboring nodes, respectively. and Represents the edge feature elements of the input micro-topographic water potential dynamic map matrix; Indicates a multimodal feature concatenation operator; This represents the linear dimension mapping transformation matrix. Subsequently, a time-dimensional self-attention encoding operation is performed on the spatially dependent features of water transport: In the formula, This represents the generated high-order spatiotemporal water demand feature expression; The time window length limit for time-series self-attention observations is indicated; j and m represent the recursive index parameters for historical time steps. , and Let represent the learnable hidden layer tensors in the self-attention network's computational units. The high-order spatiotemporal water demand feature representation is mapped to the predicted water demand flux variable through a feedforward fully connected layer. A farmland water flux balance equation is constructed that couples precipitation, evapotranspiration, deep infiltration, and soil water storage variables; higher-order spatiotemporal water demand characteristics are mapped to predicted water demand flux variables. In the physical mechanism of farmland water balance, the predicted water demand flux variable is equivalent to the actual evapotranspiration water consumption variable output by the system to the outside world. Substituting the predicted water demand flux variable into the farmland water flux balance equation generates the balance residual loss term: In the formula, This represents the generated balanced residual loss term; This represents the predicted water demand flux variable obtained from the mapping calculation, which is the equivalent evapotranspiration water consumption variable. This represents the precipitation input variable obtained from external observations; Indicates the consumption variable of deep leakage; This represents the water storage variable in shallow soil. After obtaining the equilibrium residual loss term, it is directly embedded into the subsequent objective optimization function of the network to perform physical property constraint setting.

[0036] Specifically, the process of configuring the lower bound of crop root water-seeking memory penetration as the deep infiltration flux cutoff boundary of the objective optimization function and outputting the dynamic prediction trajectory of water demand for dryland farmland is as follows: The deep infiltration flux solution domain is extracted from the farmland water flux balance equation, and the lower bound of crop root water-seeking memory penetration is introduced into the deep infiltration flux solution domain as the vertical water interaction limit depth; a water flux cutoff condition is established at the vertical water interaction limit depth to forcibly terminate invalid deep water infiltration calculation components exceeding the lower bound of crop root water-seeking memory penetration, generating the deep infiltration flux cutoff boundary; under the constraint of the deep infiltration flux cutoff boundary, the backpropagation of the objective optimization function of the spatiotemporal graph attention network is performed, updating the weight parameters of the spatiotemporal graph attention network until the objective optimization function converges, completing the adaptive optimization of network parameters; the spatiotemporal graph attention network with optimized parameters performs forward evolution extrapolation on the meteorological forced time series to generate the dynamic prediction trajectory of water demand for dryland farmland that characterizes the evolution of water demand within the future time window.

[0037] In this implementation scheme, an independent solution domain for deep infiltration flux is extracted from the farmland water flux balance equation. The lower bound of the crop root hydrotropism memory penetration, mapped from the previous coordinate system to the current node p, is introduced into the solution domain as the vertical water interaction limit depth. A water flux cutoff condition is established at this depth to forcibly terminate the calculation of deep water infiltration components. In the formula, This represents the cutoff boundary value for deep permeability flux after constrained conditions. The total number of cross-sectional space grids in the vertical dimension; f represents the derivation index variable of the vertical grid level; This represents the permeability coefficient of the reference medium in the f-th grid layer; This represents the independent water flux calculation component of node p at time t in the f-th layer of the grid; Represents a binary step cutoff function; The lower bound of penetration at coordinates (x, y) Mapped to node The upper limit threshold, when the internal calculation depth Less than or equal to hydrotropism memory penetrates the lower bound The function outputs a positive constant 1 if the condition is met, otherwise the function outputs zero. This represents the physical absolute depth value corresponding to the f-th layer of the mesh; This represents the configured lower bound threshold for root system truncation. After generating the deep penetration flux truncation boundary value, it is substituted back into the overall network objective function for supervised feedback setting: In the formula, This represents the total algebraic value of the objective optimization function of the spatiotemporal graph attention network. This represents the total number of spatial topology nodes participating in the global training computation; and These represent the dynamic balance adjustment hyperparameters for the prediction error term and the physical conservation residual term, respectively. They are determined by adaptive scaling based on the mean gradient variance fluctuation within the network training batch. This represents the true water demand vector obtained based on historical records. Under the joint constraints of the objective optimization function and the deep infiltration flux cutoff boundary, the backpropagation operation of the network's underlying logic is performed to update the weight parameters until the loss value of the objective optimization function reaches the lower bound of the convergence threshold. Finally, the spatiotemporal graph attention network architecture, after parameter optimization, is invoked to perform forward inference evolution on the newly input meteorological forced time series, outputting a dynamic prediction trajectory matrix of water demand for dryland farmland that characterizes the dynamic evolution of node water demand within a specified future time period.

[0038] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0039] This invention is described with reference to flowchart illustrations and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0040] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0041] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0042] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0043] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting the evolution characteristics of water demand in dryland farmland based on machine learning, characterized in that, Includes the following steps: S1. Collect multi-source sensor signals from dry farmland and map them to construct a feature space of root-soil-air continuum; The characteristic space of the root-soil-gas continuum is composed of surface multispectral reflectance sequence, deep profile water content dynamic sequence and meteorological forcing time series mapping; S2. Extract the radiation attenuation gradient characteristics of the red and near-infrared bands in the surface multispectral reflectance sequence under continuous rainless stress, and construct the surface crust degradation index sequence; track the water consumption penetration characteristics of the deep profile water content dynamic sequence in the vertical direction, identify the water stress retardation interface, and extract the lower boundary of crop root hydrotropism memory penetration. S3. Discretize dry farmland into spatial topological nodes, analyze the spatial heterogeneous distribution differences of surface crust degradation index sequences between adjacent spatial topological nodes, transform the spatial heterogeneous distribution differences into surface water lateral conduction impedance factors, integrate the surface water lateral conduction impedance factors into the spatial distance attributes of spatial topological nodes, and construct a micro-topography water potential dynamic map matrix. S4. Input the micro-topographic water potential dynamic map matrix and meteorological forcing time series into the spatiotemporal map attention network to perform spatiotemporal feature aggregation operation. Embed the balance residual loss term in the objective optimization function into the spatiotemporal map attention network. Configure the lower bound of crop root water-seeking memory penetration as the deep infiltration flux cutoff boundary of the objective optimization function. The specific process of outputting the dynamic prediction trajectory of water demand for dryland farmland is as follows: The deep infiltration flux solution domain is extracted from the farmland water flux balance equation, and the lower bound of the crop root water-seeking memory penetration is introduced into the deep infiltration flux solution domain as the vertical water interaction limit depth. Establish water flux cutoff conditions at the vertical water interaction limit depth, forcibly terminate invalid deep water infiltration calculation components that exceed the lower limit of crop root hydrotropism memory penetration, and generate deep infiltration flux cutoff boundaries. Under the constraint of deep penetration flux cutoff boundary, backpropagation of the spatiotemporal graph attention network objective optimization function is performed to update the weight parameters of the spatiotemporal graph attention network until the objective optimization function converges, thus completing the adaptive optimization of network parameters. By performing forward evolution extrapolation on meteorological forced time series using a spatiotemporal graph attention network with optimized parameters, dynamic prediction trajectories of water demand for dryland farmland that characterize the evolution of water demand within future time windows are generated.

2. The method for predicting the water demand evolution characteristics of dryland farmland based on machine learning according to claim 1, characterized in that: The specific process of collecting multi-source sensor signals from dryland farmland and mapping them to construct a root-soil-air continuum feature space is as follows: Acquire multispectral remote sensing images from UAVs, signals from vertical soil moisture sensor networks, and meteorological observation data; Radiometric calibration and atmospheric correction are performed on UAV multispectral remote sensing images, and temporal pixel tensors are extracted to generate surface multispectral reflectance sequences. Depth interpolation reconstruction and timestamp alignment mapping are performed on vertical soil moisture sensor network signals to generate a dynamic sequence of deep profile moisture content. Multidimensional forcing factor feature encoding is performed on meteorological observation data to generate meteorological forcing time series; root-soil-gas continuum feature space is constructed by combining surface multispectral reflectance sequence, deep profile water content dynamic sequence and meteorological forcing time series mapping.

3. The method for predicting the water demand evolution characteristics of dryland farmland based on machine learning according to claim 1, characterized in that: The specific process for extracting the radiation attenuation gradient characteristics of the red and near-infrared bands in the surface multispectral reflectance sequence under continuous rainless stress and constructing the surface crust degradation index sequence is as follows: Based on the meteorological forcing time series, the continuous rainless stress time window is determined, and the reflectance time-varying tensor of the red band and near-infrared band in the surface multispectral reflectance sequence is extracted within the continuous rainless stress time window; Nonlinear divergence feature extraction was performed on the time-varying reflectance tensor to obtain the radiation attenuation gradient features of surface soil as drought stress intensifies; The radiation attenuation gradient features are input into a pre-defined surface structure degradation response function for mapping, and the surface soil reflectance attenuation rate parameter is output. By tracking the cumulative evolution trend of the surface soil reflectance attenuation rate parameter over time, a surface crust degradation index sequence is generated.

4. The method for predicting the water demand evolution characteristics of dryland farmland based on machine learning according to claim 1, characterized in that: The specific process of tracking the vertical water consumption penetration characteristics of the dynamic sequence of deep profile moisture content, identifying the water stress retardation interface, and extracting the lower bound of crop root hydrotaxis memory penetration is as follows: The spatial partial derivative analysis of the dynamic sequence of water content in deep profiles is performed in the vertical depth dimension to obtain the vertical gradient distribution matrix of water content. Extract the spatial critical depth where the water consumption rate decreases abruptly over time from the vertical gradient distribution matrix of water content, and mark the spatial critical depth as the water stress retardation interface. Track the vertical downward migration evolution trajectory of the water stress barrier interface during historical drought stress cycles, and extract the maximum penetration depth extreme value in the vertical downward migration evolution trajectory. The maximum penetration depth extreme value is transformed into the spatial limit of root morphology adaptive downward extension, generating the lower bound of crop root hydrotropism memory penetration.

5. The method for predicting the water demand evolution characteristics of dryland farmland based on machine learning according to claim 1, characterized in that: The specific process of discretizing dryland farmland into spatial topological nodes and analyzing the spatial heterogeneous distribution differences of surface crust degradation index sequences between adjacent spatial topological nodes is as follows: Dryland farmland is mapped into a two-dimensional grid space. The coordinate system features of the grid cell center in the two-dimensional grid space are extracted to construct a set of spatial topological nodes. Perform spatial coordinate alignment mapping between the surface crust degradation index sequence and the spatial topological node set to assign dynamic degradation attributes to the spatial topological nodes; Construct an adjacency-connected network structure for a set of spatial topological nodes, and extract the spatial feature gradient tensor of the surface crust degradation index sequence between adjacent spatial topological nodes under the constraints of the adjacency-connected network structure. The nonlinear deviation of the spatial feature gradient tensor in the evolution direction is analyzed to generate spatial heterogeneous distribution differences that characterize the boundary of damaged micro-topography in arid land.

6. The method for predicting the water demand evolution characteristics of dryland farmland based on machine learning according to claim 1 or 5, characterized in that: The specific process of transforming spatial heterogeneous distribution differences into surface water lateral conduction impedance factors, and then integrating these factors into the spatial distance attributes of spatial topological nodes to construct a dynamic micro-topographic water potential map matrix is ​​as follows: A soil moisture movement hydraulic resistance response function is introduced to map the spatial heterogeneous distribution differences into a surface moisture lateral conduction impedance factor that restricts the lateral transport of surface moisture. Based on the spatial topology node set, extract the Euclidean spatial coordinate distribution features between adjacent nodes to generate a spatial topology node spatial distance attribute matrix; Spatial penalty attenuation mapping is performed on the spatial distance attribute matrix of spatial topology nodes. The surface water transverse conduction impedance factor is injected into the spatial penalty attenuation mapping mechanism to generate a dynamic spatial weight connection matrix. By combining the dynamic spatial weighted connection matrix with the micro-topographic water flow convergence topology extracted from the digital elevation model, cross-modal feature splicing is performed to construct a dynamic micro-topographic water potential map matrix.

7. The method for predicting the water demand evolution characteristics of dryland farmland based on machine learning according to claim 1, characterized in that: The specific process of inputting the micro-topographic water potential dynamic map matrix and the meteorological forcing time series into the spatiotemporal map attention network to perform spatiotemporal feature aggregation operations, and embedding the balanced residual loss term in the objective optimization function into the spatiotemporal map attention network is as follows: The micro-topographic water potential dynamic map matrix is ​​used as the spatial topological prior structure input to the spatiotemporal graph attention network to drive the spatial attention mechanism to perform neighborhood feature aggregation on meteorological forced time series and extract spatial dependence features of water transport. Perform temporal self-attention encoding on the spatial dependence features of water transport to capture the cumulative effect of temporal drought and generate a high-order spatiotemporal water demand feature representation; Construct a farmland water flux balance equation that couples precipitation, evapotranspiration, deep infiltration, and soil water storage variables; The high-order spatiotemporal water demand characteristics are mapped to predicted water demand flux variables. The predicted water demand flux variables are substituted into the farmland water flux balance equation to generate a balance residual loss term. The balance residual loss term is then embedded into the objective optimization function of the spatiotemporal graph attention network.

Images