Soil water-saving irrigation management method and system based on multi-source data

Abstract

The application discloses a soil water-saving irrigation management method and system based on multi-source data, which comprises the following steps: collecting multi-source irrigation state data, performing pretreatment, and generating a unified irrigation data set; mapping a unified spatial coordinate system, dividing irrigation grid units, constructing an adjacency relationship, and completing interpolation compensation processing; inputting an improved MTGNN model, performing double-channel interaction and dilated convolution, and outputting a water shortage priority matrix; inputting the improved MTGNN model, performing double-channel interaction and dilated convolution, and outputting a water shortage priority matrix; according to the priority of the grid unit, a time-sharing water supply window is constructed, and a peak-shaving irrigation scheduling sequence is generated; irrigation scheduling is performed and real-time monitoring is performed, water supplement deviation is calculated, and data is written back to update irrigation decision-making. Through multi-source data fusion and an improved MTGNN model, the application realizes accurate identification of water shortage trends in irrigation areas and differentiated water-saving irrigation scheduling control.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture and intelligent irrigation control technology, and in particular to a soil water-saving irrigation management method and system based on multi-source data. Background Technology

[0002] With the development of agricultural water-saving irrigation technology, Internet of Things (IoT) monitoring technology, and automatic control technology, farmland irrigation is gradually shifting from manual, timed irrigation to automated irrigation based on soil moisture, weather conditions, and crop growth status. Existing irrigation management methods typically collect relevant data through soil moisture sensors, meteorological monitoring equipment, and flow meters, and then control water pumps and valves according to preset moisture content thresholds or fixed irrigation cycles to achieve water supply management for different planting areas.

[0003] However, various monitoring data within farmland differ in sampling time, spatial location, and data scale, and some monitoring nodes are prone to missing or abnormal measurements. Existing methods typically process data from different sources separately, failing to integrate analysis with plot spatial location, water migration relationships with adjacent areas, and historical irrigation responses, making it difficult to accurately identify localized water shortages. Fixed thresholds are ill-suited to adapting to continuous changes in crop growth stages, evapotranspiration intensity, and meteorological conditions, easily leading to delayed water shortage assessments or the need for repeated irrigation.

[0004] Current irrigation scheduling often involves simultaneously activating multiple irrigation zones according to a preset sequence, lacking joint constraints on target water replenishment volume, available water supply, and remaining pipeline flow. When multiple areas are simultaneously experiencing water shortages, it can easily lead to insufficient pipeline flow, uneven water supply at the end points, and conflicts in irrigation tasks. Furthermore, after irrigation is completed, the water supply flow, valve opening, and irrigation time are usually not adjusted based on the actual water replenishment effect, making it difficult to form a continuously updated closed-loop irrigation decision-making process, thus affecting water resource utilization efficiency and irrigation management accuracy.

[0005] Therefore, how to provide soil water-saving irrigation management methods and systems based on multi-source data is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a soil water-saving irrigation management method and system based on multi-source data. This invention comprehensively utilizes multi-source data fusion, spatial grid construction, temporal feature analysis, and improved MTGNN prediction technology to accurately predict the water shortage status of farmland. It also combines the target water replenishment volume, available water supply, and remaining pipeline flow to achieve zoned staggered irrigation scheduling. At the same time, it dynamically updates irrigation decisions based on irrigation feedback. It has the advantages of accurate water shortage identification, reasonable irrigation scheduling, high water resource utilization rate, and high degree of intelligent irrigation management.

[0007] The soil water-saving irrigation management method based on multi-source data according to embodiments of the present invention includes:

[0008] Collect multi-source irrigation status data, preprocess the multi-source irrigation status data, and generate an irrigation dataset;

[0009] The irrigation dataset is mapped to a unified plot spatial coordinate system, irrigation grid units are divided according to monitoring locations, an adjacency matrix is ​​constructed, interpolation compensation is performed, and a spatially consistent multi-source grid data field is formed.

[0010] For multi-source grid data fields, the rate of change of soil moisture content, evapotranspiration potential, crop water requirement coefficient and historical water replenishment response are calculated to construct a time-series water shortage characteristic sequence;

[0011] The temporal water shortage feature sequence and adjacency matrix are input into the improved MTGNN model. The improved MTGNN model connects a dual-channel spatial interaction layer between graph convolution and gated temporal convolution. The node embedding is updated in parallel based on the dynamic crop-dependent adjacency matrix and the dynamic climate-dependent adjacency matrix, respectively. A residual dilated convolution module is inserted before each level of gated temporal convolution to expand the temporal receptive field. A prediction confidence quantization head is superimposed at the output end to output the water shortage priority matrix.

[0012] The irrigation grid units are sorted according to the water shortage priority matrix. The time-sharing water supply window is constructed by combining the target water replenishment, water supply availability and pipeline remaining flow. The continuous irrigation area is merged according to the water migration relationship of adjacent grids. The irrigation time period, water supply flow and valve opening are allocated in sequence. Irrigation tasks that exceed the water volume or flow constraints are reordered in sequence to generate a zoned staggered irrigation sequence.

[0013] The system executes a zoned staggered irrigation sequence, collects soil moisture content, water supply flow rate, and valve opening in real time for each irrigation grid unit, calculates the deviation between the actual water replenishment and the target water replenishment, and adjusts the water supply flow rate, valve opening, or irrigation period. The results are then written back to the multi-source irrigation status data to update irrigation decisions.

[0014] Optionally, the multi-source irrigation status data includes soil moisture content, soil temperature, electrical conductivity, crop growth indicators, meteorological elements, valve operating status, water supply flow rate, and historical irrigation records.

[0015] Optionally, the preprocessing of multi-source irrigation status data to generate an irrigation data set includes synchronizing the multi-source irrigation status data according to the collection time, detecting and removing missing, duplicate, and abnormal data, uniformly converting the collection formats of different data sources, completing data association based on monitoring nodes, and integrating the processed multi-source irrigation status data according to time order and monitoring location to generate an irrigation data set.

[0016] Optionally, the formation of a spatially consistent multi-source grid data field includes:

[0017] Read the latitude, longitude, elevation and measurement values ​​of each monitoring node in the irrigation data set, use the same projection coordinate system for coordinate transformation, and add the topographic slope angle and near-surface hydraulic gradient coefficient to each monitoring node;

[0018] Hierarchical clustering is performed based on the spatial density of monitoring nodes in a unified coordinate system, slope aspect consistency, and hydraulic gradient differences. The clustering results automatically delineate the grid boundaries, resulting in irrigation grid cells with side lengths that adapt to terrain complexity. The measured values ​​of each monitoring node are assigned to the center of their respective grid.

[0019] Based on the horizontal distance, slope aspect connectivity, and hydraulic gradient direction between the centers of irrigation grid units, the direct water migration probability between any two grids is calculated in real time. If the distance threshold, slope aspect, and hydraulic gradient are all satisfied at the same time, they are recorded as adjacency relationships, and an adjacency matrix that is dynamically updated with changes in surface conditions is constructed.

[0020] For irrigation grid cells with missing measurements, a set of spatially adjacent irrigation grid cells with complete data from the two most recent sampling periods is selected based on the adjacency matrix. Bidirectional spatiotemporal weighted interpolation is performed by assigning weights according to spatial proximity and temporal similarity to calculate the missing measurements and write them into the corresponding irrigation grid cells. At the same time, the interpolation confidence flag is recorded.

[0021] All irrigation grid cells are summarized according to a unified coordinate index order. The center coordinates, boundary range, adjacency matrix entries, monitoring values ​​and interpolation confidence indicators of each grid are recorded. The resulting spatially consistent multi-source grid data field is then generated and output.

[0022] Optionally, the construction of the time-series water shortage feature sequence includes:

[0023] Soil moisture content data is read from the multi-source grid data field according to the irrigation grid cell number and time index. The difference between the soil moisture content at the current sampling time and the soil moisture content at the previous sampling time is calculated. The difference is amortized according to the time interval between two adjacent samplings to obtain the rate of change of soil moisture content. A sequence of the rate of change of soil moisture content is generated in chronological order.

[0024] Meteorological elements corresponding to each irrigation grid unit at the time are read from the multi-source grid data field. Temperature, wind speed, solar radiation and air humidity are processed at the same scale. The increase in temperature, wind speed and solar radiation are taken as the evapotranspiration enhancement term, and the increase in air humidity is taken as the evapotranspiration inhibition term. The evapotranspiration potential is obtained by weighting and summarizing according to the preset weights and generating an evapotranspiration potential sequence in chronological order.

[0025] The crop growth index corresponding to each irrigation grid unit is read from the multi-source grid data field. The crop growth index is matched with the preset crop growth stage range to determine the current growth stage of each irrigation grid unit. The crop water demand coefficient corresponding to the growth stage is read and the crop water demand coefficient sequence is generated in time order.

[0026] Historical irrigation records of each irrigation grid unit are read from the multi-source grid data field. Soil moisture content before irrigation, soil moisture content after irrigation and corresponding water supply are extracted. The increase of soil moisture content after irrigation relative to soil moisture content before irrigation is converted according to the water supply to obtain the single water replenishment response. The most recent single water replenishment response is smoothed to generate a historical water replenishment response sequence.

[0027] According to the irrigation grid unit number and time index, the water content change rate sequence, evapotranspiration potential sequence, crop water requirement coefficient sequence and historical water replenishment response sequence are aligned. The four types of features of the same irrigation grid unit at the same sampling time are arranged in a fixed order to construct a single-time water shortage feature record. The single-time water shortage feature records of consecutive sampling times are spliced ​​together to generate a time-series water shortage feature sequence.

[0028] Optionally, the output water shortage priority matrix includes:

[0029] Read the time series water shortage feature sequence and adjacency matrix, arrange them according to irrigation grid cell number, continuous sampling time and water shortage feature type, write the water content change rate, evapotranspiration potential, crop water requirement coefficient and historical water replenishment response of the same irrigation grid cell into the same time slice, and generate the model input data structure;

[0030] The model input data structure and adjacency matrix are input into the graph convolutional layer of the improved MTGNN model. Based on the water migration channels marked in the adjacency matrix, the water shortage characteristics of adjacent irrigation grid cells are spatially propagated and aggregated to generate spatial propagation features.

[0031] A dual-channel spatial interaction layer is set between the graph convolutional layer and the gated temporal convolutional layer. Taking spatial propagation characteristics as input, the crop water demand coefficient, crop growth index, meteorological elements and evapotranspiration potential are read respectively to construct a dynamic crop dependency adjacency matrix and a dynamic climate dependency adjacency matrix.

[0032] Based on the dynamic crop dependency adjacency matrix, node embedding and updating are performed on irrigation grid units with similar growth stages, similar crop water demand coefficients, and similar historical water replenishment response to generate crop dependency channel features; based on the dynamic climate dependency adjacency matrix, node embedding and updating are performed on irrigation grid units with similar meteorological elements and similar evapotranspiration potential change trends to generate climate dependency channel features.

[0033] The crop-dependent channel features and climate-dependent channel features are aligned according to the irrigation grid cell number and time index. Channel splicing, feature filtering and weight mapping are performed to generate dual-channel spatial interaction features, which are then input into the residual dilated convolution module set before each level of gated temporal convolution.

[0034] The residual dilated convolution module is used to extract the water shortage trend of the dual-channel spatial interaction features at different time spans. The extracted results are then superimposed with the original dual-channel spatial interaction features by residual superposition and input into the gated temporal convolutional layer to generate water shortage temporal evolution features.

[0035] The water shortage time series evolution features are input to the output terminal. The water shortage score of each irrigation grid unit is generated through the water shortage score branch. The corresponding confidence score is generated through the prediction confidence quantification head. The water shortage score, confidence score and irrigation grid unit number are arranged from high to low according to the degree of water shortage to generate a water shortage priority matrix.

[0036] Optionally, generating the partitioned staggered irrigation sequence includes:

[0037] Read the water shortage priority matrix, extract the number of each irrigation grid unit, water shortage score and prediction confidence, use the water shortage score and prediction confidence as the sorting criteria, sort each irrigation grid unit from high to low according to the urgency of water shortage, and generate an irrigation grid unit sorting queue.

[0038] According to the irrigation grid unit sorting queue, the current soil moisture content, target soil moisture content and historical water replenishment response of each irrigation grid unit are read in sequence. The difference between the target soil moisture content and the current soil moisture content is taken as the water replenishment gap, and the target water replenishment amount of each irrigation grid unit is calculated based on the historical water replenishment response.

[0039] Read the available water supply and remaining flow in the pipeline network within the current scheduling cycle, divide the scheduling cycle into multiple continuous time-sharing water supply windows, and set the upper limit of allocable water volume and the upper limit of allocable flow for each time-sharing water supply window. Write irrigation tasks sequentially according to the irrigation grid unit sorting queue, so that the total target water replenishment volume written to the same time-sharing water supply window does not exceed the upper limit of allocable water volume and the total water supply flow does not exceed the upper limit of allocable flow.

[0040] Read the water migration channels marked in the adjacency matrix, merge the irrigation grid units that are written to the same time-sharing water supply window and are spatially adjacent, have similar target water replenishment amounts and similar water shortage levels to form a continuous irrigation area, and allocate irrigation time periods, water supply flow rates and valve openings according to the target water replenishment amounts of each irrigation grid unit in the continuous irrigation area.

[0041] Constraint verification is performed on the continuous irrigation areas within each time-sharing water supply window. Irrigation tasks exceeding the available water supply, remaining flow in the pipeline network, or window time capacity are postponed to the next time-sharing water supply window. Irrigation time periods, water supply flow, and valve opening are reallocated, and a staggered irrigation sequence is generated by zone number, execution time, duration, water supply flow, and valve opening.

[0042] Optionally, the step of calculating the deviation between the actual water replenishment volume and the target water replenishment volume and adjusting the water supply flow rate, valve opening, or irrigation period includes:

[0043] According to the execution time set in the staggered irrigation sequence, the corresponding continuous irrigation areas are started in sequence, the corresponding valves are opened to the preset opening degree, and water is continuously supplied according to the allocated water supply flow. After the irrigation is completed, the corresponding valves are closed to complete the irrigation execution of each continuous irrigation area.

[0044] During the irrigation process, the soil moisture content, water flow rate and valve opening of each irrigation grid unit are collected in real time according to the preset sampling cycle. The actual water supply during the irrigation period is obtained by accumulating the water flow rate in each sampling cycle, and the actual water replenishment of each irrigation grid unit is determined by combining the changes in soil moisture content before and after irrigation.

[0045] Read the target water replenishment amount corresponding to each irrigation grid unit, compare the actual water replenishment amount with the target water replenishment amount one by one, obtain the water replenishment deviation of each irrigation grid unit, and generate the water replenishment deviation result based on the magnitude and direction of the water replenishment deviation.

[0046] Based on the water replenishment deviation results, adjust the water supply flow, valve opening, or irrigation period of the corresponding irrigation grid unit. Specifically, for irrigation grid units where the actual water replenishment is lower than the target water replenishment, increase the water supply flow, increase the valve opening, or extend the irrigation period. For irrigation grid units where the actual water replenishment is higher than the target water replenishment, decrease the water supply flow, decrease the valve opening, or shorten the irrigation period, and generate the corrected irrigation execution result.

[0047] The soil moisture content, water supply flow, valve opening, actual water replenishment, water replenishment deviation, and irrigation period corresponding to the corrected irrigation execution results are written into the multi-source irrigation status data. The irrigation data set is updated, and the operating status of the corresponding irrigation grid unit is regenerated, providing updated input data for water shortage analysis, priority ranking, and irrigation decision-making in the next scheduling cycle.

[0048] According to an embodiment of the present invention, a soil water-saving irrigation management system based on multi-source data includes:

[0049] The data acquisition and preprocessing module is used to collect multi-source irrigation status data and perform preprocessing to generate an irrigation dataset;

[0050] The spatial grid construction module is used to complete unified coordinate mapping, irrigation grid cell division, adjacency matrix construction and interpolation compensation processing based on irrigation data sets, and generate multi-source grid data fields.

[0051] The water shortage feature construction module is used to extract soil moisture content change rate, evapotranspiration potential, crop water requirement coefficient and historical water replenishment response based on multi-source grid data field, and construct a time-series water shortage feature sequence.

[0052] The water shortage prediction and analysis module is used to input the time-series water shortage feature sequence and adjacency matrix into the improved MTGNN model to generate the water shortage priority matrix for each irrigation grid unit.

[0053] The zoned scheduling decision module is used to construct time-sharing water supply windows based on the water shortage priority matrix, combined with the target water replenishment volume, water source availability and pipeline network remaining flow, to perform staggered scheduling for continuous irrigation areas and generate zoned staggered irrigation sequences.

[0054] The irrigation feedback update module is used to execute the regional staggered irrigation sequence, collect the irrigation execution status, calculate the deviation between the actual water replenishment and the target water replenishment, adaptively adjust the irrigation parameters, and update the multi-source irrigation status data and irrigation decision results.

[0055] The beneficial effects of this invention are:

[0056] This invention maps multi-source irrigation status data to a unified plot spatial coordinate system, divides irrigation grid units according to monitoring locations, and constructs an adjacency matrix by combining the distance between grids, slope aspect, and hydraulic gradient. It performs spatiotemporal interpolation compensation for missing monitoring data, enabling soil moisture, meteorological conditions, crop growth status, and irrigation operation data to be correlated and analyzed at a unified spatial scale. This improves the consistency and completeness of multi-source data and reduces the impact of missing data, differences in sampling locations, and independent processing on water shortage judgment.

[0057] This invention constructs a time-series water shortage feature sequence by extracting soil moisture content change rate, evapotranspiration potential, crop water requirement coefficient, and historical water replenishment response. It uses an improved MTGNN model to simultaneously analyze the spatial water migration relationship, crop water requirement dependence, and climate change dependence among irrigation grid units. It combines residual dilatation convolution to expand the time analysis range of water shortage state, and evaluates the reliability of prediction results through a prediction confidence quantification head, thereby improving the accuracy of water shortage identification and prioritization under different crop growth stages and complex meteorological conditions.

[0058] This invention constructs a time-sharing water supply window based on a water shortage priority matrix, combining the target water replenishment volume, available water source, and remaining pipeline flow. It also merges continuous irrigation areas according to the water migration relationship between adjacent grids, and allocates irrigation time periods, water supply flow, and valve opening in different zones to reduce flow conflicts and insufficient end-point water supply caused by simultaneous water supply to multiple irrigation areas. At the same time, it dynamically adjusts irrigation parameters and updates irrigation decisions based on the deviation between the actual water replenishment volume and the target water replenishment volume, realizing closed-loop management of irrigation execution and feedback correction, thereby improving irrigation uniformity, water resource utilization efficiency, and water-saving irrigation management level. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is a flowchart of the soil water-saving irrigation management method based on multi-source data proposed in this invention;

[0061] Figure 2 This is a graph showing the sequence variation of water shortage characteristics in the soil water-saving irrigation management method based on multi-source data proposed in this invention.

[0062] Figure 3 This is a ranking diagram of water shortage scores and prediction confidence of irrigation grid units in the soil water-saving irrigation management method based on multi-source data proposed in this invention.

[0063] Figure 4 This is a schematic diagram of the soil water-saving irrigation management system based on multi-source data proposed in this invention. Detailed Implementation

[0064] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0065] refer to Figures 1-3 Soil water-saving irrigation management methods based on multi-source data include:

[0066] Collect multi-source irrigation status data, preprocess the multi-source irrigation status data, and generate an irrigation dataset;

[0067] The irrigation dataset is mapped to a unified plot spatial coordinate system, irrigation grid units are divided according to monitoring locations, an adjacency matrix is ​​constructed, interpolation compensation is performed, and a spatially consistent multi-source grid data field is formed.

[0068] For multi-source grid data fields, the rate of change of soil moisture content, evapotranspiration potential, crop water requirement coefficient and historical water replenishment response are calculated to construct a time-series water shortage characteristic sequence;

[0069] The temporal water shortage feature sequence and adjacency matrix are input into the improved MTGNN model. The improved MTGNN model connects a dual-channel spatial interaction layer between graph convolution and gated temporal convolution. The node embedding is updated in parallel based on the dynamic crop-dependent adjacency matrix and the dynamic climate-dependent adjacency matrix, respectively. A residual dilated convolution module is inserted before each level of gated temporal convolution to expand the temporal receptive field. A prediction confidence quantization head is superimposed at the output end to output the water shortage priority matrix.

[0070] The irrigation grid units are sorted according to the water shortage priority matrix. The time-sharing water supply window is constructed by combining the target water replenishment, water supply availability and pipeline remaining flow. The continuous irrigation area is merged according to the water migration relationship of adjacent grids. The irrigation time period, water supply flow and valve opening are allocated in sequence. Irrigation tasks that exceed the water volume or flow constraints are reordered in sequence to generate a zoned staggered irrigation sequence.

[0071] The system executes a zoned staggered irrigation sequence, collects soil moisture content, water supply flow rate, and valve opening in real time for each irrigation grid unit, calculates the deviation between the actual water replenishment and the target water replenishment, and adjusts the water supply flow rate, valve opening, or irrigation period. The results are then written back to the multi-source irrigation status data to update irrigation decisions.

[0072] In this embodiment, the multi-source irrigation status data includes soil moisture content, soil temperature, electrical conductivity, crop growth indicators, meteorological elements, valve operating status, water supply flow rate, and historical irrigation record data.

[0073] In this embodiment, the preprocessing of multi-source irrigation status data to generate an irrigation data set includes synchronizing the multi-source irrigation status data according to the collection time, detecting and removing missing, duplicate, and abnormal data, uniformly converting the collection formats of different data sources, completing data association based on monitoring nodes, and integrating the processed multi-source irrigation status data according to time order and monitoring location to generate an irrigation data set.

[0074] In this embodiment, forming a spatially consistent multi-source grid data field includes:

[0075] Read the latitude, longitude, elevation and measurement values ​​of each monitoring node in the irrigation data set, use the same projection coordinate system for coordinate transformation, and add the topographic slope angle and near-surface hydraulic gradient coefficient to each monitoring node;

[0076] Hierarchical clustering is performed based on the spatial density of monitoring nodes in a unified coordinate system, slope aspect consistency, and hydraulic gradient differences. The clustering results automatically define the grid boundaries, resulting in irrigation grid cells with side lengths adaptively varying with terrain complexity. The measurements of each monitoring node are assigned to the center of their respective grid. Specifically, the irrigation grid cells with side lengths adaptively varying with terrain complexity are as follows:

[0077] The spatial coordinates, slope aspect angles, and hydraulic gradient parameters of all monitoring nodes in a unified coordinate system are read. Hierarchical clustering is performed based on the spatial distance, slope aspect angle difference, and hydraulic gradient difference between monitoring nodes, and the clustering results are used as the initial irrigation area. The number of monitoring nodes, the range of slope aspect angle variation, and the range of hydraulic gradient variation within each initial irrigation area are counted. Initial irrigation areas with no fewer than 20 monitoring nodes and a slope aspect angle variation range greater than 25° or a hydraulic gradient variation range greater than 0.1 are identified as complex terrain areas. Initial irrigation areas with fewer than 20 monitoring nodes and a small slope aspect angle variation range are identified as complex terrain areas. An initial irrigation area with an initial angle of 25° and a hydraulic gradient variation range of no more than 0.1 is defined as a flat terrain area. For complex terrain areas, the grid is re-divided with a side length of 5 meters for irrigation grid units. For flat terrain areas, the grid is re-divided with a side length of 15 meters for irrigation grid units. When the number of monitoring nodes in adjacent irrigation grid units differs by more than 5 after re-division, the irrigation grid units with more monitoring nodes are re-divided until the difference in the number of monitoring nodes in adjacent irrigation grid units does not exceed 5, thus generating irrigation grid units with side lengths that adaptively change with terrain complexity.

[0078] Based on the horizontal distance, slope aspect connectivity, and hydraulic gradient direction between the centers of irrigation grid cells, the direct water migration probability between any two grid cells is calculated in real time. If a cell simultaneously meets the distance threshold, has the same slope aspect, and similar hydraulic gradient, it is recorded as an adjacency relationship. An adjacency matrix is ​​constructed and dynamically updated according to changes in surface conditions. Specifically, the real-time calculation of the direct water migration probability between any two grid cells involves:

[0079] Read the x-coordinate of the grid center, y-coordinate of the grid center, slope angle, hydraulic gradient value, and hydraulic gradient direction of each irrigation grid cell. Arrange the above parameters corresponding to each irrigation grid cell in a fixed order to generate a grid spatial attribute record sequence. Combine all grid spatial attribute record sequences in pairs according to the irrigation grid cell number order. Calculate the horizontal distance between any two irrigation grid cells, the slope angle difference between any two irrigation grid cells, and the hydraulic gradient difference between any two irrigation grid cells, and determine whether the hydraulic gradient direction points from the high gradient grid to the low gradient grid. Grid pairs with a horizontal distance not exceeding 15 meters, a slope angle difference not exceeding 20°, a hydraulic gradient difference not exceeding 0.08, and consistent gradient directions are identified as direct water migration grid pairs. Write 1 in the corresponding position of the direct water migration grid pair in the adjacency matrix and 0 in the other positions to generate the adjacency matrix between irrigation grid cells.

[0080] For irrigation grid cells with missing measurements, a set of spatially adjacent irrigation grid cells with complete data from the two most recent sampling periods is selected based on the adjacency matrix. Bidirectional spatiotemporal weighted interpolation is then performed, assigning weights based on spatial proximity and temporal similarity. The missing measurements are calculated and written into the corresponding irrigation grid cell, while interpolation confidence indicators are recorded. Specifically, the bidirectional spatiotemporal weighted interpolation, assigning weights based on spatial proximity and temporal similarity, is as follows:

[0081] Read the monitoring data of all adjacent irrigation grid cells in the adjacency matrix of the irrigation grid cell to be compensated. Select 3 to 5 irrigation grid cells with the closest center distance and complete monitoring values ​​in the two most recent sampling periods as interpolation reference grids. Calculate the center distance between the irrigation grid cell to be compensated and each interpolation reference grid. Sort them in ascending order of center distance, assign the highest spatial weight to the interpolation reference grid with the closest center distance, and decrease the spatial weight of the remaining interpolation reference grids in ascending order of center distance, so that the sum of all spatial weights is 1. Read the monitoring values ​​of each interpolation reference grid in the current sampling period and the previous sampling period, calculate the change in the monitoring values ​​of each interpolation reference grid in the two most recent sampling periods, and compare the differences between the changes. The interpolation reference grid with the smallest change is assigned the highest time weight, and the time weights of the remaining interpolation reference grids are decreased in order of increasing change, so that the sum of all time weights is 1. The spatial weight and time weight of each interpolation reference grid are multiplied to obtain the comprehensive weight. Based on the comprehensive weight, the monitoring values ​​of the current sampling period and the monitoring values ​​of the previous sampling period of each interpolation reference grid are weighted and accumulated to obtain the interpolation result at the current time and the interpolation result at the previous time. The average of the interpolation result at the current time and the interpolation result at the previous time is taken as the missing measurement value of the irrigation grid unit to be compensated and written into the corresponding irrigation grid unit. At the same time, the interpolation confidence flag is recorded according to the proportion of the interpolation reference grid with the largest comprehensive weight to the total comprehensive weight.

[0082] All irrigation grid cells are summarized according to a unified coordinate index order. The center coordinates, boundary range, adjacency matrix entries, monitoring values ​​and interpolation confidence indicators of each grid are recorded. The resulting spatially consistent multi-source grid data field is then generated and output.

[0083] In this embodiment, constructing the time-series water shortage feature sequence includes:

[0084] Soil moisture content data is read from the multi-source grid data field according to the irrigation grid cell number and time index. The difference between the soil moisture content at the current sampling time and the soil moisture content at the previous sampling time is calculated. The difference is amortized according to the time interval between two adjacent samplings to obtain the rate of change of soil moisture content. A sequence of the rate of change of soil moisture content is generated in chronological order.

[0085] Meteorological elements for each irrigation grid unit at the corresponding time point are read from a multi-source grid data field. Temperature, wind speed, solar radiation, and air humidity are processed at the same scale. Increased temperature, wind speed, and solar radiation are treated as evapotranspiration enhancement terms, while increased air humidity is treated as evapotranspiration inhibition terms. These are weighted and aggregated according to preset weights to obtain the evapotranspiration potential. An evapotranspiration potential sequence is generated in chronological order, where the weighted aggregation according to preset weights yields the evapotranspiration potential. Specifically:

[0086] Read the temperature, wind speed, solar radiation, and air humidity corresponding to the same irrigation grid unit at the current moment. Read the lower and upper limits of temperature, wind speed, solar radiation, and air humidity recorded in the control unit respectively. Divide the difference between the current temperature and the lower limit of temperature by the difference between the upper and lower limits of temperature to obtain the temperature scale value. Divide the difference between the current wind speed and the lower limit of wind speed by the difference between the upper and lower limits of wind speed to obtain the wind speed scale value. Divide the difference between the current solar radiation and the lower limit of solar radiation by the difference between the upper and lower limits of solar radiation to obtain the solar radiation scale value. Divide the difference between the current air humidity and the lower limit of air humidity by the difference between the upper and lower limits of air humidity to obtain the air humidity scale value. Adjust scale values ​​below 0 to 0 and scale values ​​above 1 to 1.

[0087] Multiply the temperature scale value by 0.3 to obtain the temperature enhancement value; multiply the wind speed scale value by 0.2 to obtain the wind speed enhancement value; multiply the solar radiation scale value by 0.35 to obtain the radiation enhancement value; multiply the air humidity scale value by 0.15 to obtain the humidity suppression value. Add the temperature enhancement value, wind speed enhancement value, and radiation enhancement value together, and then subtract the humidity suppression value to obtain the initial evapotranspiration potential. When the initial evapotranspiration potential is less than 0, the evapotranspiration potential is adjusted to 0; when the initial evapotranspiration potential is greater than 1, the evapotranspiration potential is adjusted to 1; otherwise, the initial evapotranspiration potential is directly used as the evapotranspiration potential of the current irrigation grid unit at the corresponding time.

[0088] The minimum temperature limit is 0℃ and the maximum temperature limit is 40℃; the minimum wind speed limit is 0 meters per second and the maximum wind speed limit is 10 meters per second; the minimum solar radiation limit is 0 watts per square meter and the maximum solar radiation limit is 1000 watts per square meter; and the minimum air humidity limit is 20% and the maximum air humidity limit is 100%.

[0089] Crop growth indicators corresponding to each irrigation grid unit are read from a multi-source grid data field. These indicators are then matched with preset crop growth stage ranges to determine the current growth stage of each irrigation grid unit. The corresponding crop water requirement coefficients for each growth stage are read, and a crop water requirement coefficient sequence is generated in chronological order. Specifically, matching crop growth indicators with preset crop growth stage ranges to determine the current growth stage of each irrigation grid unit involves:

[0090] Read the crop plant height, leaf area index, canopy coverage, and cumulative growth days since sowing for each irrigation grid unit at the current moment. Arrange the plant height, leaf area index, canopy coverage, and cumulative growth days in a fixed order to generate crop growth index records. Read the growth index ranges corresponding to the seedling stage, vegetative growth stage, flowering and fruiting stage, and maturity stage preset in the control unit, and compare the current crop growth index records with the index ranges of each growth stage.

[0091] When the cumulative growth days are 1 to 20 days after sowing, the plant height does not exceed 20 cm, the leaf area index does not exceed 1, and the canopy coverage does not exceed 30%, the corresponding irrigation grid unit is designated as the seedling stage; when the cumulative growth days are 21 to 50 days after sowing, the plant height is greater than 20 cm but not more than 60 cm, the leaf area index is greater than 1 but not more than 3, and the canopy coverage is greater than 30% but not more than 65%, the corresponding irrigation grid unit is designated as the vegetative growth stage; when the cumulative growth days are 51 to 90 days after sowing, the plant height is greater than 60 cm, the leaf area index is greater than 3 but not more than 5.5, and the canopy coverage is greater than 65% but not more than 90%, the corresponding irrigation grid unit is designated as the flowering and fruiting stage; when the cumulative growth days exceed 90 days, the leaf area index is less than 3, and the canopy coverage is less than 70%, the corresponding irrigation grid unit is designated as the maturity stage.

[0092] When crop growth indicators fall within the range of two adjacent growth stages, the number of indicators satisfied in each growth stage is counted separately, and the growth stage with more indicators satisfied is determined as the current growth stage; when two adjacent growth stages satisfy the same number of indicators, the growth stage to which the cumulative growth days belong is given priority, and the determination result is associated with the irrigation grid unit number and the current time.

[0093] Historical irrigation records for each irrigation grid unit are read from a multi-source grid data field. Soil moisture content before irrigation, soil moisture content after irrigation, and corresponding water supply are extracted for each irrigation. The increase in soil moisture content after irrigation relative to pre-irrigation soil moisture content is adjusted according to the water supply to obtain the single-time water replenishment response. The most recent single-time water replenishment response is smoothed to generate a historical water replenishment response sequence. Specifically, the adjustment of soil moisture content after irrigation relative to pre-irrigation soil moisture content is as follows:

[0094] Read the soil moisture content before irrigation, the soil moisture content after irrigation, and the corresponding water supply from the same irrigation record. Subtract the soil moisture content before irrigation from the soil moisture content after irrigation to obtain the soil moisture content recovery rate corresponding to this irrigation. When the soil moisture content recovery rate is less than 0, mark the irrigation record as an abnormal record and remove it from the conversion process. When the corresponding water supply is 0 or the water supply record is missing, mark the irrigation record as an invalid record and stop the calculation.

[0095] The soil moisture content recovery rate is converted to a unit water supply based on the corresponding water supply. Specifically, the soil moisture content recovery rate is divided by the corresponding water supply to obtain the soil moisture content recovery value caused by the unit water supply. When the area or grid side length of each irrigation grid unit is inconsistent, the corresponding water supply is first divided by the actual area of ​​the irrigation grid unit to obtain the water supply per unit area. Then, the soil moisture content recovery rate is divided by the water supply per unit area to obtain the single water replenishment response amount corresponding to the unit area and unit water supply.

[0096] The single water replenishment response quantities are sorted according to the irrigation end time. The five most recent effective single water replenishment response quantities are selected. The five most recent single water replenishment response quantities are assigned smoothing weights of 0.3, 0.25, 0.2, 0.15 and 0.1 respectively. Each single water replenishment response quantity is multiplied by its corresponding smoothing weight and then summed. The summed result is used as the historical water replenishment response quantity of the current irrigation grid unit.

[0097] According to the irrigation grid unit number and time index, the water content change rate sequence, evapotranspiration potential sequence, crop water requirement coefficient sequence and historical water replenishment response sequence are aligned. The four types of features of the same irrigation grid unit at the same sampling time are arranged in a fixed order to construct a single-time water shortage feature record. The single-time water shortage feature records of consecutive sampling times are spliced ​​together to generate a time-series water shortage feature sequence.

[0098] In this embodiment, the output water shortage priority matrix includes:

[0099] Read the time series water shortage feature sequence and adjacency matrix, arrange them according to irrigation grid cell number, continuous sampling time and water shortage feature type, write the water content change rate, evapotranspiration potential, crop water requirement coefficient and historical water replenishment response of the same irrigation grid cell into the same time slice, and generate the model input data structure;

[0100] The model input data structure and adjacency matrix are input into the graph convolutional layer of the improved MTGNN model. Based on the water migration channels marked in the adjacency matrix, the water shortage characteristics of adjacent irrigation grid cells are spatially propagated and aggregated to generate spatially propagated features. Specifically, the generation of spatially propagated features is as follows:

[0101] The soil moisture content change rate, evapotranspiration potential, crop water requirement coefficient, and historical water replenishment response of each irrigation grid unit at the same sampling time are read from the model input data structure and arranged according to the irrigation grid unit number to form node water shortage characteristics; the adjacency matrix is ​​read row by row, and the position with a value of 1 in the matrix is ​​identified as a direct water migration channel; for each irrigation grid unit, the adjacent irrigation grid units with a direct water migration relationship are found to generate the corresponding neighborhood node set;

[0102] Read the center distance, slope angle difference, and hydraulic gradient difference between the current irrigation grid cell and each adjacent irrigation grid cell. According to the rule that the smaller the center distance, the smaller the slope angle difference, and the smaller the hydraulic gradient difference, the greater the propagation weight, assign propagation weights to each water migration channel. Then, normalize all propagation weights corresponding to the same irrigation grid cell so that the sum of all propagation weights is 1.

[0103] The node water shortage characteristics of each adjacent irrigation grid unit are multiplied by their corresponding propagation weights, and then accumulated item by item in the order of soil moisture content change rate, evapotranspiration potential, crop water requirement coefficient, and historical water replenishment response to obtain the neighborhood aggregation characteristics. The node water shortage characteristics of the current irrigation grid unit are weighted and merged with the neighborhood aggregation characteristics at a ratio of 0.4 and 0.6, so that the current irrigation grid unit retains its own water shortage state and receives water shortage information propagated by adjacent irrigation grid units along the water migration channel, thus generating the spatial propagation characteristics corresponding to the current sampling time.

[0104] A dual-channel spatial interaction layer is set between the graph convolutional layer and the gated temporal convolutional layer. Using spatial propagation features as input, it reads crop water requirement coefficient, crop growth indicators, meteorological elements, and evapotranspiration potential to construct dynamic crop dependency adjacency matrices and dynamic climate dependency adjacency matrices. Specifically, the construction of these matrices is as follows:

[0105] Read the crop water requirement coefficient, crop height, leaf area index, canopy coverage and current growth stage of each irrigation grid unit at the current sampling time, arrange them according to the irrigation grid unit number and fixed field order, generate crop status records, combine the crop status records of all irrigation grid units in pairs, calculate the difference in crop water requirement coefficient, plant height, leaf area index and canopy coverage between any two irrigation grid units, and determine whether the two irrigation grid units are in the same growth stage or adjacent growth stages;

[0106] Grid pairs with differences in crop water requirement coefficients not exceeding 0.15, plant height not exceeding 20 cm, leaf area index not exceeding 0.8, and canopy coverage not exceeding 15%, and belonging to the same or adjacent growth stages, are identified as crop-dependent grid pairs. Following the rule that the smaller the differences in crop water requirement coefficients, plant height, leaf area index, and canopy coverage, the greater the dependence intensity, crop dependence weights are assigned to each crop-dependent grid pair. All crop dependence weights corresponding to each irrigation grid unit are normalized. The normalized crop dependence weights are written to the corresponding matrix positions, positions that do not meet the crop dependence conditions are written with 0, and the diagonal positions of the matrix are written with 1, generating the dynamic crop dependence adjacency matrix corresponding to the current sampling time.

[0107] Read the temperature, wind speed, solar radiation, air humidity, and evapotranspiration potential of each irrigation grid unit at the current sampling time. Arrange the above data according to the irrigation grid unit number and fixed field order to generate a climate state record. Combine the climate state records of all irrigation grid units in pairs and calculate the temperature difference, wind speed difference, solar radiation difference, air humidity difference, and evapotranspiration potential difference between any two irrigation grid units.

[0108] Grid pairs with temperature differences not exceeding 3℃, wind speed differences not exceeding 1.5 m / s, solar radiation differences not exceeding 120 watts per square meter, air humidity differences not exceeding 10%, and evapotranspiration potential differences not exceeding 0.12 are identified as climate-dependent grid pairs. Following the rule that the smaller the temperature difference, wind speed difference, solar radiation difference, air humidity difference, and evapotranspiration potential difference, the greater the dependence intensity, climate dependence weights are assigned to each climate-dependent grid pair. All climate dependence weights corresponding to each irrigation grid unit are normalized. The normalized climate dependence weights are written to the corresponding matrix positions, positions that do not meet the climate dependence conditions are written with 0, and the diagonal positions of the matrix are written with 1, generating the dynamic climate dependence adjacency matrix corresponding to the current sampling time.

[0109] Based on a dynamic crop dependency adjacency matrix, node embedding updates are performed on irrigation grid units with similar growth stages, similar crop water requirement coefficients, and similar historical water replenishment responses to generate crop dependency channel features. Similarly, based on a dynamic climate dependency adjacency matrix, node embedding updates are performed on irrigation grid units with similar meteorological elements and similar evapotranspiration potential trends to generate climate dependency channel features.

[0110] Generate crop-dependent channel features, specifically:

[0111] Read the spatial propagation characteristics, crop water requirement coefficient, current growth stage and historical water replenishment response of each irrigation grid unit at the current sampling time, and write the above data into the node crop status record according to the irrigation grid unit number; read the dynamic crop dependency adjacency matrix line by line, determine the positions with weight greater than 0 in the matrix as crop dependency connections, and extract the associated irrigation grid units with crop dependency connections for each irrigation grid unit to generate a crop dependency neighborhood set.

[0112] Read the crop dependency weights corresponding to each associated irrigation grid unit in the crop dependency neighborhood set, compare the differences in growth stage, crop water requirement coefficient, and historical water replenishment response between the current irrigation grid unit and each associated irrigation grid unit; retain the original crop dependency weights of associated irrigation grid units in the same growth stage, multiply the crop dependency weights corresponding to associated irrigation grid units in adjacent growth stages by 0.8, multiply the crop dependency weights corresponding to associated irrigation grid units with crop water requirement coefficient differences not exceeding 0.05 and historical water replenishment response differences not exceeding 0.1 by 1.2, and normalize the adjusted crop dependency weights;

[0113] The spatial propagation features of each associated irrigation grid unit are multiplied by the normalized crop dependency weights, and then accumulated sequentially according to the fields of soil moisture content change rate, evapotranspiration potential, crop water requirement coefficient, and historical water replenishment response to obtain the crop neighborhood aggregation features. The spatial propagation features of the current irrigation grid unit itself are weighted and merged with the crop neighborhood aggregation features at a ratio of 0.4 and 0.6, and the current growth stage, crop water requirement coefficient, and historical water replenishment response are written into the corresponding feature positions to generate the crop dependency channel features of the current irrigation grid unit. The above process is repeated according to the irrigation grid unit number and sampling time to form the crop dependency channel feature sequence.

[0114] Generate climate-dependent channel features, specifically:

[0115] Read the spatial propagation characteristics, temperature, wind speed, solar radiation, air humidity and evapotranspiration potential of each irrigation grid unit at the current sampling time and the previous 3 sampling times, and write them into the node climate state record according to the irrigation grid unit number and time order; read the dynamic climate dependency adjacency matrix line by line, determine the positions with weights greater than 0 in the matrix as climate dependency connections, and extract the associated irrigation grid units with climate dependency connections for each irrigation grid unit to generate a climate dependency neighborhood set.

[0116] The direction and magnitude of evapotranspiration potential change in the current irrigation grid unit and each associated irrigation grid unit are calculated over four consecutive sampling times. Continuous increase, continuous decrease, or stability are taken as the direction of evapotranspiration potential change. When the evapotranspiration potential change directions of two irrigation grid units are consistent and the cumulative difference in change magnitude does not exceed 0.15, the corresponding climate dependence weight is retained. When the change directions are consistent but the cumulative difference in change magnitude is greater than 0.15, the corresponding climate dependence weight is multiplied by 0.8. When the change directions are inconsistent, the corresponding climate dependence weight is multiplied by 0.5. The adjusted climate dependence weights are then normalized.

[0117] The spatial propagation characteristics of each associated irrigation grid unit are multiplied by the normalized climate dependence weights, and then accumulated item by item in the order of soil moisture content change rate, evapotranspiration potential, crop water requirement coefficient, and historical water replenishment response to obtain the climate neighborhood aggregation characteristics. The spatial propagation characteristics of the current irrigation grid unit itself are weighted and merged with the climate neighborhood aggregation characteristics at a ratio of 0.4 and 0.6, and the current temperature, wind speed, solar radiation, air humidity, and evapotranspiration potential are written into the corresponding feature positions to generate the climate dependence channel characteristics of the current irrigation grid unit. The process is repeated according to the irrigation grid unit number and sampling time to form a climate dependence channel characteristic sequence.

[0118] The crop-dependent channel features and climate-dependent channel features are aligned according to irrigation grid cell number and time index. Channel splicing, feature filtering, and weight mapping are performed to generate dual-channel spatial interaction features, which are then input into the residual dilated convolution module set before each level of gated temporal convolution. The specific details of generating dual-channel spatial interaction features are as follows:

[0119] According to the irrigation grid unit number and sampling time, the crop-dependent channel features and climate-dependent channel features are read one by one. A one-to-one correspondence is established between the two types of channel features corresponding to the same irrigation grid unit and the same sampling time. The features are then spliced ​​together in the order of crop-dependent channel first and climate-dependent channel second to generate dual-channel spliced ​​features.

[0120] Read each feature item in the dual-channel splicing feature, compare the response values ​​of the same feature item in the two channels, and retain the feature items with larger differences and consistent change directions for three consecutive sampling times; for feature items that are repeated in the two channels and have a difference of no more than 0.05, only retain the feature items with larger response values; remove the feature items that are missing, abnormal, or remain unchanged for three consecutive sampling times to generate the filtered dual-channel feature.

[0121] Mapping weights are assigned based on the degree of synchronization between each selected feature and the rate of change of soil moisture content. Feature items with a higher degree of synchronization are assigned larger mapping weights, while feature items with a lower degree of synchronization are assigned smaller mapping weights. All mapping weights are then normalized. Each selected feature is multiplied by its corresponding mapping weight and arranged in a fixed order to generate dual-channel spatial interaction features.

[0122] The residual dilated convolution module extracts water shortage trends across different time spans from the dual-channel spatial interaction features. The extracted results are then residually superimposed with the original dual-channel spatial interaction features and input into a gated temporal convolutional layer to generate water shortage temporal evolution features. Specifically, the generation of water shortage temporal evolution features involves:

[0123] The dual-channel spatial interaction features corresponding to continuous sampling times of each irrigation grid unit are read and arranged according to the irrigation grid unit number and time order. Three convolution branches with expansion rates of 1, 2 and 4 are set respectively. The convolution branch with expansion rate of 1 reads the feature changes of consecutive adjacent sampling times to extract short-term water shortage fluctuations. The convolution branch with expansion rate of 2 reads the feature changes at intervals of one sampling time to extract the water shortage evolution trend over a medium time span. The convolution branch with expansion rate of 4 reads the feature changes at intervals of three sampling times to extract the long-term water shortage accumulation trend, generating short-term trend features, medium-term trend features and long-term trend features.

[0124] The short-term, medium-term, and long-term trend features are aligned and stitched together according to the corresponding irrigation grid cell number, sampling time, and feature position. Channel compression is performed on the stitched result to make the compressed feature dimension consistent with the original dual-channel spatial interaction feature dimension. The compressed multi-time span trend features are added to the original dual-channel spatial interaction features item by item, and the superposition result is normalized to generate residual time series features.

[0125] The residual temporal features are input into the gated temporal convolutional layer to generate candidate temporal features and gate weights respectively. The candidate temporal features are used to record the changes in water shortage status of each irrigation grid unit in continuous sampling time, and the gate weights are used to judge the degree of feature retention at each sampling time. The candidate temporal features are multiplied with the corresponding gate weights one by one to retain the effective temporal information corresponding to the continuous decrease in soil moisture content, the continuous increase in evapotranspiration potential, the increase in crop water demand coefficient, or the weakening of historical water replenishment response, while suppressing instantaneous fluctuations and abnormal sampling information. The water shortage temporal evolution features are generated by arranging them according to the irrigation grid unit number and time order.

[0126] The water shortage time-series evolution features are input to the output. A water shortage score is generated for each irrigation grid cell through a water shortage scoring branch. A corresponding confidence level is generated through a prediction confidence quantification head. The water shortage score, confidence level, and irrigation grid cell number are arranged from highest to lowest water shortage severity to generate a water shortage priority matrix, where:

[0127] The water shortage score for each irrigation grid cell is generated through a water shortage score branch, specifically as follows:

[0128] Read the water shortage time series evolution characteristics of each irrigation grid unit at the current sampling time, extract the soil moisture content decreasing trend, evapotranspiration potential changing trend, crop water requirement coefficient, historical water replenishment response amount and continuous water shortage duration, and map each feature item to a unified numerical range of 0 to 1; the feature items corresponding to continuous decrease in soil moisture content, continuous increase in evapotranspiration potential, increase in crop water requirement coefficient and extension of continuous water shortage time are regarded as water shortage enhancement items, and the feature items corresponding to increase in historical water replenishment response amount are regarded as water shortage inhibition items.

[0129] Multiply the characteristic value corresponding to the soil moisture content decrease trend by 0.3, the characteristic value corresponding to the evapotranspiration potential change trend by 0.25, the characteristic value corresponding to the crop water requirement coefficient by 0.2, the characteristic value corresponding to the duration of continuous water shortage by 0.15, and the characteristic value corresponding to the historical water replenishment response by 0.1. Sum the weighted results of the four water shortage enhancement items, and then subtract the weighted result of the historical water replenishment response to obtain the initial water shortage score.

[0130] When the initial water shortage score is less than 0, the water shortage score is adjusted to 0; when the initial water shortage score is greater than 1, the water shortage score is adjusted to 1; otherwise, the initial water shortage score is used as the water shortage score of the current irrigation grid unit, and the water shortage score is associated with the irrigation grid unit number and the current sampling time.

[0131] The corresponding confidence level is generated by predicting the confidence level quantization head, specifically as follows:

[0132] Read the water shortage time series evolution characteristics, interpolation confidence flag, water shortage score of the last 4 sampling times, and the number of effective connections in the dynamic crop-dependent adjacency matrix and dynamic climate-dependent adjacency matrix corresponding to the current sampling time of each irrigation grid unit; calculate the difference between the maximum and minimum values ​​of the water shortage score of the last 4 sampling times, use this difference as the score fluctuation, and calculate the proportion of effective monitoring data of the current irrigation grid unit to all input data to generate data completeness;

[0133] Irrigation grid cells with data integrity of at least 90%, score fluctuation of no more than 0.1, interpolated data ratio of no more than 20%, and at least 3 effective connections in both the dynamic crop-dependent adjacency matrix and the dynamic climate-dependent adjacency matrix are defined as stable prediction grids; irrigation grid cells with data integrity of less than 90%, score fluctuation of more than 0.1, interpolated data ratio of more than 20%, or fewer than 3 effective connections in any dynamic adjacency matrix are defined as unstable prediction grids.

[0134] A confidence weight of 0.35 is assigned to data completeness, a confidence weight of 0.3 is assigned to score stability, a confidence weight of 0.2 is assigned to the proportion of non-interpolated data, and a confidence weight of 0.15 is assigned to the connection completeness of the two dynamic adjacency matrices. Each indicator is multiplied by its corresponding confidence weight and then summed to obtain the initial confidence level. The initial confidence level is then limited to the range of 0 to 1 to generate the confidence level corresponding to the current irrigation grid cell.

[0135] When the confidence level is not lower than 0.8, the corresponding water shortage score is marked as a high-confidence prediction result; when the confidence level is lower than 0.8 but not lower than 0.6, the corresponding water shortage score is marked as a moderately confident prediction result; when the confidence level is lower than 0.6, the corresponding water shortage score is marked as a low-confidence prediction result, and the confidence level, confidence mark, water shortage score and irrigation grid cell number are associated and written into the water shortage priority matrix.

[0136] To address the problem that traditional MTGNN models, which rely solely on fixed adjacency relationships for graph structure modeling, struggle to reflect the dynamic changes in water demand between different crop growth stages and irrigation areas, this invention introduces a dual-channel spatial interaction layer between the graph convolutional layer and the gated temporal convolutional layer. This layer constructs dynamic crop-dependent adjacency matrices and dynamic climate-dependent adjacency matrices, respectively. The dynamic crop-dependent adjacency matrix establishes node associations based on crop water demand coefficients, growth stages, and historical water replenishment responses. The dynamic climate-dependent adjacency matrix establishes dynamic associations based on evapotranspiration potential, air temperature, wind speed, solar radiation, and air humidity. After aggregating neighborhood information, these layers are interactively fused, enabling the model to simultaneously learn the coupling relationship between crop growth patterns and environmental change patterns. This move beyond fixed spatial adjacency relationships enhances the ability to express real water shortage relationships between different irrigation grids and improves the accuracy of water shortage identification in complex plots.

[0137] To address the limitation of traditional temporal convolutional models, which can only extract features at a single time scale and struggle to simultaneously perceive short-term fluctuations and long-term water scarcity accumulation trends, this invention incorporates a residual dilated convolutional module before each gated temporal convolutional layer. Multi-branch convolutions with dilation rates of 1, 2, and 4 extract short-term water scarcity fluctuations, medium-term water scarcity evolution trends, and long-term water scarcity accumulation trends, respectively. These features from different time scales are then residually fused with the original dual-channel spatial interaction features before being input into the gated temporal convolutional layer. Through collaborative modeling of information at different time scales, this approach not only preserves long-term trends during continuous water scarcity evolution but also effectively suppresses interference from instantaneous sampling fluctuations and local anomalies. This results in a more complete temporal evolution feature set for water scarcity, improving the stability and continuity of water scarcity trend prediction.

[0138] To address the issue that traditional MTGNN models only provide predicted values ​​and fail to reflect the reliability of those predictions, this invention adds a prediction confidence quantification head to the model output, operating in parallel with the water shortage scoring branch. This prediction confidence quantification head comprehensively utilizes information such as data completeness, interpolated data ratio, water shortage score stability, and the completeness of the dynamic adjacency matrix to assess the reliability of the prediction results for each irrigation grid unit. This assessment, combined with the water shortage score, then generates a water shortage priority matrix. By introducing a prediction reliability evaluation mechanism, irrigation scheduling not only prioritizes water shortage levels but also comprehensively considers the reliability of the prediction results. This allows for conservative scheduling or subsequent corrections for low-reliability prediction areas, reducing the impact of abnormal data, missing sensors, and environmental fluctuations on irrigation decisions. Ultimately, this improves the reliability, stability, and engineering application capability of irrigation scheduling strategies.

[0139] In this embodiment, generating the partitioned staggered irrigation sequence includes:

[0140] Read the water shortage priority matrix, extract the number of each irrigation grid unit, water shortage score and prediction confidence, use the water shortage score and prediction confidence as the sorting criteria, sort each irrigation grid unit from high to low according to the urgency of water shortage, and generate an irrigation grid unit sorting queue.

[0141] According to the irrigation grid unit sorting queue, the current soil moisture content, target soil moisture content and historical water replenishment response of each irrigation grid unit are read in sequence. The difference between the target soil moisture content and the current soil moisture content is taken as the water replenishment gap, and the target water replenishment amount of each irrigation grid unit is calculated based on the historical water replenishment response.

[0142] The system reads the available water supply and remaining pipeline flow within the current scheduling period, divides the scheduling period into multiple consecutive time-sharing water supply windows, and sets an upper limit for allocable water volume and an upper limit for allocable flow for each time-sharing water supply window. Irrigation tasks are then written sequentially according to the irrigation grid unit sorting queue, ensuring that the total target water replenishment volume written to the same time-sharing water supply window does not exceed the upper limit for allocable water volume, and the total water supply flow does not exceed the upper limit for allocable flow. Specifically, the scheduling period is divided into multiple consecutive time-sharing water supply windows as follows:

[0143] Read the start time, end time, water supply availability change record, and pipeline remaining flow change record of the current scheduling cycle. Based on a base time length of 30 minutes, set the start time and end time of the window sequentially from the start time of the scheduling cycle. Connect the beginning and end of two adjacent time-sharing water supply windows until the current scheduling cycle is covered, and generate an initial set of time-sharing water supply windows. When the last segment of the scheduling cycle is less than 30 minutes, merge the last segment into the previous time-sharing water supply window.

[0144] The available water supply and remaining pipeline flow rate at each sampling time within each initial time-sharing water supply window are read separately. The minimum available water supply in the window is taken as the water supply constraint for that window, and the minimum remaining pipeline flow rate in the window is taken as the pipeline constraint flow rate for that window. When the change in the water supply constraint and the change in the pipeline constraint flow rate of adjacent initial time-sharing water supply windows do not exceed 10%, the adjacent windows are merged into one time-sharing water supply window. When either change exceeds 10%, the window boundary position is retained.

[0145] When the change in the available water supply or the remaining flow of the pipeline network within any initial time-sharing water supply window exceeds 15% for two consecutive sampling times, the sampling time at which the change occurred is used as the new window boundary point, and the corresponding initial time-sharing water supply window is split into two consecutive time-sharing water supply windows; the merged or split windows are renumbered according to the chronological order, and the start time, end time, water source constraint quantity, and pipeline network constraint flow of each time-sharing water supply window are recorded to generate multiple consecutive time-sharing water supply windows;

[0146] Read the water migration channels marked in the adjacency matrix, merge the irrigation grid units that are written to the same time-sharing water supply window and are spatially adjacent, have similar target water replenishment amounts and similar water shortage levels to form a continuous irrigation area, and allocate irrigation time periods, water supply flow rates and valve openings according to the target water replenishment amounts of each irrigation grid unit in the continuous irrigation area.

[0147] Constraint verification is performed on the continuous irrigation areas within each time-sharing water supply window. Irrigation tasks exceeding the available water supply, remaining flow in the pipeline network, or window time capacity are postponed to the next time-sharing water supply window. Irrigation time periods, water supply flow, and valve opening are reallocated, and a staggered irrigation sequence is generated by zone number, execution time, duration, water supply flow, and valve opening.

[0148] In this embodiment, the step of calculating the deviation between the actual water replenishment volume and the target water replenishment volume and adjusting the water supply flow rate, valve opening, or irrigation period includes:

[0149] The corresponding continuous irrigation zones are started sequentially according to the execution time set in the staggered irrigation sequence. The corresponding valves are opened to the preset opening degree, and water is continuously supplied according to the allocated water supply flow rate. After irrigation is completed, the corresponding valves are closed, thus completing the irrigation execution of each continuous irrigation zone. The preset opening degree is:

[0150] Read the allocated water supply flow rate of the corresponding continuous irrigation area in the staggered irrigation sequence, and read the rated maximum flow rate of the valve connected to the continuous irrigation area; convert the allocated water supply flow rate to the rated maximum flow rate to obtain the initial valve opening, and convert the initial valve opening to the corresponding percentage opening.

[0151] When the initial valve opening is less than 30%, the preset opening is set to 30% to avoid unstable water supply caused by the valve opening being too small; when the initial valve opening is greater than 90%, the preset opening is set to 90% to retain the flow regulation margin of the pipeline network; when the initial valve opening is between 30% and 90%, the initial valve opening is used as the preset opening.

[0152] When multiple valves are set up in the same continuous irrigation area, the water supply flow is allocated according to the number of irrigation grid units controlled by each valve and the corresponding target water replenishment. The preset opening degree of each valve is calculated separately, and the preset opening degree is rounded up in increments of 5% to generate valve opening control values ​​between 30%, 35%, 40% and 90%.

[0153] During the irrigation process, the soil moisture content, water flow rate and valve opening of each irrigation grid unit are collected in real time according to the preset sampling period. The actual water supply during the irrigation period is obtained by accumulating the water flow rate within each sampling period. Combined with the change in soil moisture content before and after irrigation, the actual water replenishment of each irrigation grid unit is determined. The preset sampling period is 1 minute.

[0154] The actual water replenishment amount for each irrigation grid unit is determined as follows:

[0155] Read the soil moisture content of each irrigation grid unit for three consecutive sampling periods before the valve is opened, and take the average of the three soil moisture contents as the soil moisture content before irrigation; read the soil moisture content at 10 minutes, 20 minutes and 30 minutes after the valve is closed, and take the average of the three soil moisture contents as the soil moisture content after irrigation, and calculate the increase of soil moisture content after irrigation relative to soil moisture content before irrigation.

[0156] The grid area, effective soil wetting depth, and soil bulk density of the irrigation grid unit are read. The increase in soil moisture content is converted into water volume according to the grid area, effective soil wetting depth, and soil bulk density to obtain the increase in soil moisture content corresponding to the change in soil moisture content. The effective soil wetting depth is determined according to the distribution of crop roots: 0.2 meters during the seedling stage, 0.3 meters during the vegetative growth stage, 0.4 meters during the flowering and fruiting stage, and 0.3 meters during the maturity stage.

[0157] Using a 1-minute sampling cycle, the water supply flow rate at each sampling moment is converted into the water supply volume within the corresponding sampling cycle. The water supply volumes for all sampling cycles during the irrigation period are then summed to obtain the actual water supply volume. When the increase in water content is not greater than the actual water supply volume, the increase in water content is taken as the actual replenishment volume. When the increase in water content is greater than the actual water supply volume, the actual water supply volume is taken as the actual replenishment volume. When the soil moisture content after irrigation is not higher than the soil moisture content before irrigation, the actual replenishment volume is recorded as 0, and the corresponding irrigation record is marked as an invalid replenishment record.

[0158] Read the target water replenishment amount corresponding to each irrigation grid unit, compare the actual water replenishment amount with the target water replenishment amount one by one, obtain the water replenishment deviation of each irrigation grid unit, and generate the water replenishment deviation result based on the magnitude and direction of the water replenishment deviation.

[0159] Based on the water replenishment deviation results, adjust the water supply flow, valve opening, or irrigation period of the corresponding irrigation grid unit. Specifically, for irrigation grid units where the actual water replenishment is lower than the target water replenishment, increase the water supply flow, increase the valve opening, or extend the irrigation period. For irrigation grid units where the actual water replenishment is higher than the target water replenishment, decrease the water supply flow, decrease the valve opening, or shorten the irrigation period, and generate the corrected irrigation execution result.

[0160] The soil moisture content, water supply flow, valve opening, actual water replenishment, water replenishment deviation, and irrigation period corresponding to the corrected irrigation execution results are written into the multi-source irrigation status data. The irrigation data set is updated, and the operating status of the corresponding irrigation grid unit is regenerated, providing updated input data for water shortage analysis, priority ranking, and irrigation decision-making in the next scheduling cycle.

[0161] refer to Figure 4 A soil water-saving irrigation management system based on multi-source data includes:

[0162] The data acquisition and preprocessing module is used to collect multi-source irrigation status data and perform preprocessing to generate an irrigation dataset;

[0163] The spatial grid construction module is used to complete unified coordinate mapping, irrigation grid cell division, adjacency matrix construction and interpolation compensation processing based on irrigation data sets, and generate multi-source grid data fields.

[0164] The water shortage feature construction module is used to extract soil moisture content change rate, evapotranspiration potential, crop water requirement coefficient and historical water replenishment response based on multi-source grid data field, and construct a time-series water shortage feature sequence.

[0165] The water shortage prediction and analysis module is used to input the time-series water shortage feature sequence and adjacency matrix into the improved MTGNN model to generate the water shortage priority matrix for each irrigation grid unit.

[0166] The zoned scheduling decision module is used to construct time-sharing water supply windows based on the water shortage priority matrix, combined with the target water replenishment volume, water source availability and pipeline network remaining flow, to perform staggered scheduling for continuous irrigation areas and generate zoned staggered irrigation sequences.

[0167] The irrigation feedback update module is used to execute the regional staggered irrigation sequence, collect the irrigation execution status, calculate the deviation between the actual water replenishment and the target water replenishment, adaptively adjust the irrigation parameters, and update the multi-source irrigation status data and irrigation decision results.

[0168] Example 1: In one farmland water-saving irrigation scheduling cycle, the system receives multi-source irrigation status data uploaded from soil sensor nodes, valve controllers, pipeline flow meters, and meteorological acquisition devices. The simulated plot area is 3.6 hectares, with 192 soil moisture sensor nodes, 18 pipeline flow meters, and 24 electric valves deployed. A total of 2304 soil moisture records, 432 water supply flow records, 576 valve opening records, and 96 meteorological records were received. The original data contained 41 missing soil moisture records, 12 water supply flow records with instantaneous jumps, and 9 valve opening records with delayed uploads. The system first converts soil moisture content, water supply flow, air temperature, wind speed, solar radiation, and air humidity to the 0–1 range, while removing abnormal flow values ​​exceeding 35% of the average of two adjacent sampling points. After processing, the effective data ratio increased from 96.8% to 99.1%, and the data entered the subsequent grid division process.

[0169] The system reads the spatial coordinates, slope angle, and hydraulic gradient parameters of 192 monitoring nodes. Clustering results show that region 4 contains 28 monitoring nodes, with a slope angle variation range of 32° and a hydraulic gradient variation range of 0.13, meeting the criteria for complex terrain. Therefore, this region is divided into irrigation grid units with a side length of 5m. The remaining regions have slope angle variations ranging from 8° to 21° and hydraulic gradient variations ranging from 0.03 to 0.09, and are divided into irrigation grid units with a side length of 15m. The first division yields 124 grids, among which 6 groups of adjacent grids have a difference of more than 5 monitoring nodes, with a maximum difference of 9. The system continues to perform a second division on the grids with denser nodes. After completion, 136 irrigation grid units are formed, including 48 grids for complex terrain and 88 grids for gentle terrain.

[0170] During the adjacency matrix construction process, the system reads the center distance, aspect difference, hydraulic gradient difference, and gradient direction for each of the 136 irrigation grid cells. When the center distance is ≤15m, aspect difference is ≤20°, hydraulic gradient difference is ≤0.08, and the gradient directions are consistent, a 1 is written into the adjacency matrix; otherwise, a 0 is written. A 136×136 adjacency matrix is ​​generated with 412 effective adjacency edges. For the missing soil moisture content at the current sampling point G037, the system selects G033, G038, G041, and G042 from the adjacency matrix as reference grids, with comprehensive weights of 0.34, 0.27, 0.22, and 0.17, respectively. Interpolation yields a current soil moisture content of 21.6% for G037, and the interpolation confidence value of 0.34 is recorded.

[0171] In the water-deficient characteristic construction stage, the system read the changes in soil moisture content at 12 consecutive sampling points. The soil moisture content of G018 decreased from 21.8% to 18.9%, a drop of 2.9 percentage points; G122 decreased from 20.6% to 17.8%, a drop of 2.8 percentage points; and G097 changed from 26.4% to 26.1%, a drop of only 0.3 percentage points. In the meteorological records, the temperature scale value was 0.74, the wind speed scale value was 0.62, the solar radiation scale value was 0.81, and the air humidity scale value was 0.29. After processing with a temperature suppression factor of 0.3, a wind speed suppression factor of 0.2, a solar radiation suppression factor of 0.35, and an air humidity suppression factor of 0.15, the evapotranspiration potential of the area where G122 is located was obtained as 0.72. The system continued to read plant height, leaf area index, canopy coverage, and cumulative growing days, classifying G018 and G122 as being in the flowering and fruiting stage, with corresponding crop water requirement coefficients of 0.92 and 0.95, respectively. The five most recent water replenishment responses per unit area for G122 were 0.68, 0.64, 0.61, 0.59, and 0.55. After weighting by 0.3, 0.25, 0.2, 0.15, and 0.1, the historical water replenishment response amount is 0.62.

[0172] The training samples consist of irrigation grid cells and 12 consecutive sampling points, totaling 8160 samples, including 5712 training samples, 1632 validation samples, and 816 test samples. Sample G018 recorded a 2.9 percentage point decrease in moisture content, an evapotranspiration potential of 0.75, a water demand coefficient of 0.92, a historical water replenishment response of 0.58, and an actual water replenishment demand of 0.83; Sample G046 recorded a 0.6 percentage point decrease in moisture content, an evapotranspiration potential of 0.56, a water demand coefficient of 0.84, a historical water replenishment response of 0.72, and an actual water replenishment demand of 0.51; Sample G097 recorded a 0.3 percentage point decrease in moisture content, an evapotranspiration potential of 0.47, a water demand coefficient of 0.78, a historical water replenishment response of 0.81, and an actual water replenishment demand of 0.22; Sample G122 recorded a 2.8 percentage point decrease in moisture content, an evapotranspiration potential of 0.83, a water demand coefficient of 0.95, a historical water replenishment response of 0.52, and an actual water replenishment demand of 0.88. Through this sample structure, the system can distinguish between different water shortage types: "currently low moisture content but good water replenishment response" and "continuously decreasing moisture content and weak water replenishment response."

[0173] The system inputs the temporal water shortage feature sequence and adjacency matrix into the improved MTGNN model. During the spatial propagation phase, the system merges features with a value of 0.4 for itself and 0.6 for its neighbors. G122 has effective adjacency relationships with G118, G119, G123, and G126. Before fusion, the mean spatial propagation feature value of G122 was 0.71, which increased to 0.78 after fusion. The dynamic crop-dependent adjacency matrix screened out 156 connections at the same stage with a water requirement difference ≤ 0.05 and a historical response difference ≤ 0.10; the dynamic climate-dependent adjacency matrix screened out 184 connections with consistent evapotranspiration directions. After dual-channel spatial interaction, the crop channel response of G122 was 0.82, and the climate channel response was 0.79. The system retained 9 effective feature terms and removed 2 feature terms that remained unchanged for 3 consecutive sampling points.

[0174] In the residual dilated convolution module, the system uses dilation rates of 1, 2, and 4 to extract the water shortage trend. The short-term trend response of G122 is 0.76, the medium-term trend response is 0.81, and the long-term trend response is 0.86, indicating that the grid does not fluctuate at a single point, but rather accumulates continuously due to water shortage. After residual stacking, the data enters a gated temporal convolutional layer with a mean gate weight of 0.73. The weights of abnormal fluctuation sampling points are reduced to 0.28, while the weights of continuously decreasing sampling points are maintained between 0.71 and 0.83, ultimately generating the temporal evolution features of water shortage.

[0175] When calculating the water shortage score at the output end, G122 showed a moisture content decreasing trend of 0.88, an evapotranspiration potential trend of 0.83, a crop water requirement coefficient of 0.95, a continuous water shortage duration of 0.80, and a historical water replenishment response of 0.52. After processing with values ​​of 0.3, 0.25, 0.2, 0.15, and 0.1, a water shortage score of 0.89 was obtained. The confidence quantification head read data with a completeness of 94.2%, a score fluctuation of 0.07 over the last four sampling points, an interpolated data ratio of 6.4%, and at least three effective connections in the dynamic adjacency matrix, generating a confidence score of 0.87. Ultimately, G122 ranked 1st, G018 ranked 2nd with a score of 0.82 and a confidence score of 0.84, G046 ranked 41st with a score of 0.59 and a confidence score of 0.79, and G097 ranked 126th with a score of 0.21 and a confidence score of 0.91.

[0176] During the scheduling phase, the system reads the available water supply as 280 m³, the available flow rate at the pipeline inlet as 48 m³ / h, and the remaining flow rate at the end branch pipe as 6.5 m³ / h. The system initially divides the water supply windows into 30-minute intervals. It is found that the remaining flow rate in the third window changes by more than 15% between two consecutive sampling points. Therefore, this window is split, ultimately forming six continuous time-sharing water supply windows. The system merges G118–G126 into Region A, with a target water supply of 32.4 m³ and an allocated flow rate of 12.0 m³ / h; merges G015–G023 into Region B, with a target water supply of 28.6 m³ and an allocated flow rate of 10.5 m³ / h; and merges G044–G052 into Region C, with a target water supply of 19.8 m³ and an allocated flow rate of 8.2 m³ / h. Valve opening is limited to 30%–90% and rounded in 5% increments. The valve opening is 80% and 75% in Zone A, 60%, 65% and 55% in Zone B, and 50% and 45% in Zone C.

[0177] During irrigation, the system collects soil moisture content, water supply flow rate, and valve opening at 1-minute sampling intervals. In Area A, a total of 31.7 m³ of water was supplied. Before the valves were opened, the average moisture content at the three sampling points was 18.3%. After the valves were closed, the moisture contents at 10, 20, and 30 minutes were 22.9%, 23.5%, and 24.7%, respectively, with a post-irrigation average of 23.7%. The actual replenishment volume was 30.9 m³. In Area B, a total of 28.1 m³ of water was supplied. The pre-irrigation moisture content was 19.1%, and the post-irrigation average moisture content was 24.2%, with an actual replenishment volume of 27.5 m³. In Area C, a total of 19.4 m³ of water was supplied. The pre-irrigation moisture content was 20.4%, and the post-irrigation average moisture content was 23.6%, with an actual replenishment volume of 18.7 m³. The system writes back the actual replenishment volume, water shortage score, confidence level, and valve opening to update the historical replenishment response for the next round.

[0178] On the same batch of test data, the traditional fixed-period irrigation method used 246.8 m³ of water, while the method of this invention used 189.5 m³, saving 57.3 m³, a water saving ratio of 23.2%. After irrigation using the traditional method, 21 grids still had water content below the target lower limit, while the method of this invention had 5. The traditional method had 34 grids with excessive water replenishment, while the method of this invention had 9. The peak flow rate of the pipeline network using the traditional method was 51.6 m³ / h, while the method of this invention was 46.9 m³ / h. The traditional method had 68 valve opening and closing operations, while the method of this invention had 41. The accuracy rate of high water shortage identification using the traditional method was 76.4%, while the method of this invention was 91.8%. The average deviation of water shortage scoring using the traditional method was 0.157, while the method of this invention was 0.068.

[0179] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A soil water-saving irrigation management method based on multi-source data, characterized in that, include: Collect multi-source irrigation status data, preprocess the multi-source irrigation status data, and generate an irrigation dataset; The irrigation dataset is mapped to a unified plot spatial coordinate system, irrigation grid units are divided according to monitoring locations, an adjacency matrix is ​​constructed, interpolation compensation is performed, and a spatially consistent multi-source grid data field is formed. For multi-source grid data fields, the rate of change of soil moisture content, evapotranspiration potential, crop water requirement coefficient and historical water replenishment response are calculated to construct a time-series water shortage characteristic sequence; The temporal water shortage feature sequence and adjacency matrix are input into the improved MTGNN model. The improved MTGNN model connects a dual-channel spatial interaction layer between graph convolution and gated temporal convolution. The node embedding is updated in parallel based on the dynamic crop-dependent adjacency matrix and the dynamic climate-dependent adjacency matrix, respectively. A residual dilated convolution module is inserted before each level of gated temporal convolution to expand the temporal receptive field. A prediction confidence quantization head is superimposed at the output end to output the water shortage priority matrix. The irrigation grid units are sorted according to the water shortage priority matrix. The time-sharing water supply window is constructed by combining the target water replenishment, water supply availability and pipeline remaining flow. The continuous irrigation area is merged according to the water migration relationship of adjacent grids. The irrigation time period, water supply flow and valve opening are allocated in sequence. Irrigation tasks that exceed the water volume or flow constraints are reordered in sequence to generate a zoned staggered irrigation sequence. The system executes a zoned staggered irrigation sequence, collects soil moisture content, water supply flow rate, and valve opening in real time for each irrigation grid unit, calculates the deviation between the actual water replenishment and the target water replenishment, and adjusts the water supply flow rate, valve opening, or irrigation period. The results are then written back to the multi-source irrigation status data to update irrigation decisions.

2. The soil water-saving irrigation management method based on multi-source data according to claim 1, characterized in that, The multi-source irrigation status data includes soil moisture content, soil temperature, electrical conductivity, crop growth indicators, meteorological elements, valve operating status, water supply flow rate, and historical irrigation records.

3. The soil water-saving irrigation management method based on multi-source data according to claim 1, characterized in that, The preprocessing of multi-source irrigation status data to generate an irrigation data set includes: synchronizing the multi-source irrigation status data according to the collection time; detecting and removing missing, duplicate, and abnormal data; uniformly converting the collection formats of different data sources; completing data association based on monitoring nodes; and integrating the processed multi-source irrigation status data according to time sequence and monitoring location to generate an irrigation data set.

4. The soil water-saving irrigation management method based on multi-source data according to claim 1, characterized in that, The formation of a spatially consistent multi-source grid data field includes: Read the latitude, longitude, elevation and measurement values ​​of each monitoring node in the irrigation data set, use the same projection coordinate system for coordinate transformation, and add the topographic slope angle and near-surface hydraulic gradient coefficient to each monitoring node; Hierarchical clustering is performed based on the spatial density of monitoring nodes in a unified coordinate system, slope aspect consistency, and hydraulic gradient differences. The clustering results automatically delineate the grid boundaries, resulting in irrigation grid cells with side lengths that adapt to terrain complexity. The measured values ​​of each monitoring node are assigned to the center of their respective grid. Based on the horizontal distance, slope aspect connectivity, and hydraulic gradient direction between the centers of irrigation grid units, the direct water migration probability between any two grids is calculated in real time. If the distance threshold, slope aspect, and hydraulic gradient are all satisfied at the same time, they are recorded as adjacency relationships, and an adjacency matrix that is dynamically updated with changes in surface conditions is constructed. For irrigation grid cells with missing measurements, a set of spatially adjacent irrigation grid cells with complete data from the two most recent sampling periods is selected based on the adjacency matrix. Bidirectional spatiotemporal weighted interpolation is performed by assigning weights according to spatial proximity and temporal similarity to calculate the missing measurements and write them into the corresponding irrigation grid cells. At the same time, the interpolation confidence flag is recorded. All irrigation grid cells are summarized according to a unified coordinate index order. The center coordinates, boundary range, adjacency matrix entries, monitoring values ​​and interpolation confidence indicators of each grid are recorded. The resulting spatially consistent multi-source grid data field is then generated and output.

5. The soil water-saving irrigation management method based on multi-source data according to claim 1, characterized in that, The construction of the time-series water shortage feature sequence includes: Soil moisture content data is read from the multi-source grid data field according to the irrigation grid cell number and time index. The difference between the soil moisture content at the current sampling time and the soil moisture content at the previous sampling time is calculated. The difference is amortized according to the time interval between two adjacent samplings to obtain the rate of change of soil moisture content. A sequence of the rate of change of soil moisture content is generated in chronological order. Meteorological elements corresponding to each irrigation grid unit at the time are read from the multi-source grid data field. Temperature, wind speed, solar radiation and air humidity are processed at the same scale. The increase in temperature, wind speed and solar radiation are taken as the evapotranspiration enhancement term, and the increase in air humidity is taken as the evapotranspiration inhibition term. The evapotranspiration potential is obtained by weighting and summarizing according to the preset weights and generating an evapotranspiration potential sequence in chronological order. The crop growth index corresponding to each irrigation grid unit is read from the multi-source grid data field. The crop growth index is matched with the preset crop growth stage range to determine the current growth stage of each irrigation grid unit. The crop water demand coefficient corresponding to the growth stage is read and the crop water demand coefficient sequence is generated in time order. Historical irrigation records of each irrigation grid unit are read from the multi-source grid data field. Soil moisture content before irrigation, soil moisture content after irrigation and corresponding water supply are extracted. The increase of soil moisture content after irrigation relative to soil moisture content before irrigation is converted according to the water supply to obtain the single water replenishment response. The most recent single water replenishment response is smoothed to generate a historical water replenishment response sequence. According to the irrigation grid unit number and time index, the water content change rate sequence, evapotranspiration potential sequence, crop water requirement coefficient sequence and historical water replenishment response sequence are aligned. The four types of features of the same irrigation grid unit at the same sampling time are arranged in a fixed order to construct a single-time water shortage feature record. The single-time water shortage feature records of consecutive sampling times are spliced ​​together to generate a time-series water shortage feature sequence.

6. The soil water-saving irrigation management method based on multi-source data according to claim 1, characterized in that, The output water shortage priority matrix includes: Read the time series water shortage feature sequence and adjacency matrix, arrange them according to irrigation grid cell number, continuous sampling time and water shortage feature type, write the water content change rate, evapotranspiration potential, crop water requirement coefficient and historical water replenishment response of the same irrigation grid cell into the same time slice, and generate the model input data structure; The model input data structure and adjacency matrix are input into the graph convolutional layer of the improved MTGNN model. Based on the water migration channels marked in the adjacency matrix, the water shortage characteristics of adjacent irrigation grid cells are spatially propagated and aggregated to generate spatial propagation features. A dual-channel spatial interaction layer is set between the graph convolutional layer and the gated temporal convolutional layer. Taking spatial propagation characteristics as input, the crop water demand coefficient, crop growth index, meteorological elements and evapotranspiration potential are read respectively to construct a dynamic crop dependency adjacency matrix and a dynamic climate dependency adjacency matrix. Based on the dynamic crop dependency adjacency matrix, node embedding and updating are performed on irrigation grid units with similar growth stages, similar crop water demand coefficients, and similar historical water replenishment response to generate crop dependency channel features; based on the dynamic climate dependency adjacency matrix, node embedding and updating are performed on irrigation grid units with similar meteorological elements and similar evapotranspiration potential change trends to generate climate dependency channel features. The crop-dependent channel features and climate-dependent channel features are aligned according to the irrigation grid cell number and time index. Channel splicing, feature filtering and weight mapping are performed to generate dual-channel spatial interaction features, which are then input into the residual dilated convolution module set before each level of gated temporal convolution. The residual dilated convolution module is used to extract the water shortage trend of the dual-channel spatial interaction features at different time spans. The extracted results are then superimposed with the original dual-channel spatial interaction features by residual superposition and input into the gated temporal convolutional layer to generate water shortage temporal evolution features. The water shortage time series evolution features are input to the output terminal. The water shortage score of each irrigation grid unit is generated through the water shortage score branch. The corresponding confidence score is generated through the prediction confidence quantification head. The water shortage score, confidence score and irrigation grid unit number are arranged from high to low according to the degree of water shortage to generate a water shortage priority matrix.

7. The soil water-saving irrigation management method based on multi-source data according to claim 1, characterized in that, The generation of partitioned staggered irrigation sequences includes: Read the water shortage priority matrix, extract the number of each irrigation grid unit, water shortage score and prediction confidence, use the water shortage score and prediction confidence as the sorting criteria, sort each irrigation grid unit from high to low according to the urgency of water shortage, and generate an irrigation grid unit sorting queue. According to the irrigation grid unit sorting queue, the current soil moisture content, target soil moisture content and historical water replenishment response of each irrigation grid unit are read in sequence. The difference between the target soil moisture content and the current soil moisture content is taken as the water replenishment gap, and the target water replenishment amount of each irrigation grid unit is calculated based on the historical water replenishment response. Read the available water supply and remaining flow in the pipeline network within the current scheduling cycle, divide the scheduling cycle into multiple continuous time-sharing water supply windows, and set the upper limit of allocable water volume and the upper limit of allocable flow for each time-sharing water supply window. Write irrigation tasks sequentially according to the irrigation grid unit sorting queue, so that the total target water replenishment volume written to the same time-sharing water supply window does not exceed the upper limit of allocable water volume and the total water supply flow does not exceed the upper limit of allocable flow. Read the water migration channels marked in the adjacency matrix, merge the irrigation grid units that are written to the same time-sharing water supply window and are spatially adjacent, have similar target water replenishment amounts and similar water shortage levels to form a continuous irrigation area, and allocate irrigation time periods, water supply flow rates and valve openings according to the target water replenishment amounts of each irrigation grid unit in the continuous irrigation area. Constraint verification is performed on the continuous irrigation areas within each time-sharing water supply window. Irrigation tasks exceeding the available water supply, remaining flow in the pipeline network, or window time capacity are postponed to the next time-sharing water supply window. Irrigation time periods, water supply flow, and valve opening are reallocated, and a staggered irrigation sequence is generated by zone number, execution time, duration, water supply flow, and valve opening.

8. The soil water-saving irrigation management method based on multi-source data according to claim 1, characterized in that, The calculation of the deviation between the actual water replenishment and the target water replenishment, and the adjustment of the water supply flow rate, valve opening, or irrigation period, includes: According to the execution time set in the staggered irrigation sequence, the corresponding continuous irrigation areas are started in sequence, the corresponding valves are opened to the preset opening degree, and water is continuously supplied according to the allocated water supply flow. After the irrigation is completed, the corresponding valves are closed to complete the irrigation execution of each continuous irrigation area. During the irrigation process, the soil moisture content, water flow rate and valve opening of each irrigation grid unit are collected in real time according to the preset sampling cycle. The actual water supply during the irrigation period is obtained by accumulating the water flow rate in each sampling cycle, and the actual water replenishment of each irrigation grid unit is determined by combining the changes in soil moisture content before and after irrigation. Read the target water replenishment amount corresponding to each irrigation grid unit, compare the actual water replenishment amount with the target water replenishment amount one by one, obtain the water replenishment deviation of each irrigation grid unit, and generate the water replenishment deviation result based on the magnitude and direction of the water replenishment deviation. Based on the water replenishment deviation results, adjust the water supply flow, valve opening, or irrigation period of the corresponding irrigation grid unit. Specifically, for irrigation grid units where the actual water replenishment is lower than the target water replenishment, increase the water supply flow, increase the valve opening, or extend the irrigation period. For irrigation grid units where the actual water replenishment is higher than the target water replenishment, decrease the water supply flow, decrease the valve opening, or shorten the irrigation period, and generate the corrected irrigation execution result. The soil moisture content, water supply flow, valve opening, actual water replenishment, water replenishment deviation, and irrigation period corresponding to the corrected irrigation execution results are written into the multi-source irrigation status data. The irrigation data set is updated, and the operating status of the corresponding irrigation grid unit is regenerated, providing updated input data for water shortage analysis, priority ranking, and irrigation decision-making in the next scheduling cycle.

9. A soil water-saving irrigation management system based on multi-source data, executing the soil water-saving irrigation management method based on multi-source data as described in any one of claims 1 to 8, characterized in that, Includes the following modules: The data acquisition and preprocessing module is used to collect multi-source irrigation status data and perform preprocessing to generate an irrigation dataset; The spatial grid construction module is used to complete unified coordinate mapping, irrigation grid cell division, adjacency matrix construction and interpolation compensation processing based on irrigation data sets, and generate multi-source grid data fields. The water shortage feature construction module is used to extract soil moisture content change rate, evapotranspiration potential, crop water requirement coefficient and historical water replenishment response based on multi-source grid data field, and construct a time-series water shortage feature sequence. The water shortage prediction and analysis module is used to input the time-series water shortage feature sequence and adjacency matrix into the improved MTGNN model to generate the water shortage priority matrix for each irrigation grid unit. The zoned scheduling decision module is used to construct time-sharing water supply windows based on the water shortage priority matrix, combined with the target water replenishment volume, water source availability and pipeline network remaining flow, to perform staggered scheduling for continuous irrigation areas and generate zoned staggered irrigation sequences. The irrigation feedback update module is used to execute the regional staggered irrigation sequence, collect the irrigation execution status, calculate the deviation between the actual water replenishment and the target water replenishment, adaptively adjust the irrigation parameters, and update the multi-source irrigation status data and irrigation decision results.

Images