Irrigation District Water Demand Prediction Method Based on Multi-Source Remote Sensing Data
By improving the irrigation district water demand prediction method, a high-density mutually exclusive connected domain identification mechanism and a dual-branch structure are adopted, which solves the problem of insufficient differentiation of spatial heterogeneity and temporal characteristics in traditional methods, and realizes high-precision irrigation district water demand prediction and zoned precision irrigation.
Patent Information
- Application Number
- CN202611613182.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-19
- Publication Date
- 2026-11-17
AI Technical Summary
Traditional irrigation district water demand forecasting methods do not fully consider the spatial heterogeneity of crop growth status within the irrigation district, resulting in insufficient water demand forecasting accuracy and poor regional adaptability of the forecast results. Existing clustering algorithms cannot effectively handle inter-cluster adhesion, and the time weight sharing mechanism is difficult to distinguish the degree of feature contribution at different time steps, resulting in low accuracy of long-term water demand forecasting.
The clustering algorithm is improved by adopting a high-density mutually exclusive connected component identification mechanism, a neighborhood distance adaptive calculation strategy, and an automatic boundary stripping strategy. A dual-branch structure with gated cyclic branches and time-specific dynamic memory branches running in parallel is constructed. Combined with a time attention weighting layer, a combined coding architecture is formed for irrigation area water demand prediction.
It improved the accuracy and spatial continuity of crop growth heterogeneity identification within irrigation areas, enhanced the ability to express the correlation between crop growth characteristics and water demand variation patterns in different regions, improved the accuracy and spatial matching ability of water demand forecasting, and improved the accuracy of long-term water demand forecasting.
Smart Images

Figure CN123117732A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, specifically to a method for predicting water demand in irrigation areas based on multi-source remote sensing data. Background Technology
[0002] The irrigation district water demand prediction method based on multi-source remote sensing data is an intelligent agricultural hydrological prediction technology for irrigation district water resource management. This method relies on core technologies of artificial intelligence and big data analysis. By collecting multi-source remote sensing observation data of the entire irrigation district and combining it with big data analysis algorithms, it accurately extrapolates the irrigation water demand of the irrigation district, realizing high-precision dynamic prediction of water demand at the irrigation district scale. This effectively improves the scientific nature of irrigation water resource allocation and the accuracy of irrigation scheduling, reduces the waste of agricultural irrigation water resources, and provides reliable data support and decision-making basis for refined water resource management in irrigation districts.
[0003] However, traditional irrigation district water demand forecasting methods use a uniform scale modeling across the entire irrigation district, failing to fully consider the spatial heterogeneity of crop growth states within the irrigation district. This results in technical problems such as the inability to accurately express the differences in crop water demand characteristics within different irrigation districts, insufficient water demand forecasting accuracy, and poor regional adaptability of the forecast results. Existing clustering algorithms applicable to identifying crop growth heterogeneity within irrigation districts use isolated single points as cluster centers and cannot effectively handle inter-cluster adhesion, leading to inaccurate identification results of crop growth states within the irrigation district. Existing water demand forecasting models employ a uniform time weight sharing mechanism, making it difficult to distinguish the contribution levels of features at different time steps, resulting in low accuracy of long-term water demand forecasting. Summary of the Invention
[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a method for predicting irrigation district water demand based on multi-source remote sensing data. Traditional irrigation district water demand prediction methods employ a uniform scale modeling across the entire irrigation district, failing to adequately consider the spatial heterogeneity of crop growth states within the district. This results in inaccurate representation of differences in crop water demand characteristics across different irrigation districts, insufficient prediction accuracy, and poor regional adaptability of the prediction results. This innovative approach first proposes a zoning method based on the heterogeneity of crop growth within the irrigation district, and then uses each growth zone as an independent unit to conduct dynamic water demand prediction. This transforms the traditional overall-scale irrigation district water demand prediction into a spatially zoning-scale prediction based on differences in crop growth states. This enhances the correlation between crop growth characteristics and water demand variation patterns in different regions, improves the adaptability and accuracy of the prediction model to regional water demand differences, and enhances the accuracy and spatial matching ability of future water demand prediction. It achieves dynamic water demand prediction and precise zoning of irrigation districts that considers the spatial heterogeneity of crop growth. Furthermore, this approach addresses the limitations of existing clustering algorithms for identifying crop growth heterogeneity within irrigation districts, which use isolated single points as cluster centers and cannot effectively... To address the technical problem of inaccurate crop growth status identification and insufficient spatial partitioning caused by inter-cluster adhesion, which reduces the reliability of subsequent zoning water demand prediction and the adaptability to precision irrigation, this solution innovatively adopts a high-density mutually exclusive connected component identification mechanism, a neighborhood distance adaptive calculation strategy, and an automatic boundary stripping strategy to improve the clustering algorithm. This enhances the accuracy and spatial continuity of crop growth heterogeneity identification within the irrigation area, strengthens the ability to distinguish regions with different crop growth statuses and the ability to express boundaries, and achieves adaptive and refined zoning of crop growth status within the irrigation area. Furthermore, to address the technical problem of low accuracy in long-term water demand prediction due to the difficulty in distinguishing the contribution levels of features at different time steps using a unified time weight sharing mechanism in existing water demand prediction models, this solution innovatively employs a dual-branch structure with a gated cyclic branch and a time-specific dynamic memory branch to construct a dynamic memory-enhanced temporal coding layer, and superimposes a time attention weighted layer to form a combined coding architecture. This enhances the model's adaptability to complex temporal changes, improves the accuracy of long-term irrigation area water demand prediction, and achieves dynamic water demand prediction for irrigation areas based on the spatial heterogeneity of crop growth.
[0005] The technical solution adopted by this invention is as follows: The irrigation area water demand prediction method based on multi-source remote sensing data provided by this invention includes the following steps:
[0006] Step S1: Standardize multi-source data acquisition;
[0007] Step S2: Zoning of crop growth heterogeneity within the irrigation district;
[0008] Step S3: Construct the water demand feature set for each irrigation district;
[0009] Step S4: Construct a regional water demand prediction model for irrigation areas;
[0010] Step S5: Zonal dynamic water demand forecasting.
[0011] Further, in step S1, the standardized multi-source data acquisition specifically involves obtaining raw irrigation district water demand management data through data acquisition operations, and performing data optimization operations on the raw irrigation district water demand management data to obtain standardized irrigation district water demand management data. The standardized irrigation district water demand management data includes standardized historical crop growth monitoring data within the irrigation district, standardized real-time crop growth monitoring data within the irrigation district, standardized historical irrigation district regional water demand prediction data, and standardized real-time irrigation district regional water demand prediction data. The data optimization operation specifically involves sequentially performing data cleaning and data standardization processing on the raw irrigation district water demand management data to obtain standardized irrigation district water demand management data.
[0012] Furthermore, in step S2, the heterogeneous zoning of crop growth within the irrigation area specifically includes the following steps:
[0013] Step S21: Irrigation area basic spatial unit division. Specifically, the target irrigation area is divided into spatial units according to a preset spatial scale using a regular grid division method. The standardized real-time crop growth monitoring data of each spatial unit is extracted to form a crop growth monitoring data set with spatial units as the basic objects, and the real-time crop growth monitoring data of each spatial unit is obtained.
[0014] Step S22: Construct a crop growth heterogeneity identification model within the irrigation district, which specifically includes the following steps:
[0015] Step S221: Adaptive determination of natural neighborhood feature values. Specifically, firstly, a clustered sample dataset is constructed based on standardized crop growth monitoring data within the irrigation area. Then, a KD-tree spatial index is constructed on the clustered sample dataset. Based on the KD-tree spatial index, an iterative natural neighborhood search is performed until the natural neighborhood stability convergence condition is met, thus obtaining the natural neighborhood feature values of the clustered sample dataset. ;
[0016] The natural neighborhood stable convergence condition is set with two parallel judgment rules: the full sample mutual nearest neighbor coverage rule and the non-neighbor sample number stability rule. If either rule is satisfied, it is determined that a stable convergence state has been reached and the iteration is terminated.
[0017] Step S222: Calculate the adaptive neighborhood distance value of the sample, specifically by determining the adaptive neighborhood scale based on the natural neighborhood feature values. Then, the reverse neighbor set of each sample data point is calculated based on the adaptive neighborhood scale. Finally, based on the number of sample data points in the reverse neighbor set, the adaptive neighborhood distance value of each sample data point is calculated for each scenario. The formula used is as follows:
[0018] ;
[0019] In the formula, Represents sample data points The reverse neighbor set, Represents sample data points The number of sample data points in the reverse neighbor set. Represents sample data points and The Euclidean distance between them Represents sample data points The adaptive neighborhood distance value, Represents sample data points of The Kth nearest neighbor sample data point in the nearest neighbor set, This represents the function for calculating Euclidean distance. and Let i and j be the i-th and j-th sample data points in the clustered sample dataset;
[0020] Step S223: Construction of the neighbor graph of mutual nearest neighbors. Specifically, the clustered sample dataset is used as the vertex set. An initial undirected complete graph is constructed based on the Euclidean distance between the sample data points. Then, based on the adaptive neighborhood distance value of each sample data point, adaptive pruning is performed on the initial undirected complete graph. All sample data point pairs are traversed and effective edges are filtered. A pure neighbor graph of mutual nearest neighbors is constructed based on the filtered set of mutual nearest neighbors.
[0021] The adaptive pruning rule is that the adjacent edge between two sample data points is retained only when the Euclidean distance between them is less than the minimum value of their adaptive neighborhood distance.
[0022] The formula used is as follows:
[0023] ;
[0024] In the formula, Represents sample data points A set of mutual nearest neighbors, in which all sample data points are adjacent to each other. Preserve adjacent edges. Represents sample data points The adaptive neighborhood distance value;
[0025] Step S224: High-density mutually exclusive connected component identification. Specifically, firstly, connected component analysis is performed based on the neighboring graph to obtain M initial connected subgraphs that are mutually disjoint and cover all sample data points. Then, vertex mutual exclusion partitioning is performed on all initial connected subgraphs to make the vertex set of each connected subgraph independent and non-overlapping. Then, the maximum density core data point in each connected subgraph is calculated. Finally, high-density mutually exclusive connected components are partitioned according to the high-density mutual exclusion rule.
[0026] The vertex mutual exclusion partitioning specifically means that the vertex sets of any two different connected subgraphs must satisfy the non-intersection relationship, and the vertex sets of all connected subgraphs must completely cover all sample data points;
[0027] The highest density core data point is specifically the sample data point that is most representative of the connected subgraph;
[0028] The high-density mutual exclusion rules specifically include single-core constraints for single connected domains, inter-domain mutual exclusion constraints, and intra-domain sample data point consistency constraints.
[0029] Step S225: Adaptive segmentation of the boundary region. Specifically, first, an adaptive segmentation trigger condition is set. When the number of initial connected subgraphs is less than the preset number of crop growth states, the density conflict domain adaptive segmentation process is triggered. Then, based on the adaptive neighborhood distance values of all sample data points, the global adaptive neighborhood distance mean is calculated to identify boundary sample data points. Next, all adjacent edges associated with the boundary sample data points are located. The boundary sample data points and their associated edges are temporarily stripped from the mutual nearest neighbor adjacency graph, cutting off the adhesion paths of the core regions of different growth states, completing the topological segmentation of the adhesion connected subgraphs. Connectivity component analysis is re-executed on the segmented mutual nearest neighbor adjacency graph to update the number of connected subgraphs and high-density mutually exclusive connected regions. The segmentation process is repeated until the number of connected subgraphs equals the preset number of crop growth states, at which point the segmentation process is terminated. The final core regions of the growth states are determined by sorting according to the high-density mutually exclusive connected regions. The formula used is as follows:
[0030] ;
[0031] ;
[0032] In the formula, This represents the set of identified boundary sample data points. This represents the mean of the adaptive neighborhood distances for all sample data points. Represents a clustered sample dataset;
[0033] Step S226: Sample data point reassignment. Specifically, the temporarily separated boundary sample data points are first reintegrated into the sample set to be assigned. Then, based on the high-density mutually exclusive connected components (HDMs) sorting as the core benchmark, and then based on the adaptive neighborhood scale, the global adaptive K-nearest neighbor sample set is searched for each boundary sample data point to be assigned. Sample data points in the nearest neighbor set that have been assigned to the corresponding HDMs are selected. Finally, a density-weighted adaptive selection rule is used to calculate the matching degree between the boundary samples to be assigned and each HDM, and the boundary sample data points are assigned to the HDMs with the highest matching degree. After the full assignment of all sample data points is completed, each HDM represents a growth state, thus outputting the final adaptive clustering result. The formula used is as follows:
[0034] ;
[0035] In the formula, Represents the boundary sample data points to be assigned. The degree of matching with high-density mutually exclusive connected domains, Let M represent the m-th high-density mutually exclusive connected component, where M represents the number of connected components.
[0036] Step S23: Adaptive clustering identification of crop growth status, specifically, inputting real-time crop growth monitoring data of each spatial unit and standardized historical crop growth monitoring data within the irrigation area into the crop growth heterogeneity identification model within the irrigation area, and outputting the adaptive clustering results of crop growth status within the irrigation area.
[0037] Step S24: Generate crop growth status identification results within the irrigation area. Specifically, based on the adaptive clustering results of crop growth status within the irrigation area, frequency statistics are performed on the clustering labels corresponding to all sample data points in each high-density mutually exclusive connected domain. The clustering label with the highest frequency in the connected domain is selected as the growth status label corresponding to the high-density mutually exclusive connected domain, thereby obtaining the real-time crop growth identification results within the irrigation area for each spatial unit.
[0038] Step S25: Generate crop growth heterogeneity partitioning results. Specifically, based on the real-time crop growth identification results within the irrigation area of each spatial unit, adjacent spatial units with the same crop growth status category are merged based on spatial adjacency relationships. Spatial units with continuous spatial location and consistent growth status are assigned to the same crop growth status region, forming continuously distributed regions of the same crop growth status, thus obtaining the crop growth heterogeneity partitioning results within the irrigation area.
[0039] Further, in step S3, the construction of the irrigation area water demand feature set specifically involves, based on the results of crop growth heterogeneity zoning within the irrigation area, taking each crop growth status zoning as the basic analysis object, extracting real-time irrigation area water demand prediction data within the corresponding spatial range of the zoning, and combining the corresponding crop growth status information, remote sensing crop water demand data, remote sensing environmental data, and soil environmental data within the same spatial zoning to obtain the water demand feature set of each zoning area.
[0040] Furthermore, in step S4, the construction of the irrigation area regional water demand prediction model specifically includes the following steps:
[0041] Step S41: Construct the model input layer, specifically based on the water demand feature set of the region, using a preset time step as the basic unit, and perform temporal alignment and ordered splicing of the features in the water demand feature set of the region in chronological order to obtain a continuous temporal feature sequence;
[0042] Step S42: Construct a dynamic memory-enhanced temporal coding layer. Specifically, a dual-branch structure with gated cyclic branch and dynamic memory branch running in parallel is adopted to perform time-step-by-time deep coding on the input continuous temporal feature sequence. At each time step, the calculation of gated cyclic branch and dynamic memory branch are completed synchronously. Finally, the two outputs are fused to obtain the dynamic memory-enhanced temporal feature of that time step.
[0043] The specific calculation of the gated loop branch is as follows: taking the water demand feature vector of the partition at time step t and the hidden state and cell state at time step t-1 as input, the gated outputs of the forget gate, input gate and output gate are calculated in sequence, the candidate cell state and the final cell state are updated, and the hidden state at that time step is output.
[0044] The dynamic memory branch calculation specifically involves assigning dynamic memory weights to each time step of the input sequence. The memory weights of all time steps constitute a global dynamic memory matrix. Based on the partition water requirement feature vector and the corresponding dedicated memory weight of the t-th time step, the dynamic memory state of that time step is calculated. The formula used is as follows:
[0045] ;
[0046] ;
[0047] In the formula, Let represent the hidden state at time step t. This represents the dynamic memory state at time step t. This represents the dynamic memory weight matrix at time step t. This represents the bias vector of the dynamic memory branch. This represents the dynamic memory-enhanced temporal feature of the final output at time step t;
[0048] Step S43: Construct a time attention weighted layer. Specifically, the attention score of the dynamic memory-enhanced temporal features is calculated step by step through the time attention mechanism. After normalization, the contribution weight of each time step is obtained. Finally, the dynamic memory-enhanced temporal features of all time steps are weighted and summed to obtain the water demand prediction weighted features.
[0049] Step S44: Construct a zone water demand prediction output layer, specifically by performing a linear dimension transformation through a fully connected layer to directly output the zone irrigation water demand prediction value for the corresponding future time period;
[0050] Step S45: Water demand prediction model training. Specifically, an initial irrigation area water demand prediction model is obtained by constructing a model input layer, a dynamic memory-enhanced temporal coding layer, a temporal attention weighting layer, and a regional water demand prediction output layer. The standardized historical irrigation area water demand prediction data is used as the training data for the prediction model to train the prediction model, resulting in the trained irrigation area water demand prediction model.
[0051] Furthermore, in step S5, the dynamic water demand prediction for each zone specifically involves inputting the water demand feature sets corresponding to each zone of the irrigation district into the trained irrigation zone water demand prediction model, outputting the predicted irrigation water demand values for each zone in the future time period, and finally integrating all the predicted irrigation water demand values to form a real-time irrigation zone water demand result covering the entire irrigation district. Based on the real-time irrigation zone water demand result, differentiated irrigation water allocation is performed for different spatial zones to achieve dynamic water demand prediction and precise irrigation based on the spatial heterogeneity of crop growth.
[0052] The beneficial effects achieved by the present invention using the above solution are as follows:
[0053] (1) In response to the technical problems of traditional irrigation district water demand prediction methods, which adopt a unified scale modeling of the entire irrigation district and do not fully consider the spatial heterogeneity of crop growth status within the irrigation district, resulting in the inability to accurately express the differences in crop water demand characteristics in different irrigation districts, insufficient water demand prediction accuracy, and poor regional adaptability of prediction results, this scheme innovatively proposes a zoning of crop growth heterogeneity within the irrigation district, and then uses each growth zone as an independent unit to carry out dynamic water demand prediction separately. This transforms the traditional overall scale water demand prediction of the irrigation district into a spatial zoning scale water demand prediction oriented towards differences in crop growth status, enhances the ability to express the correlation between crop growth characteristics and water demand change patterns in different regions, improves the adaptability and prediction accuracy of the prediction model to differences in water demand in different regions, improves the accuracy and spatial matching ability of future water demand prediction, and realizes dynamic water demand prediction and zoning precision irrigation of irrigation districts that considers the spatial heterogeneity of crop growth.
[0054] (2) To address the technical problems of existing clustering algorithms for identifying crop growth heterogeneity within irrigation districts, which use isolated single points as cluster centers and cannot effectively handle inter-cluster adhesion, resulting in inaccurate crop growth status identification results and insufficient spatial partitioning integrity, and consequently reducing the reliability of subsequent partitioned water demand prediction and the suitability for precise irrigation, this solution innovatively adopts a high-density mutually exclusive connected component identification mechanism, a neighborhood distance adaptive calculation strategy, and an automatic boundary stripping strategy to improve the clustering algorithm. Through the neighborhood distance adaptive calculation strategy, the appropriate neighborhood scale is automatically determined based on the spatial distribution characteristics of crop growth monitoring data, reducing the partitioning bias caused by the reliance on manual parameter settings in traditional clustering methods. By employing a high-density mutually exclusive connected component identification mechanism, the core region of crop growth status is determined jointly by spatial connectivity structure and high-density features. This avoids the unstable category division problem caused by using isolated sample points as cluster centers. Through an automatic boundary stripping strategy, the transitional spatial units between different crop growth status regions are identified and separated, reducing erroneous connections between different growth status regions. This improves the accuracy and spatial continuity of crop growth heterogeneity identification within the irrigation area, enhances the distinguishability and boundary expression ability of different crop growth status regions, and realizes adaptive and refined zoning of crop growth status within the irrigation area. This provides an accurate spatial division basis for subsequent zoning water demand prediction and precise irrigation scheduling.
[0055] (3) To address the technical problem that existing water demand prediction models employ a unified time weight sharing mechanism, making it difficult to distinguish the contribution levels of features at different time steps, and causing historical key information to gradually decay and gradient transfer ability to decline during long-term series prediction, making it difficult to effectively capture long-range water consumption dependence across the fertility period, resulting in low accuracy of long-term water demand prediction, this solution innovatively adopts a dual-branch structure with gated cyclic branches and time-specific dynamic memory branches to construct a dynamic memory-enhanced time-series coding layer, and superimposes a time attention weighting layer to form a combined coding architecture. The gated cyclic branches retain the traditional long short-term memory network's ability to connect time series data. The modeling capability for continuous variation patterns is enhanced by allocating independent dynamic memory parameters to different time steps through a time-specific dynamic memory branch. This breaks through the limitation of time step parameter sharing in traditional time series models, improves the ability to capture long-range water consumption dependencies across growth periods, and enhances the retention of key historical water demand features. By using a time attention weighting layer to adaptively allocate feature contributions from different historical time stages, the model reduces the interference of invalid time information on prediction results, enhances its adaptability to complex temporal variation patterns, improves the accuracy of long-term irrigation area water demand prediction, and realizes dynamic water demand prediction for irrigation areas based on the spatial heterogeneity of crop growth. Attached Figure Description
[0056] Figure 1 This is a flowchart illustrating the irrigation area water demand prediction method based on multi-source remote sensing data provided by the present invention.
[0057] Figure 2 This is a flowchart illustrating step S2;
[0058] Figure 3 This is a flowchart illustrating step S4;
[0059] Figure 4 This is a flowchart illustrating step S22;
[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation
[0061] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0062] In the description of this invention, it should be understood that the terms upper, lower, front, back, left, right, top, bottom, inner, and outer, etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the system or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0063] Example 1, see Figure 1 The technical solution adopted by this invention is as follows: The irrigation area water demand prediction method based on multi-source remote sensing data provided by this invention includes the following steps:
[0064] Step S1: Standardized multi-source data acquisition, specifically, obtaining standardized irrigation district water demand management data through data acquisition and data optimization operations;
[0065] Step S2: Heterogeneous zoning of crop growth within the irrigation area. Specifically, the irrigation area is first divided into basic spatial units, and real-time crop growth monitoring data of each spatial unit is obtained. Then, a clustering algorithm is improved by adopting a high-density mutually exclusive connected component identification mechanism, a neighborhood distance adaptive calculation strategy, and an automatic boundary stripping strategy to construct a heterogeneous identification model of crop growth within the irrigation area. Subsequently, based on the real-time crop growth monitoring data of each spatial unit, adaptive clustering identification of crop growth status is performed to obtain adaptive clustering results. Finally, the heterogeneous zoning results of crop growth within the irrigation area are obtained based on the adaptive clustering results.
[0066] Step S3: Construct the water requirement feature set for each irrigation district, specifically by obtaining the water requirement feature set for each irrigation district based on the results of crop growth heterogeneity partitioning within the irrigation district;
[0067] Step S4: Construct a regional water demand prediction model for irrigation areas, which specifically includes constructing a model input layer, constructing a dynamic memory enhanced temporal coding layer using a dual-branch structure with gated cyclic branches and dynamic memory branches in parallel, constructing a temporal attention weighting layer, and constructing a regional water demand prediction output layer to obtain an initial regional water demand prediction model for irrigation areas, and training the prediction model to obtain a trained regional water demand prediction model for irrigation areas.
[0068] Step S5: Regional dynamic water demand prediction. Specifically, the regional water demand feature sets corresponding to each region of the irrigation district are input into the trained regional water demand prediction model of the irrigation district, and the predicted irrigation water demand values of each region for the corresponding future time period are output. The predicted irrigation water demand values of each region are then integrated to obtain the real-time regional water demand results of the irrigation district.
[0069] By performing the above operations, this solution addresses the technical problems of traditional irrigation district water demand forecasting methods, which use a uniform scale modeling across the entire irrigation district and fail to fully consider the spatial heterogeneity of crop growth states within the irrigation district. This results in inaccurate expression of differences in crop water demand characteristics within different irrigation districts, insufficient water demand forecasting accuracy, and poor regional adaptability of the forecast results. This innovative approach first proposes a zoning of crop growth heterogeneity within the irrigation district, and then uses each growth zone as an independent unit to conduct dynamic water demand forecasting. This transforms the traditional overall irrigation district-scale water demand forecasting into a spatially zoning-scale water demand forecasting oriented towards differences in crop growth states. This enhances the ability to express the correlation between crop growth characteristics and water demand variation patterns in different regions, improves the adaptability and forecasting accuracy of the prediction model to differences in water demand across different regions, and enhances the accuracy and spatial matching ability of future water demand forecasting. Ultimately, this solution achieves dynamic water demand forecasting and precise zoning of irrigation districts that considers the spatial heterogeneity of crop growth.
[0070] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, the standardized multi-source data acquisition specifically involves data collection through the irrigation district management platform to obtain the original irrigation district water demand management data, and data optimization operation on the original irrigation district water demand management data to obtain standardized irrigation district water demand management data.
[0071] The standardized irrigation district water demand management data includes historical standardized data on crop growth monitoring within the irrigation district, real-time standardized data on crop growth monitoring within the irrigation district, historical standardized data on water demand forecasting for different irrigation districts, and real-time standardized data on water demand forecasting for different irrigation districts.
[0072] The raw data for irrigation district water demand management includes historical crop growth monitoring data within the irrigation district, real-time crop growth monitoring data within the irrigation district, historical water demand forecast data for irrigation districts, and real-time water demand forecast data for irrigation districts.
[0073] Both the historical crop growth monitoring data and the real-time crop growth monitoring data within the irrigation district include basic crop information data and remote sensing monitoring data.
[0074] The historical crop growth monitoring data within the irrigation district also includes the identification results of crop growth within the historical irrigation district.
[0075] The crop growth identification results within the irrigation area are used to represent the crop growth and development level in different spatial areas within the irrigation area during historical periods. Specifically, based on a 3m×3m basic spatial unit as the basic evaluation object, attribute labels are assigned to the crop growth status corresponding to each spatial unit through manual annotation. These labels are respectively vigorous growth status, normal growth status, water stress growth status, and low growth status. These labels are used as reference labels for clustering operations and are only used for cluster label selection, without participating in the clustering data processing process.
[0076] The basic crop information data of the irrigation area includes crop planting type, crop height, crop coverage and crop planting time;
[0077] The remote sensing monitoring data includes optical remote sensing data, thermal infrared remote sensing data, and microwave remote sensing data;
[0078] The optical remote sensing data includes NDVI normalized vegetation index, EVI enhanced vegetation index, NDWI normalized water index, and LAI leaf area index.
[0079] The thermal infrared remote sensing data includes land surface temperature and crop canopy temperature;
[0080] The microwave remote sensing data includes soil moisture and soil dielectric parameters;
[0081] Both the historical irrigation area regional water demand prediction data and the real-time irrigation area regional water demand prediction data include remote sensing crop water demand data, remote sensing environmental data, and soil environmental data.
[0082] The remote sensing crop water requirement data is obtained by fusing optical remote sensing data, thermal infrared remote sensing data, and microwave remote sensing data, including crop evapotranspiration and actual crop water consumption.
[0083] The remote sensing environmental data is obtained through thermal infrared remote sensing data and optical remote sensing images, including rainfall, air temperature, air humidity, surface temperature, soil moisture, and crop cover.
[0084] The soil environmental data includes soil moisture content, soil type, soil texture, and soil water retention capacity;
[0085] The historical irrigation area water demand prediction data also includes remote sensing spatial zoning data and historical irrigation area water demand results. The irrigation area water demand results are used to represent the actual water demand corresponding to different spatial zoning in historical periods. They are obtained by manually labeling the water demand of different spatial zoning by combining historical irrigation records, agricultural production management data and expert experience. They are used as target label data in the subsequent training process of the zoning water demand prediction model. Specifically, they are the water demand corresponding to each spatial zoning, in liters.
[0086] The remote sensing spatial zoning data is obtained by extracting spatial information, georeferencing, and dividing spatial units from multi-source remote sensing images, including spatial zoning number, zoning spatial range, zoning area information, and crop growth status corresponding to the zoning.
[0087] The data optimization operation is used to eliminate noise, abnormal information, redundant information, and scale differences between different data sources in the original data of irrigation district water demand management, and to improve the consistency and effectiveness of multi-source heterogeneous data. Specifically, data cleaning and data standardization are performed on the original data of irrigation district water demand management in sequence to obtain standardized irrigation district water demand management data.
[0088] The data cleaning process is used to remove invalid, redundant, and erroneous information from the original data, fill in missing data, eliminate data noise, and ensure the integrity, validity, and logical consistency of the dataset. Specifically, the original data of irrigation district water demand management is subjected to missing value imputation, outlier removal, and redundant data removal operations in sequence. The missing value imputation is specifically to fill in missing values using the mean imputation method. The outlier removal is specifically to detect and remove extreme values in the original data using the interquartile range method.
[0089] The redundant data removal operation specifically involves detecting duplicate records in the original data of irrigation area water demand management and deleting duplicate data records to reduce data redundancy and improve data processing efficiency.
[0090] The data standardization process is used to eliminate the differences in dimensionality and numerical scale between different fields in the cleaned data, and to convert multi-source heterogeneous data into a standardized dataset with a unified format and scale, thereby improving the stability and computational accuracy in subsequent cluster analysis and prediction model training. Specifically, the label encoding method is first used to map the category fields in the original data to corresponding integer values according to the category values, and then the Z-Score standardization method is used to standardize all continuous variables so that each continuous variable meets a unified numerical scale range.
[0091] Example 3, see Figure 1 , Figure 2 and Figure 4 This embodiment is based on the above embodiment. In step S2, the crop growth heterogeneity partitioning within the irrigation area is used to identify and classify the crop growth status in different spatial regions within the irrigation area; specifically, it includes the following steps:
[0092] Step S21: Irrigation area basic spatial unit division, which is used to convert large-scale irrigation area into basic spatial units that can be used for crop growth status monitoring; specifically, a regular grid division method is used to divide the target irrigation area into spatial units according to a preset spatial scale, and the real-time crop growth monitoring standardized data of each spatial unit is extracted to form a crop growth monitoring data set with spatial units as the basic objects, and the real-time crop growth monitoring data of each spatial unit is obtained.
[0093] The preferred spatial scale is 3m×3m. A 3m×3m regular grid is used as the basic spatial analysis unit of the irrigation area. The target irrigation area is divided into multiple continuous spatial units, so that each spatial unit corresponds to a crop group within a certain range, rather than a single crop individual.
[0094] Step S22: Construct a crop growth heterogeneity identification model within the irrigation area. This model is used to analyze the similarity of crop growth states between different spatial units, enabling automatic identification of heterogeneous regions within the irrigation area. Specifically, it involves improving the clustering algorithm through a high-density mutually exclusive connected component identification mechanism, an adaptive neighborhood distance calculation strategy, and an automatic boundary stripping strategy, thereby constructing the crop growth heterogeneity identification model within the irrigation area. This includes the following steps:
[0095] Step S221: Adaptive determination of natural neighborhood feature values is used to automatically determine the optimal neighborhood scale that fits the data distribution, eliminating the dependence of traditional clustering methods on the manual setting of the nearest neighbor parameter K and avoiding hierarchical bias caused by manual parameter setting. Specifically, firstly, a clustering sample dataset is constructed based on standardized crop growth monitoring data within the irrigation area. Then, a KD-tree spatial index is constructed on the clustering sample dataset. Based on the KD-tree spatial index, an iterative natural neighborhood search is performed until the natural neighborhood stability convergence condition is met, thus obtaining the natural neighborhood feature values of the clustering sample dataset. ;
[0096] The clustering sample dataset is composed of a set of real-time monitoring sample data points and a set of historical monitoring sample data points.
[0097] The set of real-time monitoring sample data points is composed of standardized real-time crop growth monitoring data within the irrigation area at the current time phase, with each spatial unit's real-time crop growth monitoring data corresponding to a real-time monitoring sample data point.
[0098] The set of historical monitoring sample data points consists of standardized crop growth monitoring data within historical irrigation districts, obtained by dividing the data according to the same spatial scale.
[0099] Each sample data point includes 11 feature dimensions: crop planting type, crop coverage, crop planting time, NDVI normalized vegetation index, EVI enhanced vegetation index, NDWI normalized water index, LAI leaf area index, surface temperature, crop canopy temperature, soil moisture, and soil dielectric parameters.
[0100] The construction of the KD tree spatial index is used to hierarchically divide the multidimensional feature space of the monitored samples and reduce the time complexity of the nearest neighbor search. Specifically, the KD tree spatial index is constructed based on the multidimensional feature vector of each sample data point in the clustered sample dataset as spatial coordinates.
[0101] The iterative natural neighborhood search specifically involves first initializing the iteration parameters, setting the initial number of nearest neighbors... Set the value to 1, and simultaneously create a reverse neighbor counting array. The length of this array is consistent with the total number of sample data points in the clustering sample dataset, and the initial value of all elements in the array is set to 0. Then, based on the KD-tree spatial index, quickly search to obtain the t-nearest neighbor set corresponding to each sample data point, thus obtaining the t-nearest neighbor set of the i-th sample data point. ; Traverse the t-nearest neighbor set of all sample data points, if the sample data points Belongs to sample data points t nearest neighbor set Then the element at the corresponding position in the reverse neighbor count array will be... Increment the value by 1. After completing a single round of count updates, determine whether the current iteration satisfies the natural neighborhood stable convergence condition. If the current iteration does not satisfy the natural neighborhood stable convergence condition, increment the number of nearest neighbors t by 1 and reset the reverse neighbor count array. If all values are 0, proceed to the next iteration. If the natural neighborhood stability convergence condition is met, terminate the iteration and output the nearest neighbor number t of the current iteration as the feature value of the natural neighborhood; where, This represents the total number of sample data points that include the j-th sample data point in its own nearest neighbor set, given the current nearest neighbor number t.
[0102] The natural neighborhood stability convergence condition is set with two parallel decision rules: a full-sample mutual nearest neighbor coverage rule and a non-neighbor sample number stability rule. Satisfying either rule indicates that a stable convergence state has been reached, and the iteration terminates. The formula used is as follows:
[0103] ;
[0104] ;
[0105] In the formula, Represents a clustered sample dataset. and Let i and j be the i-th and j-th sample data points in the clustered sample dataset, respectively. This represents the number of isolated samples with a value of 0 in the reverse neighbor count array at the current nearest neighbor number t. Represents the number of nearest neighbors in the previous iteration. Below, the number of isolated samples with a value of 0 in the reverse neighbor count array;
[0106] Step S222: Calculate the adaptive neighborhood distance value of the samples to effectively solve the core problems of traditional clustering algorithms, which rely on manually set cutoff distances and nearest neighbor numbers, resulting in partitioning results being greatly affected by subjective parameters and poor adaptability to different regional datasets; specifically, determine the adaptive neighborhood scale based on natural neighborhood feature values. Then, the reverse neighbor set of each sample data point is calculated based on the adaptive neighborhood scale. Finally, based on the number of sample data points in the reverse neighbor set, the adaptive neighborhood distance value of each sample data point is calculated for each scenario. The formula used is as follows:
[0107] ;
[0108] ;
[0109] ;
[0110] In the formula, Represents sample data points The set of 2x K nearest neighbors, i.e., the search results obtained based on the adaptive neighborhood scale K, are... The nearest 2K sample data points in Euclidean distance. Represents sample data points The reverse neighbor set refers to all sets of adjacent pairs. The sample data points are included in a set of twice their own adaptive K-nearest neighbors. Represents sample data points The number of sample data points in the reverse neighbor set. Represents sample data points of Nearest neighbor set, i.e., with The nearest K sample data points are Euclidean distances. Represents sample data points and The Euclidean distance between them Represents sample data points The adaptive neighborhood distance value, Represents sample data points of The Kth nearest neighbor sample data point in the nearest neighbor set, This represents the function for calculating Euclidean distance. Represents sample data points The Euclidean distance to the farthest sample data point in its reverse neighbor set;
[0111] Step S223: Construct a neighbor graph to preserve the adjacency relationships between spatial units with high feature similarity. This effectively solves the problems of blurred sample boundaries in the transition zone and incorrect connectivity between samples of different levels, ensuring that samples within each connected subgraph in the adjacency graph have a highly consistent state, thus avoiding sample confusion from different partitioning results at the topological level. Specifically, the clustered sample dataset is used as the vertex set, and an initial undirected complete graph is constructed based on the Euclidean distance between sample data points. Then, based on the adaptive neighborhood distance values of each sample data point, adaptive pruning is performed on the initial undirected complete graph. All sample data point pairs are traversed and valid edges are filtered. A clean neighbor graph is constructed based on the filtered set of nearest neighbors. This represents the vertex set, i.e., the clustering sample dataset. Let represent the edge set, which consists of undirected edges between all pairs of sample data points;
[0112] The adaptive pruning rule is that the adjacent edge between two sample data points is retained only when the Euclidean distance between them is less than the minimum value of their adaptive neighborhood distance.
[0113] The formula used is as follows:
[0114] ;
[0115] In the formula, Represents sample data points A set of mutual nearest neighbors, in which all sample data points are adjacent to each other. Preserve adjacent edges. Represents sample data points The adaptive neighborhood distance value;
[0116] Step S224: High-density mutually exclusive connected component identification, used to achieve automatic partitioning of the core region. One connected component corresponds to a standardized growth state. Specifically, firstly, connected component analysis is performed based on the nearest neighbor adjacency graph to obtain M disjoint initial connected subgraphs that cover all sample data points. Then, vertex mutual exclusion partitioning is performed on all initial connected subgraphs to ensure that the vertex sets of each connected subgraph are independent and non-overlapping. Next, the maximum density core data points within each connected subgraph are calculated. Finally, high-density mutually exclusive connected components are partitioned according to the high-density mutual exclusion rule. The connected subgraphs... , Indicates the first The vertex set of a connected subgraph, that is, the set of all sample data points contained in the connected subgraph. Indicates the first The edge set of a connected subgraph, that is, the set of effective nearest neighbor edges between sample data points within the connected subgraph;
[0117] The connected component analysis is specifically based on the nearest neighbor adjacency graph. It uses the connected component analysis algorithm in graph theory to traverse all vertices and edges in the graph and automatically identify and separate all interconnected topological substructures.
[0118] The vertex mutual exclusion partitioning specifically means that the vertex sets of any two different connected subgraphs must satisfy the non-intersection relationship, and the vertex sets of all connected subgraphs must completely cover all sample data points;
[0119] The maximum density core data points Specifically, the most representative sample data points of the connected subgraph are used as the core benchmark for dividing the growth state; the formula used is as follows:
[0120] ;
[0121] The high-density mutual exclusion rules are used to ensure that each partition level benchmark is unique, the boundaries between domains are clear, and the features within a domain are consistent. All three constraints must be satisfied simultaneously. Specifically, they include single core constraint for a single connected domain, mutual exclusion constraint between domains, and consistency constraint for sample data points within a domain.
[0122] The single-core constraint of a single connected domain specifically allows only one maximum density core point to be selected within each high-density mutually exclusive connected domain as the benchmark sample data point for that growth state.
[0123] Specifically, the inter-domain mutual exclusion constraint means that the maximum density core point of different connected subgraphs must correspond uniquely, and multiple connected domains must not share the same core point, so as to ensure that the inter-domain structures are independent.
[0124] The consistency constraint of sample data points within the domain specifically means that all sample data points within the same high-density mutually exclusive connected domain share the same crop growth status label.
[0125] Step S225: Adaptive boundary region segmentation is used to address the issue of insufficient connected components and missed growth state classifications in the initial connected subgraph due to sample adhesion in the transition region of growth states. This is achieved by adaptively separating and segmenting the adhered connected components at the boundary, ensuring the integrity and accuracy of the growth state classification. Specifically, an adaptive segmentation trigger condition is first set. When the number of initial connected subgraphs is less than the preset number of crop growth states, the density conflict domain adaptive segmentation process is triggered. Then, based on the adaptive neighborhood distance values of all sample data points, the global adaptive neighborhood distance mean is calculated to identify boundary sample data points. Next, all adjacent edges associated with the boundary sample data points are located, and the boundary sample data points and their associated edges are temporarily separated from the nearest neighbor adjacency graph. This cuts off the adhesion paths in the core regions of different growth states, completing the topological segmentation of the adhered connected subgraph. Connectivity component analysis is re-executed on the segmented nearest neighbor adjacency graph to update the number of connected subgraphs and high-density mutually exclusive connected components. The segmentation process is repeated until the number of connected subgraphs equals the preset number of crop growth states, at which point the segmentation process terminates. The final core regions of growth states are determined by sorting according to the high-density mutually exclusive connected components. The formula used is as follows:
[0126] The sorting of high-density mutually exclusive connected components by size specifically refers to sorting them in descending order according to the number of sample data points contained in each high-density mutually exclusive connected component after the segmentation is completed.
[0127] The number of preset crop growth states is determined according to the needs of crop growth state categories in the irrigation area; in this embodiment, it is set to 4.
[0128] The high-density mutually exclusive connected domain refers to the final hierarchical unit that uniquely corresponds to a standardized growth state after the connected subgraph is used as the topological basis carrier, and after full constraint verification of high-density mutually exclusive rules and adaptive segmentation of connected domains.
[0129] ;
[0130] ;
[0131] In the formula, This represents the set of identified boundary sample data points. Sample data points whose adaptive neighborhood distance value is greater than the global adaptive neighborhood distance mean are considered boundary sample data points. The mean of the adaptive neighborhood distances of all sample data points is used as the adaptive threshold for determining boundary sample data points. This represents the set of all edges associated with the low-density boundary sample data points. This represents an undirected adjacent edge between vertices u and v, corresponding to the nearest neighbor association edge between the two samples. If either of the two endpoints u or v belongs to the boundary sample set B, the edge will be classified as a boundary-related edge and added to the set of edges to be removed. ;
[0132] Step S226: Sample data point reassignment, used to complete the assignment of growth state labels for boundary samples; specifically, firstly, the temporarily stripped boundary sample data points are re-included into the sample set to be assigned, then the high-density mutually exclusive connected components are sorted as the core benchmark, and then, based on the adaptive neighborhood scale, the global adaptive K-nearest neighbor sample set of each boundary sample data point to be assigned is searched, and the sample data points in the nearest neighbor set that have been assigned to the corresponding high-density mutually exclusive connected components are selected. Finally, the density-weighted adaptive selection rule is used to calculate the matching degree between the boundary samples to be assigned and each high-density mutually exclusive connected component, and the boundary sample data points are assigned to the high-density mutually exclusive connected components with the highest matching degree. After the full assignment of all sample data points is completed, each high-density mutually exclusive connected component is formed, and each high-density mutually exclusive connected component represents a growth state, thereby outputting the final adaptive clustering result;
[0133] The formula used is as follows:
[0134] ;
[0135] In the formula, Represents the boundary sample data points to be assigned. The degree of matching with high-density mutually exclusive connected domains, Let M represent the m-th high-density mutually exclusive connected component, where M represents the number of connected components.
[0136] Step S23: Adaptive clustering identification of crop growth status, used to realize automatic clustering and classification of crop growth status within the irrigation area based on the multi-source crop growth monitoring characteristics corresponding to different spatial units within the irrigation area; specifically, the real-time crop growth monitoring data of each spatial unit and the standardized historical crop growth monitoring data within the irrigation area are input into the crop growth heterogeneity identification model within the irrigation area, and the adaptive clustering results of crop growth status within the irrigation area are output.
[0137] Step S24: Generate crop growth status identification results within the irrigation area. Specifically, based on the adaptive clustering results of crop growth status within the irrigation area, frequency statistics are performed on the clustering labels corresponding to all sample data points in each high-density mutually exclusive connected domain. The clustering label with the highest frequency in the connected domain is selected as the growth status label corresponding to the high-density mutually exclusive connected domain, thereby obtaining the real-time crop growth identification results within the irrigation area for each spatial unit.
[0138] Step S25: Generate crop growth heterogeneity partitioning results. This is used to spatially merge spatial units with the same growth status characteristics based on the crop growth status identification results, forming spatial partitioning results corresponding to different crop growth statuses within the irrigation area. Specifically, based on the real-time crop growth identification results within the irrigation area for each spatial unit, adjacent spatial units with the same crop growth status category are merged into connected regions based on spatial adjacency relationships. Spatial units with continuous spatial locations and consistent growth statuses are assigned to the same crop growth status region, forming continuously distributed regions of the same crop growth status, thus obtaining the crop growth heterogeneity partitioning results within the irrigation area.
[0139] The spatial adjacency relationship is determined based on the spatial positional relationship between basic spatial units. When two spatial units have a common boundary, it is determined that there is a spatial adjacency relationship between the two spatial units.
[0140] The results of crop growth heterogeneity zoning within the irrigation area include spatial zoning number, spatial range of zoning, zoning area information, and crop growth status corresponding to the zoning.
[0141] The spatial range of the partition is determined by the spatial boundaries of the corresponding basic spatial units; the area information of the partition is calculated by the number of basic spatial units and the spatial scale of the units.
[0142] By performing the above operations, this solution addresses the technical problems of existing clustering algorithms for identifying crop growth heterogeneity within irrigation districts. These algorithms use isolated single points as cluster centers and cannot effectively handle inter-cluster adhesion, leading to inaccurate crop growth status identification results and insufficient spatial partitioning integrity. This, in turn, reduces the reliability of subsequent zoning water demand prediction and the suitability for precise irrigation. This solution innovatively improves the clustering algorithm by employing a high-density mutually exclusive connected component identification mechanism, an adaptive neighborhood distance calculation strategy, and an automatic boundary stripping strategy. The adaptive neighborhood distance calculation strategy automatically determines the appropriate neighborhood scale based on the spatial distribution characteristics of crop growth monitoring data, reducing the reliance on manual parameter settings in traditional clustering methods. The region bias is addressed by using a high-density mutually exclusive connected domain identification mechanism. This mechanism uses spatial connectivity and high-density features to jointly determine the core region of crop growth status, avoiding the instability of category division caused by using isolated sample points as cluster centers. Through an automatic boundary stripping strategy, the transitional spatial units between different crop growth status regions are identified and separated, reducing erroneous connections between different growth status regions. This improves the accuracy and spatial continuity of crop growth heterogeneity identification within the irrigation area, enhances the distinguishability and boundary representation of different crop growth status regions, and realizes adaptive and refined zoning of crop growth status within the irrigation area. This provides an accurate spatial division basis for subsequent zoning water demand prediction and precise irrigation scheduling.
[0143] Example 4, see Figure 1 This embodiment is based on the above embodiment. In step S3, the construction of irrigation area water demand feature set is used to construct a water demand feature set representing the water demand variation law of different crop growth states, and to provide input data for the subsequent irrigation area water demand prediction model. Specifically, based on the crop growth heterogeneity zoning results within the irrigation area, each crop growth state zoning is used as the basic analysis object. Real-time irrigation area water demand prediction data within the corresponding zoning spatial range is extracted, and the corresponding crop growth state information, remote sensing crop water demand data, remote sensing environmental data and soil environmental data within the same spatial zoning are combined to obtain the water demand feature set of each zoning area of the irrigation area.
[0144] The irrigation area water demand feature set includes the spatial range of the area, the area information of the area, and the crop growth status, crop evapotranspiration, actual crop water consumption, rainfall, air temperature, air humidity, surface temperature, soil moisture, crop cover, soil moisture content, soil type, soil texture, and soil water storage capacity corresponding to the area.
[0145] Example 5, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In step S4, the construction of the irrigation area regional water demand prediction model specifically includes the following steps:
[0146] Step S41: Construct the model input layer to convert the static partition water demand feature set of a single time step dimension into standardized time series data that conforms to the input specification of the dynamic memory-enhanced LSTM temporal coding layer; specifically, based on the partition water demand feature set, with a preset time step as the basic unit, perform temporal alignment and ordered concatenation of the features in the partition water demand feature set in chronological order to obtain a continuous temporal feature sequence.
[0147] The preferred time step is 24 hours, which meets the operational needs of daily irrigation scheduling in the irrigation district and matches the data sampling cycle of conventional meteorological stations and soil moisture monitoring stations.
[0148] The time-series alignment specifically involves using a preset time step as the base time unit, and processing dynamic time-series features and spatial static attributes separately for the water demand characteristics within the corresponding crop growth zone. For features that dynamically change over time, such as crop growth status, crop evapotranspiration, actual crop water consumption, rainfall, air temperature, air humidity, surface temperature, soil moisture, crop cover, and soil moisture content, a unified time base and time granularity alignment are performed to eliminate sampling time deviations from different data sources. For spatial static background attributes that do not change over time, such as soil type, soil texture, and soil water storage capacity, the dimensions are expanded according to the time series length, so that the static attributes correspond one-to-one with the dynamic features of each time step. Ultimately, this ensures that each time step corresponds to a set of zone water demand feature vectors with complete dimensions and consistent time bases.
[0149] The ordered splicing specifically involves sequentially splicing the aligned single-step partition water demand feature vectors in chronological order to form a continuous and uninterrupted temporal feature sequence.
[0150] Step S42: Construct a dynamic memory-enhanced temporal coding layer to perform deep encoding of regional water demand temporal features, overcome the gradient vanishing defect of the standard long short-term memory neural network under long-term water demand sequences, and enhance the ability to capture long-range water consumption dependencies across the reproductive period. Specifically, a dual-branch structure with gated recurrent branch and dynamic memory branch running in parallel is adopted to perform time-step deep encoding of the input continuous temporal feature sequence. At each time step, the gated recurrent branch calculation and dynamic memory branch calculation are completed simultaneously. Finally, the two outputs are fused to obtain the dynamic memory-enhanced temporal features of that time step.
[0151] The specific calculation of the gated loop branch is as follows: taking the water demand feature vector of the partition at time step t and the hidden state and cell state at time step t-1 as input, the gated outputs of the forget gate, input gate and output gate are calculated in sequence, the candidate cell state and the final cell state are updated, and the hidden state at that time step is output.
[0152] The dynamic memory branch calculation specifically involves assigning dynamic memory weights to each time step of the input sequence. The memory weights of all time steps constitute a global dynamic memory matrix. Based on the partition water requirement feature vector and the corresponding dedicated memory weight of the t-th time step, the dynamic memory state of that time step is calculated. The formula used is as follows:
[0153] ;
[0154] ;
[0155] ;
[0156] ;
[0157] ;
[0158] ;
[0159] ;
[0160] ;
[0161] In the formula, This represents the input gate value at time step t. This represents the forgetting gate value at time step t. This represents the output gate value at time step t. This represents the state value of the candidate cell at time step t. This represents the cell unit state value at time step t. This indicates the state of the cell unit at the previous moment. This represents the environmental smoothing factor at time step t. This represents the water demand feature vector of the partition corresponding to the t-th time step. , , and These represent the weight matrices from the input data to the input gate, forget gate, output gate, and candidate unit, respectively, used to map the input data to the corresponding gated units. , , and These represent the weight matrices from the hidden state to the input gate, forget gate, output gate, and candidate unit, respectively, used to incorporate historical information into the calculation of the current gated unit. , , and These represent the bias vectors for the input gate, forget gate, output gate, and candidate unit, respectively. This indicates element-wise multiplication. Let represent the hidden state at time step t. Indicates the first The hidden state at each time step. This represents the dynamic memory state at time step t, and is the output of the dynamic memory branch. This represents the dynamic memory weight matrix at time step t, taken from the t-th row of the global dynamic memory matrix. This represents the bias vector of the dynamic memory branch. This represents the dynamic memory-enhanced temporal feature of the final output at time step t;
[0162] Step S43: Construct a time attention weighting layer to adaptively allocate feature weights for each historical time step, focusing on periods that significantly affect water demand, such as critical crop growth periods, high-temperature periods, and rainfall events, while weakening noise interference from irrelevant periods and improving the accuracy of water demand prediction. Specifically, the time attention mechanism is used to calculate the attention score of dynamic memory-enhanced time-series features step by step, and the contribution weight of each time step is obtained after normalization. Finally, the dynamic memory-enhanced time-series features of all time steps are weighted and summed to obtain the weighted features for water demand prediction.
[0163] Step S44: Construct a partitioned water demand prediction output layer to map high-dimensional abstract features into partitioned irrigation water demand values with actual physical meaning, and output the partitioned water demand results for the target prediction period. Specifically, a linear dimensional transformation is performed through a fully connected layer to directly output the partitioned irrigation water demand prediction value for the corresponding future period.
[0164] Step S45: Water demand prediction model training, specifically by constructing a model input layer, a dynamic memory-enhanced temporal coding layer, a temporal attention weighting layer, and a regional water demand prediction output layer to obtain an initial irrigation area regional water demand prediction model. The standardized historical irrigation area regional water demand prediction data is used as the training data for the prediction model to train the prediction model and obtain the trained irrigation area regional water demand prediction model.
[0165] The training of the prediction model specifically involves using mean squared error as the loss function to optimize the gap between the model's prediction results and the true values. The prediction model parameters are then jointly optimized using a backpropagation algorithm. The loss function value is continuously monitored and optimized until it converges, at which point the model training is stopped, resulting in the trained irrigation district water demand prediction model.
[0166] Example 6, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In step S5, the dynamic water demand prediction of the zoning is used to dynamically extrapolate the future water demand of different crop growth states in different zoning areas, and output the spatiotemporally matched zoning water demand prediction results. This solves the problems of traditional unified prediction of the whole irrigation area not considering crop growth heterogeneity and insufficient prediction accuracy and regional adaptability. Specifically, based on the water demand feature set of each zoning area, the corresponding zoning water demand feature set of each zoning area is input into the trained irrigation zoning water demand prediction model, and the predicted irrigation water demand value of each zoning area for the corresponding future time period is output. Finally, all the predicted irrigation water demand values of the zoning areas are integrated to form the real-time irrigation zoning water demand result covering the whole irrigation area. Based on the real-time irrigation zoning water demand result, the irrigation water volume is allocated to different spatial zoning areas in a differentiated manner, so as to realize the dynamic water demand prediction of the irrigation area and precise zoning based on the spatial heterogeneity of crop growth.
[0167] By performing the above operations, this solution addresses the technical problems of low accuracy in long-term water demand prediction due to the difficulty in distinguishing the contribution levels of features at different time steps, the gradual decay of historical key information and the decline in gradient transfer ability during long-term series prediction, which makes it difficult to effectively capture long-range water consumption dependencies across the fertility period. This solution innovatively employs a dual-branch structure with a gated recurrent branch and a time-specific dynamic memory branch to construct a dynamic memory-enhanced time-series coding layer, and superimposes a time attention weighting layer to form a combined coding architecture. The gated recurrent branch preserves the time-specific characteristics of traditional long short-term memory networks. The modeling capability of continuous variation patterns is enhanced by allocating independent dynamic memory parameters to different time steps through a time-specific dynamic memory branch. This breaks through the limitation of time step parameter sharing in traditional time series models, improves the ability to capture long-range water consumption dependence across growth periods, and enhances the retention of key historical water demand features. By adaptively allocating feature contributions from different historical time stages through a time attention weighting layer, the interference of invalid time information on prediction results is reduced, enhancing the model's adaptability to complex temporal variation patterns and improving the accuracy of long-term irrigation area water demand prediction. This enables dynamic water demand prediction for irrigation areas based on the spatial heterogeneity of crop growth.
[0168] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "include," "contain," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.
[0169] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.
[0170] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.
Claims
1. A method for predicting water demand in irrigation districts based on multi-source remote sensing data, characterized in that: The method includes the following steps: Step S1: Standardized multi-source data acquisition, specifically, obtaining standardized irrigation district water demand management data through data acquisition and data optimization operations; Step S2: Heterogeneous zoning of crop growth within the irrigation area. Specifically, the irrigation area is first divided into basic spatial units, and real-time crop growth monitoring data of each spatial unit is obtained. Then, a clustering algorithm is improved by adopting a high-density mutually exclusive connected component identification mechanism, a neighborhood distance adaptive calculation strategy, and an automatic boundary stripping strategy to construct a heterogeneous identification model of crop growth within the irrigation area. Subsequently, based on the real-time crop growth monitoring data of each spatial unit, adaptive clustering identification of crop growth status is performed to obtain adaptive clustering results. Finally, the heterogeneous zoning results of crop growth within the irrigation area are obtained based on the adaptive clustering results. Step S3: Construct the water requirement feature set for each irrigation district, specifically by obtaining the water requirement feature set for each irrigation district based on the results of crop growth heterogeneity partitioning within the irrigation district; Step S4: Construct a regional water demand prediction model for irrigation areas, which specifically includes constructing a model input layer, constructing a dynamic memory enhanced temporal coding layer using a dual-branch structure with gated cyclic branches and dynamic memory branches in parallel, constructing a temporal attention weighting layer, and constructing a regional water demand prediction output layer to obtain an initial regional water demand prediction model for irrigation areas, and training the prediction model to obtain a trained regional water demand prediction model for irrigation areas. Step S5: Regional dynamic water demand prediction. Specifically, the regional water demand feature sets corresponding to each region of the irrigation district are input into the trained regional water demand prediction model of the irrigation district, and the predicted irrigation water demand values of each region for the corresponding future time period are output. The predicted irrigation water demand values of each region are then integrated to obtain the real-time regional water demand results of the irrigation district.
2. The irrigation district water demand prediction method based on multi-source remote sensing data according to claim 1, characterized in that: In step S2, the heterogeneous zoning of crop growth within the irrigation area specifically includes the following steps: Step S21: Irrigation area basic spatial unit division. Specifically, the target irrigation area is divided into spatial units according to a preset spatial scale using a regular grid division method. The standardized real-time crop growth monitoring data of each spatial unit is extracted to form a crop growth monitoring data set with spatial units as the basic objects, and the real-time crop growth monitoring data of each spatial unit is obtained. Step S22: Construct a crop growth heterogeneity identification model within the irrigation area; Step S23: Adaptive clustering identification of crop growth status, specifically, inputting real-time crop growth monitoring data of each spatial unit and standardized historical crop growth monitoring data within the irrigation area into the crop growth heterogeneity identification model within the irrigation area, and outputting the adaptive clustering results of crop growth status within the irrigation area. Step S24: Generate crop growth status identification results within the irrigation area. Specifically, based on the adaptive clustering results of crop growth status within the irrigation area, frequency statistics are performed on the clustering labels corresponding to all sample data points in each high-density mutually exclusive connected domain. The clustering label with the highest frequency in the connected domain is selected as the growth status label corresponding to the high-density mutually exclusive connected domain, thereby obtaining the real-time crop growth identification results within the irrigation area for each spatial unit. Step S25: Generate crop growth heterogeneity partitioning results. Specifically, based on the real-time crop growth identification results within the irrigation area of each spatial unit, adjacent spatial units with the same crop growth status category are merged based on spatial adjacency relationships. Spatial units with continuous spatial location and consistent growth status are assigned to the same crop growth status region, forming continuously distributed regions of the same crop growth status, thus obtaining the crop growth heterogeneity partitioning results within the irrigation area.
3. The irrigation district water demand prediction method based on multi-source remote sensing data according to claim 2, characterized in that: In step S22, the construction of the crop growth heterogeneity identification model within the irrigation area specifically includes the following steps: Step S221: Adaptive determination of natural neighborhood feature values. Specifically, firstly, a clustered sample dataset is constructed based on standardized crop growth monitoring data within the irrigation area. Then, a KD-tree spatial index is constructed on the clustered sample dataset. Based on the KD-tree spatial index, an iterative natural neighborhood search is performed until the natural neighborhood stability convergence condition is met, thus obtaining the natural neighborhood feature values of the clustered sample dataset. ; The natural neighborhood stable convergence condition is set with two parallel judgment rules: the full sample mutual nearest neighbor coverage rule and the non-neighbor sample number stability rule. If either rule is satisfied, it is determined that a stable convergence state has been reached and the iteration is terminated. Step S222: Calculate the adaptive neighborhood distance value of the sample, specifically by determining the adaptive neighborhood scale based on the natural neighborhood feature values. Then, the reverse neighbor set of each sample data point is calculated based on the adaptive neighborhood scale. Finally, based on the number of sample data points in the reverse neighbor set, the adaptive neighborhood distance value of each sample data point is calculated for each scenario. The formula used is as follows: ; In the formula, Represents sample data points The reverse neighbor set, Represents sample data points The number of sample data points in the reverse neighbor set. Represents sample data points and The Euclidean distance between them Represents sample data points The adaptive neighborhood distance value, Represents sample data points of The Kth nearest neighbor sample data point in the nearest neighbor set, This represents the function for calculating Euclidean distance. and Let i and j be the i-th and j-th sample data points in the clustered sample dataset; Step S223: Construct the adjacency graph of mutual nearest neighbors; Step S224: Identification of high-density mutually exclusive connected components; Step S225: Adaptive segmentation of the boundary region; Step S226: Redistribute sample data points.
4. The irrigation district water demand prediction method based on multi-source remote sensing data according to claim 3, characterized in that: The construction of the mutual nearest neighbor adjacency graph specifically involves using the clustered sample dataset as the vertex set, constructing an initial undirected complete graph based on the Euclidean distance between sample data points, then performing adaptive pruning on the initial undirected complete graph based on the adaptive neighborhood distance value of each sample data point, traversing all sample data point pairs and filtering valid edges, and constructing a pure mutual nearest neighbor adjacency graph based on the filtered mutual nearest neighbor set. The adaptive pruning rule is that the adjacent edge between two sample data points is retained only when the Euclidean distance between them is less than the minimum value of their adaptive neighborhood distance. The formula used is as follows: ; In the formula, Represents sample data points A set of mutual nearest neighbors, in which all sample data points are adjacent to each other. Preserve adjacent edges. Represents sample data points The adaptive neighborhood distance value; The high-density mutually exclusive connected component identification specifically involves first performing connected component analysis based on the neighbor graph to obtain M initial connected subgraphs that are mutually disjoint and cover all sample data points. Then, vertex mutual exclusion partitioning is performed on all initial connected subgraphs to make the vertex set of each connected subgraph independent and non-overlapping. Next, the maximum density core data point in each connected subgraph is calculated. Finally, high-density mutually exclusive connected components are partitioned according to the high-density mutual exclusion rule. The vertex mutual exclusion partitioning specifically means that the vertex sets of any two different connected subgraphs must satisfy the non-intersection relationship, and the vertex sets of all connected subgraphs must completely cover all sample data points; The highest density core data point is specifically the sample data point that is most representative of the connected subgraph; The high-density mutual exclusion rules specifically include single-core constraints for simply connected domains, inter-domain mutual exclusion constraints, and intra-domain sample data point consistency constraints.
5. The irrigation district water demand prediction method based on multi-source remote sensing data according to claim 3, characterized in that: The adaptive segmentation of the boundary region specifically involves first setting adaptive segmentation trigger conditions. When the number of initial connected subgraphs is less than the preset number of crop growth states, the density conflict domain adaptive segmentation process is triggered. Then, based on the adaptive neighborhood distance values of all sample data points, the global adaptive neighborhood distance mean is calculated to identify boundary sample data points. Next, all adjacent edges associated with the boundary sample data points are located, and the boundary sample data points and their associated edges are temporarily removed from the nearest neighbor adjacency graph. This cuts off the adhesion paths between core regions of different growth states, completing the topological segmentation of the adhesion-connected subgraphs. Connectivity component analysis is then re-executed on the segmented nearest neighbor adjacency graph to update the number of connected subgraphs and high-density mutually exclusive connected regions. The segmentation process is repeated until the number of connected subgraphs equals the preset number of crop growth states, at which point the segmentation process terminates. The final core regions of the growth states are determined by sorting according to the high-density mutually exclusive connected regions. The formula used is as follows: ; ; In the formula, This represents the set of identified boundary sample data points. This represents the mean of the adaptive neighborhood distances for all sample data points. Represents a clustered sample dataset; The sample data point redistribution process involves first re-incorporating temporarily removed boundary sample data points into the sample set to be assigned. Then, based on the high-density mutually exclusive connected components (HDMs) sorting as the core benchmark, and next, using an adaptive neighborhood scale, searching the global adaptive K-nearest neighbor set for each boundary sample data point to be assigned. Sample data points already assigned to the corresponding HDMs are then selected from the nearest neighbors. Finally, a density-weighted adaptive selection rule is used to calculate the matching degree between the boundary sample to be assigned and each HDM, assigning the boundary sample data point to the HDM with the highest matching degree. After completing the full allocation of all sample data points, each HDM represents a growth state, thus outputting the final adaptive clustering result. The formula used is as follows: ; In the formula, Represents the boundary sample data points to be assigned. The degree of matching with high-density mutually exclusive connected domains, Let M represent the m-th high-density mutually exclusive connected component, where M represents the number of connected components.
6. The irrigation district water demand prediction method based on multi-source remote sensing data according to claim 1, characterized in that: In step S3, the construction of the irrigation area water demand feature set specifically involves, based on the results of crop growth heterogeneity zoning within the irrigation area, taking each crop growth status zoning as the basic analysis object, extracting real-time irrigation area water demand prediction data within the corresponding spatial range of the zoning, and combining the corresponding crop growth status information, remote sensing crop water demand data, remote sensing environmental data, and soil environmental data within the same spatial zoning to obtain the water demand feature set of each zoning area.
7. The irrigation district water demand prediction method based on multi-source remote sensing data according to claim 1, characterized in that: In step S4, the construction of the irrigation area regional water demand prediction model specifically includes the following steps: Step S41: Construct the model input layer, specifically based on the water demand feature set of the region, using a preset time step as the basic unit, and perform temporal alignment and ordered splicing of the features in the water demand feature set of the region in chronological order to obtain a continuous temporal feature sequence; Step S42: Construct a dynamic memory-enhanced temporal coding layer. Specifically, a dual-branch structure with gated cyclic branch and dynamic memory branch running in parallel is adopted to perform time-step-by-time deep coding on the input continuous temporal feature sequence. At each time step, the calculation of gated cyclic branch and dynamic memory branch are completed synchronously. Finally, the two outputs are fused to obtain the dynamic memory-enhanced temporal feature of that time step. The specific calculation of the gated loop branch is as follows: taking the water demand feature vector of the partition at time step t and the hidden state and cell state at time step t-1 as input, the gated outputs of the forget gate, input gate and output gate are calculated in sequence, the candidate cell state and the final cell state are updated, and the hidden state at that time step is output. The dynamic memory branch calculation specifically involves assigning dynamic memory weights to each time step of the input sequence. The memory weights of all time steps constitute a global dynamic memory matrix. Based on the partition water requirement feature vector and the corresponding dedicated memory weight of the t-th time step, the dynamic memory state of that time step is calculated. The formula used is as follows: ; ; In the formula, Let represent the hidden state at time step t. This represents the dynamic memory state at time step t. This represents the dynamic memory weight matrix at time step t. This represents the bias vector of the dynamic memory branch. This represents the dynamic memory-enhanced temporal feature of the final output at time step t; Step S43: Construct a time attention weighted layer. Specifically, the attention score of the dynamic memory-enhanced temporal features is calculated step by step through the time attention mechanism. After normalization, the contribution weight of each time step is obtained. Finally, the dynamic memory-enhanced temporal features of all time steps are weighted and summed to obtain the water demand prediction weighted features. Step S44: Construct a zone water demand prediction output layer, specifically by performing a linear dimension transformation through a fully connected layer to directly output the zone irrigation water demand prediction value for the corresponding future time period; Step S45: Water demand prediction model training. Specifically, an initial irrigation area water demand prediction model is obtained by constructing a model input layer, a dynamic memory-enhanced temporal coding layer, a temporal attention weighting layer, and a regional water demand prediction output layer. The standardized historical irrigation area water demand prediction data is used as the training data for the prediction model to train the prediction model, resulting in the trained irrigation area water demand prediction model.
8. The irrigation district water demand prediction method based on multi-source remote sensing data according to claim 1, characterized in that: In step S1, the standardized multi-source data acquisition specifically involves obtaining raw irrigation district water demand management data through data acquisition operations, and then performing data optimization operations on the raw irrigation district water demand management data to obtain standardized irrigation district water demand management data. The standardized irrigation district water demand management data includes standardized historical crop growth monitoring data within the irrigation district, standardized real-time crop growth monitoring data within the irrigation district, standardized historical irrigation district regional water demand prediction data, and standardized real-time irrigation district regional water demand prediction data. The data optimization operation specifically involves sequentially performing data cleaning and data standardization processing on the raw irrigation district water demand management data to obtain standardized irrigation district water demand management data.
9. The irrigation district water demand prediction method based on multi-source remote sensing data according to claim 1, characterized in that: In step S5, the dynamic water demand prediction for each zoning area specifically involves inputting the water demand feature sets corresponding to each zoning area into the trained irrigation zoning water demand prediction model, outputting the predicted irrigation water demand values for each zoning area in the future time period, and finally integrating all the predicted irrigation water demand values to form a real-time irrigation zoning water demand result covering the entire irrigation area. Based on the real-time irrigation zoning water demand result, differentiated irrigation water allocation is performed for different spatial zoning areas to achieve dynamic water demand prediction and precise irrigation based on the spatial heterogeneity of crop growth.
