A planning method for a rainwater collecting and supplemental irrigation system of a slope farmland in an arid region

Abstract

The application provides a rainwater collection and supplementary irrigation system planning method for arid region slope farmland, and relates to the technical field of farmland water conservancy, comprising the following steps: obtaining multi-source geographic spatial data of a research region, and performing data preprocessing on the multi-source geographic spatial data to obtain preprocessed data; inputting the preprocessed data into a pre-trained rainwater collection and supplementary irrigation system planning model for system planning processing, and outputting a spatial coordinate set of a water storage pool; verifying the spatial coordinate set of the water storage pool in combination with a terrain constraint condition, eliminating spatial coordinates that do not satisfy the constraint condition, and outputting a final site selection result that passes the verification; wherein the pre-trained rainwater collection and supplementary irrigation system planning model comprises a terrain perception coding and decoding feature extraction network, a node mapping module and a graph attention network connected in sequence. In this way, the prediction accuracy of the confluence potential heat map and the rationality of the water storage pool site selection result are improved, and end-to-end automatic processing of the rainwater collection and supplementary irrigation system planning process for arid region slope farmland is realized.

Description

Technical Field

[0001] This invention relates to the field of farmland irrigation technology, specifically to a planning method for a rainwater harvesting and supplementary irrigation system for sloping farmland in arid areas. Background Technology

[0002] Sloping farmland resources are widely distributed in arid and semi-arid regions, and the control of soil erosion and efficient utilization of rainwater resources have always been core research topics in the field of agricultural water conservancy. As an important production carrier for agriculture in arid areas, the topographic parameters of sloping farmland, such as slope aspect, slope gradient, and catchment area, directly determine the convergence path and collection efficiency of surface runoff. However, with the intensification of climate change and the frequent occurrence of drought events, how to scientifically plan rainwater harvesting and irrigation systems for sloping farmland to effectively collect, store, and utilize dispersed surface runoff has become a key technical challenge to ensure the stability of agricultural production in arid areas. The planning of rainwater harvesting and irrigation systems involves comprehensive analysis and decision-making based on multi-dimensional information such as topography, precipitation characteristics, soil infiltration capacity, and crop water requirements. Among these, there is a close physical correlation and spatial coupling between the site selection of rainwater harvesting facilities, the configuration of water storage capacity, and the matching of water demand in the irrigation area. Decision deviations in any single link may affect the overall operational efficiency and engineering economy of the system.

[0003] Geographic Information System (GIS) platforms, with their powerful spatial data processing and analysis capabilities, have become the most widely used technical tool in the planning of rainwater harvesting and supplementary irrigation systems for sloping farmland. Existing technologies typically employ multi-factor overlay analysis methods based on GIS platforms for planning. The basic process includes: assigning scores to influencing factors such as slope, aspect, catchment area, and land use type of the study area; determining the weight of each factor using the Analytic Hierarchy Process (AHP) or expert scoring; and using a GIS raster calculator to weightedly overlay the factor layers to generate a comprehensive suitability evaluation layer, which then determines the spatial location and capacity of rainwater harvesting facilities such as cisterns and reservoirs. For facility layout optimization, some technical solutions introduce K-means clustering algorithms to spatially cluster candidate points, or use Genetic Algorithms (GA) to iteratively optimize capacity configuration, aiming to minimize engineering investment while meeting irrigation needs.

[0004] The planning techniques described above, which combine GIS multi-factor overlay analysis with traditional optimization algorithms, have gradually revealed several technical limitations in practical engineering applications. Regarding subjectivity, factor weights and suitability level standards rely heavily on the experience of planners, lacking unified objective criteria. Different personnel often produce significantly different analysis results for the same area, resulting in insufficient scientific rigor and repeatability of the plans. In terms of feature recognition, these planning techniques have limited ability to extract micro-topographic features. Detailed topographic information such as gully distribution and road interception surfaces contained in high-resolution remote sensing images is difficult to fully utilize, yet this information significantly impacts the accurate assessment of local runoff paths and runoff potential. Regarding system correlation, there is a lack of effective spatial topological correlation analysis among slope runoff areas, rainwater harvesting facilities, and irrigation areas, making it difficult to achieve global layout coordination and capacity matching. In terms of solution mechanisms, the K-means algorithm is sensitive to initial cluster centers and prone to getting trapped in local optima, while the genetic algorithm has slow convergence speed and complex parameter calibration. The accuracy and efficiency of solutions under complex terrain constraints are insufficient to meet the needs of large-scale engineering planning. The aforementioned technical limitations are mutually restrictive, seriously affecting the planning quality and construction benefits of rainwater harvesting and supplementary irrigation systems for sloping farmland in arid areas, and there is an urgent need to optimize and upgrade such planning technologies. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a planning method for rainwater harvesting and supplementary irrigation systems on sloping farmland in arid areas.

[0006] The technical problem to be solved by this invention is achieved through the following technical solution: This invention provides a planning method for rainwater harvesting and supplemental irrigation systems on sloping farmland in arid areas, comprising: Multi-source geospatial data of the study area were acquired, and the multi-source geospatial data were preprocessed to obtain preprocessed data. The preprocessed data is input into the pre-trained rainwater harvesting and irrigation system planning model for system planning processing, and the spatial coordinate set of the reservoir is output. The spatial coordinate set of the reservoir is verified in combination with the terrain constraints. Spatial coordinates that do not meet the constraints are removed, and the final site selection result that passes the verification is output. The pre-trained rainwater harvesting and irrigation system planning model includes a terrain-aware encoding / decoding feature extraction network, a node mapping module, and a graph attention network connected in sequence. The terrain-aware encoding / decoding feature extraction network is a network structure containing an encoding branch, a terrain-aware attention module, and a transposed convolutional branch. The terrain-aware attention module is connected between the encoding branch and the transposed convolutional branch and is used to generate a spatial attention weight map based on the preprocessed data and apply the spatial attention weight map to the transposed convolutional branch. The node mapping module is used to map the processing results of the terrain-aware encoding / decoding feature extraction network to nodes and generate heterogeneous graph structure data. The graph attention network performs spatial dependency learning on the heterogeneous graph structure data to obtain the spatial coordinate set of the reservoir.

[0007] The beneficial effects of this invention are: 1. This invention breaks away from the reliance on planners' experience in traditional methods. It outputs a comprehensive predicted score for each candidate node of a reservoir through a multi-layer fully connected regression network, and automatically verifies the spatial coordinate set of the reservoirs by incorporating terrain constraints. This transforms traditional manual qualitative scoring and weighting into data-driven objective calculation. This process reduces human intervention, enabling different personnel to arrive at consistent planning results for the same study area, significantly improving the scientific rigor and repeatability of the planning scheme.

[0008] 2. This invention addresses the shortcomings of traditional methods in fully utilizing micro-topographic information such as gullies and road cutoff surfaces in high-resolution remote sensing imagery. It employs a terrain-aware attention module within a terrain-aware encoding / decoding feature extraction network to extract multi-dimensional target terrain factors and generate a spatial attention weight map. This spatial attention weight map adaptively applies to each level of the input feature map in the transposed convolutional branch, enhancing the response in key confluence areas. This enables the model to accurately capture refined terrain information, thereby improving the prediction accuracy of confluence potential heatmaps for surface runoff collection characteristics and providing an accurate data foundation for subsequent site selection.

[0009] 3. This invention addresses the pain point of existing technologies lacking effective spatial topological correlation analysis among the slope runoff area, reservoir candidate area, and irrigation area. This invention maps these areas to different nodes using a node mapping module and constructs heterogeneous graph structure data containing spatial topological relationships. By utilizing a graph attention network and its pre-constructed physical element paths to learn spatial dependencies in this heterogeneous graph structure data, it fully explores the spatial coupling relationships between functional areas, overcoming the limitations of local optimization in traditional methods and achieving layout coordination and capacity matching of the three major areas from a global perspective.

[0010] 4. This invention addresses the shortcomings of traditional genetic algorithms, such as slow convergence, the tendency of K-means algorithms to get trapped in local optima, and the complexity of parameter calibration. It employs a complete end-to-end neural network architecture. By integrating data preprocessing, terrain feature extraction, spatial topology analysis, and result verification into a unified framework, it directly outputs the spatial coordinate set of the reservoir. This end-to-end processing significantly shortens the planning computation time and enhances the global optimization capability under complex terrain constraints, fully meeting the practical engineering requirements for efficiency and accuracy in large-scale rainwater harvesting and irrigation projects on sloping farmland in arid areas.

[0011] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0012] Figure 1 A flowchart illustrating a planning method for a rainwater harvesting and supplemental irrigation system for sloping farmland in arid areas, provided by an embodiment of the present invention; Figure 2 An exemplary schematic diagram of the structure of a terrain-aware encoding / decoding feature extraction network is shown. Detailed Implementation

[0013] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0014] To improve the prediction accuracy of heat maps of runoff potential and the rationality and engineering applicability of reservoir site selection, and to realize end-to-end automated processing and global collaborative optimization of the planning process of rainwater harvesting and supplementary irrigation systems for sloping farmland in arid areas, this invention provides a planning method for rainwater harvesting and supplementary irrigation systems for sloping farmland in arid areas. Figure 1 This is a flowchart illustrating a planning method for a rainwater harvesting and supplemental irrigation system on sloping farmland in arid areas, provided by an embodiment of the present invention. Figure 1 As shown, it includes: S101. Obtain multi-source geospatial data of the study area and perform data preprocessing on the multi-source geospatial data to obtain preprocessed data.

[0015] Optionally, the multi-source geospatial data includes: remote sensing image data and digital elevation model data, S101 including: The remote sensing image data is sequentially subjected to radiometric correction, atmospheric correction and geometric fine correction to obtain preprocessed remote sensing image data. The digital elevation model data is sequentially subjected to hole filling and outlier correction processing to obtain preprocessed digital elevation model data. Preprocessed remote sensing image data and preprocessed digital elevation model data together constitute the preprocessed data.

[0016] In one specific embodiment, the study area in step S101 is a typical arid sloping farmland distribution area. The acquired multi-source geospatial data includes remote sensing image data and digital elevation model (DEM) data. The remote sensing image data can be medium-to-high resolution optical remote sensing images such as Gaofen-1 or operational land imagers. The DEM data can be SRTM (Shuttle Radar Topography Mission) DEM (Digital Elevation Model) products, with a preferred spatial resolution of 30 meters. After acquisition, the remote sensing image data undergoes radiometric correction, atmospheric correction, and geometric fine correction sequentially to eliminate the effects of atmospheric scattering, sensor response differences, and geometric distortion, resulting in preprocessed remote sensing image data. Radiometric correction uses a radiometric calibration tool to convert digital quantization values ​​into radiance values. Atmospheric correction uses the FLAASH (Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes) model. Geometric fine correction uses polynomial fitting based on ground control points to ensure that the image's planar position accuracy is better than one pixel.

[0017] The digital elevation model (DEM) data undergoes sequential hole-filling and outlier correction processing to obtain preprocessed DEM data. Hole filling can be performed using inverse distance weighted interpolation or spline interpolation methods, utilizing surrounding valid elevation values ​​to fill missing areas. Outlier correction identifies and corrects peaks or depressions caused by data noise by setting slope or elevation change thresholds. After these processing steps, the preprocessed remote sensing image data and the preprocessed DEM data are combined to form preprocessed data, which serves as input to the subsequent terrain-aware encoding / decoding feature extraction network.

[0018] S102. Input the preprocessed data into the pre-trained rainwater harvesting and irrigation system planning model for system planning processing, and output the set of spatial coordinates of the water storage tank.

[0019] The pre-trained rainwater harvesting and irrigation system planning model includes a terrain-aware encoding / decoding feature extraction network, a node mapping module, and a graph attention network connected in sequence. The terrain-aware encoding / decoding feature extraction network is a network structure containing an encoding branch, a terrain-aware attention module, and a transposed convolutional branch. The terrain-aware attention module is connected between the encoding branch and the transposed convolutional branch and is used to generate a spatial attention weight map based on the preprocessed data and apply the spatial attention weight map to the transposed convolutional branch. The node mapping module is used to map the processing results of the terrain-aware encoding / decoding feature extraction network to nodes and generate heterogeneous graph structure data. The graph attention network performs spatial dependency learning on the heterogeneous graph structure data to obtain the spatial coordinate set of the reservoir.

[0020] Optionally, S102 includes: A terrain-aware encoding and decoding feature extraction network is used to perform multi-scale feature extraction on preprocessed data and feature reuse processing based on skip connections to obtain a heat map of confluence potential. The node mapping module uses the heat map of confluence potential to identify the pre-planned functional areas in the study area and maps the pre-planned functional areas to different nodes. Different physical and hydrological attribute vectors are assigned according to the attribute type of the nodes to obtain the mapped node information. The mapped node information is used to construct an adjacency matrix according to the spatial topological relationship between each node to generate heterogeneous graph structure data. Graph attention networks use pre-constructed physical metapaths to learn spatial dependencies in heterogeneous graph structure data, resulting in a set of spatial coordinates for the reservoir.

[0021] Optionally, the pre-planned functional areas include: slope runoff area, reservoir candidate area and irrigation area.

[0022] In one specific embodiment, the node mapping module utilizes a convergence potential heatmap. Identify pre-planned functional areas within the study area, among which... Heatmap showing confluence potential Represents the set of real numbers. The pixel width of the heatmap representing confluence potential. This represents the pixel height of the runoff potential heatmap. The runoff potential heatmap is a two-dimensional probability map with the same spatial resolution as the original study area; each pixel value represents the magnitude of the surface runoff runoff potential at that location. The node mapping module uses a preset runoff potential threshold. ( The value ranges from 0 to 1, and the extracted value from the heatmap of confluence potential is higher than... The connected regions were designated as slope runoff areas; based on the land use map of the study area, cultivated land patches were identified as irrigation areas; combined with topographic runoff accumulation analysis, areas downstream of the slope runoff areas and upstream of the irrigation areas, with a topographic slope less than a preset slope threshold, were identified. On flat terrain, candidate areas for reservoirs are delineated. Preferably, the catchment potential threshold is... Preset slope threshold .

[0023] The above identification process involves scanning the heat map of confluence potential grid by grid, extracting connected regions using a connected component labeling algorithm, and binding each region to the geographical coordinates of the study area through spatial overlay analysis to obtain a set of pre-defined planned functional areas. ,in This represents a set of pre-planned functional areas. Indicates the slope confluence area. Indicates the candidate area for the reservoir. Indicates the irrigated area.

[0024] Identified slope confluence areas Candidate areas for reservoirs and irrigation areas Each node is mapped to a different node, and each node corresponds to an independent functional unit, thus obtaining a node set. ,in Represents the set of all nodes. Indicates the first Nodes ( For node index, the value range is 1 to ), This represents the total number of nodes. Different physical hydrological attribute vectors are assigned to nodes according to their attribute types to obtain mapped node information.

[0025] Specifically, define the node type mapping function. ,in This represents a mapping function from a set of nodes to a set of type labels. Indicates the type of slope runoff area label. Indicates the candidate area type label for the water storage tank. This indicates the irrigation area type label. The node type mapping function is used to map each node... Map to the corresponding type tag Assign attribute vectors based on the attribute type of the node. ,in Indicates the first The physical hydrological attribute vector of each node, The dimension is The real vector space, The type is The attribute vector dimension of a node can vary depending on the node type. For example, the physical hydrological attribute vector corresponding to a node in a slope runoff area. This may include: catchment area, mean slope, aspect, surface roughness, soil saturated hydraulic conductivity, and runoff coefficient; and the physical and hydrological attribute vectors corresponding to the candidate nodes of the reservoir area. This may include: terrain flatness, catchment area, Euclidean distance to the nearest irrigated area, length of the runoff path from the slope catchment area, and groundwater level depth; physical hydrological attribute vectors corresponding to the nodes in the irrigated area. This can include: crop type, irrigation water requirement, soil field capacity, and irrigated area. The physical hydrological attribute vector of each node can also include the spatial coordinates of the node's center point within the study area, serving as the node feature input for subsequent graph attention network processing.

[0026] After completing node mapping and attribute assignment, the mapped node information is used to construct an adjacency matrix according to the spatial topological relationships between the nodes. Generate heterogeneous graph structure data, in which This represents the adjacency matrix.

[0027] Adjacency Matrix Middle node With nodes Corresponding elements Defined as: Represents a node With nodes There are spatial topological connections between them. Represents a node With nodes There is no connection between them. and All are node indices, with values ​​ranging from 1 to... , Indicates the first 1 node Indicates the first There are several nodes. Specifically, if the outlet location of a node in the slope runoff area is connected to the inlet location of a node in the reservoir candidate area through a surface runoff path, an edge is established between the two nodes; if the water supply range of a node in the reservoir candidate area covers the geometric range of a node in the irrigation area, an edge is established between the two nodes. The above spatial topology relationships can be automatically determined based on the flow direction grid and water system network extracted from the preprocessed digital elevation model data.

[0028] Map the node information of each node Adjacency Matrix Together they constitute heterogeneous graph structure data ,in Represents heterogeneous graph structure data. This represents the set of edges established between all nodes (an edge represents a spatial topological connection between two nodes). The node feature matrix is ​​represented by the row-wise stacking of the physical and hydrological attribute vectors of all nodes. (representing matrix transpose) This represents a node type vector (composed of all node type labels stacked row by row). This heterogeneous graph structure data contains three types of heterogeneous nodes: slope runoff area nodes, reservoir candidate area nodes, and irrigation area nodes, as well as the spatial topological relationships between them, providing a data foundation for subsequent graph attention network processing.

[0029] Optionally, the terrain-aware encoding and decoding feature extraction network also includes a spatial transformation network, wherein the encoding branch includes a first residual block, a second residual block and a third residual block connected in sequence, and the transposed convolution branch includes a first transposed convolutional layer, a second transposed convolutional layer and a third transposed convolutional layer connected in sequence.

[0030] Figure 2 An exemplary schematic diagram of a terrain-aware encoding / decoding feature extraction network is shown. Combined with... Figure 2 By utilizing a terrain-aware encoding / decoding feature extraction network to perform multi-scale feature extraction on preprocessed data and feature reuse processing based on skip connections, a heatmap of confluence potential is obtained, including: The spatial transformation network is used to spatially register the preprocessed remote sensing image data and preprocessed digital elevation model data in the preprocessed data, and then the data is stitched together according to the channel dimension to form the input feature tensor. The input feature tensor is passed through the first residual block, the second residual block, and the third residual block in sequence for downsampling and deep feature extraction, and the corresponding outputs are shallow coding feature map, medium coding feature map, and deep coding feature map; The preprocessed digital elevation model data is input into the terrain-aware attention module to extract the target terrain factors, and a spatial attention weight map is generated based on the target terrain factors. The deep encoding feature map and the spatial attention weight map are weighted element-wise through the first element-wise weighting unit to generate a deep weighted feature map. The deep weighted feature map is then input into the first transposed convolutional layer for upsampling to obtain a deep upsampled feature map. The deep upsampled feature map and the middle-layer encoding feature map are concatenated along the channel dimension through the first feature concatenation unit to obtain a deep-middle-layer concatenated feature map. The deep-middle layer concatenated feature map is obtained by performing convolution processing on the deep-middle layer concatenated feature map through the first convolution processing unit; The deep-to-middle layer decoding feature map and the spatial attention weight map are weighted element-wise by the second element-wise weighting unit to generate a middle layer weighted feature map. The middle layer weighted feature map is then input into the second transposed convolutional layer for upsampling to obtain a middle layer upsampled feature map. The middle layer upsampled feature map and the shallow layer encoding feature map are concatenated along the channel dimension by the second feature concatenation unit to obtain a middle-to-shallow concatenated feature map. The middle-shallow layer concatenated feature map is obtained by performing convolution processing on the middle-shallow layer concatenated feature map using the second convolution processing unit; The middle-shallow decoding feature map and the spatial attention weight map are weighted element-wise using the third element-wise weighting unit to generate a shallow weighted feature map. The shallow weighted feature map is then input into the third transposed convolutional layer for upsampling to obtain a shallow upsampled feature map. After mapping the shallow upsampled feature map using an activation function layer, a heatmap of confluence potential is generated.

[0031] In this invention, after completing the basic preprocessing of multi-source geospatial data, the spatial transformation network receives preprocessed remote sensing image data and preprocessed digital elevation model (DEM) data as input. First, the spatial transformation network performs spatial registration on the preprocessed remote sensing image data and preprocessed DEM data. By unifying the preprocessed remote sensing image data and preprocessed DEM data to the same projected coordinate system and aligning them with a spatial reference, it ensures that the spatial coordinates of the preprocessed remote sensing image data and preprocessed DEM data achieve a precise pixel-by-pixel correspondence, eliminating spatial positional deviations caused by different data sources. Subsequently, the spatial transformation network stitches the spatially registered preprocessed remote sensing image data and preprocessed DEM data according to the channel dimension. That is, the single-channel preprocessed DEM data is treated as an additional data channel and superimposed on the multi-channel preprocessed remote sensing image data. Through the above channel-dimensional stitching operation, a final input feature tensor containing surface spectral information and elevation topographic information is generated.

[0032] In practice, the terrain perception attention module uses preprocessed digital elevation model data as input to extract multi-dimensional target terrain factors. Specifically, the third-order inverse distance squared weighted difference algorithm is used to obtain the slope factor, the eight-way neighborhood method is used to calculate the aspect factor, and the local quadratic surface fitting method is used to quantify the terrain curvature factor. Simultaneously, a deterministic eight-neighborhood single-flow direction algorithm is used to progressively accumulate the confluence contribution area of ​​each grid cell to obtain the confluence accumulation, which is then combined with the slope factor to derive the terrain humidity index. Because there are significant differences in the physical dimensions and numerical magnitudes among the various target terrain factors, such as the slope factor, aspect factor, terrain curvature factor, confluence accumulation, and terrain humidity index, after extraction, Min-Max numerical normalization is performed on each target terrain factor to uniformly compress them into the numerical range [0,1]. Subsequently, the normalized slope factor, aspect factor, topographic curvature factor, runoff accumulation, and topographic humidity index, among other single-channel factor matrices, are stacked along the channel dimension to form a multi-channel topographic feature tensor. Then, a micro-convolutional mapping network configured within the topographic awareness attention module is used to perform a nonlinear feature transformation on this topographic feature tensor in the spatial dimension. The transformed feature map is then mapped and compressed using a sigmoid activation function, ultimately generating a spatial attention weight map that perfectly matches the spatial resolution of the input feature maps at each level of the transposed convolutional branch. This spatial attention weight map can intuitively reflect the contribution of topographic conditions at each spatial location to surface runoff accumulation, providing accurate weight guidance for the subsequent decoding process.

[0033] The micro-convolutional mapping network consists of three sequentially connected convolutional layers, with an activation function layer between every two convolutional layers.

[0034] In one specific embodiment, the first residual block, the second residual block, and the third residual block have the same structure; the first residual block, the second residual block, and the third residual block are all composed of a first convolutional layer, a first batch normalization layer, a first ReLU activation function layer, a second convolutional layer, a second batch normalization layer, and an element-wise addition layer connected in sequence. The first, second, and third transposed convolutional layers have the same structure. They are all based on a transposed convolutional layer, a third batch normalization layer, and a second ReLU activation function layer connected in sequence.

[0035] Optionally, the graph attention network employs the GATv2 model, using pre-constructed physical meta-paths to learn spatial dependencies in the heterogeneous graph structure data, resulting in a set of spatial coordinates for the reservoir, including: According to each physical element path, the corresponding node sequence and edge sequence are extracted from the heterogeneous graph structure data to generate the isomorphic subgraph corresponding to each physical element path; Input the physical hydrological attribute vectors of each node in each isomorphic subgraph into the GATv2 model to obtain the path-level encoded feature vectors of each node under each physical metapath. The path-level encoded feature vectors belonging to the same node are fused to generate a fused feature vector for each node. The fused feature vectors of each candidate node in the reservoir are input into a multilayer fully connected regression network, which outputs the comprehensive score and predicted volume value of each candidate node in the reservoir. The spatial coordinates of the candidate nodes in the reservoir that meet the preset score threshold are selected to form the spatial coordinate set of the reservoir.

[0036] After obtaining the path-level encoded feature vectors of each node under each physical metapath, the graph attention network needs to fuse multiple sets of path-level encoded feature vectors belonging to the same node to generate a fused feature vector that can comprehensively represent the spatial dependencies of the node under multiple physical metapaths.

[0037] The pre-constructed physical element paths are node type connection patterns pre-defined based on the physical-hydrological relationships between functional areas in the rainwater harvesting and irrigation system for arid sloping farmland. Specifically, based on the physical-hydrological relationships between the slope runoff area, the candidate reservoir area, and the irrigation area, node type connection patterns for each physical element path are pre-defined. For example, the node type connection patterns include a first pattern where nodes in the slope runoff area point to nodes in the candidate reservoir area, a second pattern where nodes in the candidate reservoir area point to nodes in the irrigation area, and a third pattern where nodes in the slope runoff area point to nodes in the irrigation area via nodes in the candidate reservoir area. After the node mapping module generates heterogeneous graph structure data, according to the pre-defined node type connection patterns for each physical element path, node sequences and edge sequences conforming to the patterns are extracted from the heterogeneous graph structure data to generate isomorphic subgraphs corresponding to each physical element path.

[0038] For example, a weighted fusion method based on weights can be adopted. For the path-level encoded feature vectors corresponding to each physical meta-path of the same node, a set of learnable adaptive fusion weight coefficients is pre-defined. The path-level encoded feature vectors of the same node are weighted and summed according to the corresponding fusion weight coefficients to generate the fused feature vector of the node. When implementing adaptive weight learning, a cross-path attention fusion mechanism can be further introduced. That is, by constructing a miniature cross-path attention mapping network, the path-level encoded feature vectors of the same node under different physical meta-paths are used as input to this cross-path attention mapping network. The attention mechanism automatically calculates the contribution score of each physical meta-path to the feature expression of the node, and then the path-level encoded feature vectors are weighted and aggregated according to the contribution score to obtain the fused feature vector of the node.

[0039] In addition, a fusion method combining feature concatenation and linear transformation can be used. Graph attention networks sequentially concatenate the path-level encoded feature vectors of each physical element path belonging to the same node along the vector dimension to form a higher-dimensional concatenated feature vector. Then, a single-layer or multi-layer fully connected mapping network is used to perform dimensionality reduction and nonlinear transformation on this higher-dimensional concatenated feature vector, finally generating a compact fused feature vector for that node.

[0040] All of the above-mentioned fusion methods can effectively integrate and refine the spatial dependency information of multiple physical path elements under the same node, thereby providing richer and more accurate feature inputs for inputting the fusion feature vectors of candidate nodes of each reservoir into a multilayer fully connected regression network.

[0041] S103. Verify the spatial coordinate set of the reservoir in conjunction with the terrain constraints, remove spatial coordinates that do not meet the constraints, and output the final site selection result that passes the verification.

[0042] In practice, after the graph attention network outputs the set of spatial coordinates of the reservoir, this set of coordinates is used as a candidate set of coordinates to be validated. Subsequently, the spatial coordinate set of the reservoir is validated in conjunction with terrain constraints. These terrain constraints specifically include slope constraints, geological stability constraints, and catchment area constraints.

[0043] In the specific verification process, for each candidate spatial coordinate in the candidate coordinate set, the corresponding topographic spatial features are extracted, and its slope gradient, geological stability score, and catchment area are calculated. These calculated values ​​are then compared one by one with preset slope gradient thresholds, geological stability thresholds, and catchment area thresholds. If the slope gradient, geological stability score, and catchment area of ​​a candidate spatial coordinate all simultaneously meet the corresponding topographic constraints, the candidate spatial coordinate is deemed to meet the engineering site selection requirements and is retained. If any one of these indicators does not meet the corresponding topographic constraints, the candidate spatial coordinate is removed from the reservoir's spatial coordinate set. After verifying all candidate spatial coordinates one by one, all the retained candidate spatial coordinates that meet the topographic constraints are collected, and the final site selection result that passes the verification is output. Preferably, the slope gradient threshold can be 5°, the geological stability threshold can be 200 kPa, and the catchment area threshold can be 0.1 hectares.

[0044] Optionally, the training process of the pre-trained rainwater harvesting and supplemental irrigation system planning model includes: Acquire multi-source geospatial sample data, as well as labeled heatmaps of confluence potential, labeled final site selection coordinates, and labeled planned volume values ​​corresponding to the multi-source geospatial sample data; Data preprocessing is performed on multi-source geospatial sample data to obtain preprocessed sample data; The preprocessed sample data is input into the initial rainwater harvesting and irrigation system planning model for the first stage of training through multiple iterations. During the first stage of training, the heat map of the labeled runoff potential is used as a supervision signal, and the terrain perception encoding and decoding feature extraction network in the initial rainwater harvesting and irrigation system planning model is iteratively trained using the first preset loss function. The network parameters of the trained terrain perception encoding and decoding feature extraction network are frozen to obtain the current rainwater harvesting and irrigation system planning model. The preprocessed sample data is input into the current rainwater harvesting and irrigation system planning model for a second stage of training with multiple rounds of iteration. The predicted comprehensive score and predicted volume value of each candidate node of the reservoir are output. The predicted comprehensive score and predicted volume value, along with the labeled final site selection coordinates and labeled planned volume value, are substituted into the second preset loss function for calculation. The network parameters of the current rainwater harvesting and irrigation system planning model are iteratively adjusted according to the calculation results. Repeat the above iterative training process until the preset iteration stopping condition is met. Use the current rainwater harvesting and irrigation system planning model when the preset iteration stopping condition is met as the pre-trained rainwater harvesting and irrigation system planning model. The preset iteration stopping condition is that the number of iterations meets the iteration number threshold or the values ​​of the first preset loss function and the second preset loss function are continuously less than the corresponding loss threshold.

[0045] Optionally, the first preset loss function is expressed as: ; in, This represents the value of the first preset loss function. This indicates the potential for convergence loss. This indicates the loss of structural similarity in the heatmap of confluence potential. Indicates the regularization loss. This represents the regression loss coefficient of the confluence potential. The coefficient representing the structural similarity loss in the heatmap of confluence potential. This represents the regularization loss coefficient.

[0046] The second preset loss function is a weighted combination of location regression loss and volume regression loss; both location regression loss and volume regression loss use the mean squared error loss function.

[0047] Alternatively, the convergence potential regression loss can be expressed as: ; in, This represents the total number of pixels in the heatmap indicating confluence potential. The heatmap indicating the potential of confluence points is shown below. The annotation value of each pixel. The first element in the heatmap representing the predicted confluence potential is... The predicted value of each pixel; The structural similarity loss of the confluence potential heatmap is represented as: ; ; in, The total number of image patches in the heatmap of confluence potential. To mark the first in the heat map of confluence potential Image blocks, To predict the confluence potential heat map, the first Image blocks, express and Structural similarity index between them and They represent and The average pixel value, and They represent and pixel variance for and covariance, Represents the first numerical stability constant. This represents the second numerical stability constant. , .

[0048] The regularization loss is expressed as: ; in, This represents the total number of network layers to be regularized in the initial rainwater harvesting and supplemental irrigation system planning model. For the first The weight parameter matrix of the layer to be regularized network, This represents the L2 norm.

[0049] In practical implementation, the three loss coefficients in the aforementioned first preset loss function are addressed as follows: , and The method for determining its value needs to comprehensively consider the order-of-magnitude differences of each loss component during actual training. This is due to the convergence potential regression loss. It reflects pixel-level numerical deviations, and its converged magnitude is usually small (e.g., to Level); Structural similarity loss of confluence potential heatmap It reflects the differences in the overall structure of the image, and its value is generally in the range of to Between; and regularization loss The magnitude of the loss coefficient is related to the number of network layers and the size of the weight matrix, and is usually relatively small. If the loss coefficient is set improperly, the optimization of other loss terms may be neglected due to the excessive magnitude of a single loss term during network training, thus affecting the overall prediction quality of the confluence potential heatmap.

[0050] Therefore, this embodiment uses a combination of magnitude balancing and grid search tuning to determine these three loss coefficients. First, before formal training, an initial baseline coefficient is set (e.g., all set to 1), and a small number of pre-training rounds are run. The individual loss components are then statistically analyzed. , , The initial value is on the order of magnitude of ). Then, the loss coefficient is adjusted based on the statistical results, so that... , and Three weighted terms The contributions can be kept within a similar order of magnitude. After initial order of magnitude matching, grid search cross-validation is performed within a preset coefficient search range to finally determine a set of hyperparameters that best predicts the confluence potential heatmap on the validation set.

[0051] For example, in one specific implementation of this embodiment, the final loss coefficient value is as follows: 1. Convergence potential regression loss coefficient The value is set to 1 to ensure the accuracy weight of pixel-level numerical prediction; 2. Structural similarity loss coefficient of confluence potential heatmap A value of 1 is used to ensure structural similarity while avoiding overweighting of data fitting. 3. Regularization loss coefficient The value is set to 0.001, which means that only a weak L2 regularization constraint is applied to prevent the model from overfitting, while avoiding the introduction of excessive regularization penalty that would interfere with the effective learning of the confluence potential features.

[0052] By setting the aforementioned loss coefficients, the guiding effects of the three loss functions on network parameter updates can be effectively balanced, enabling the trained rainwater harvesting and irrigation system planning model to achieve optimal structural and numerical accuracy in the heatmap of runoff potential. Of course, in actual engineering implementation, the three loss coefficients can be fine-tuned proportionally based on the terrain complexity of the specific study area, the amount of sample data, and the network training convergence.

[0053] The second pre-defined loss function is expressed as follows: ; ; ; in, This represents the value of the second preset loss function. This represents the location regression loss coefficient. Indicates the site selection regression loss. This represents the volumetric regression loss coefficient. Indicates volumetric regression loss; This indicates the number of candidate nodes for the water storage pool. Indicates the first Candidate nodes for water storage tanks Indicates the first The predicted comprehensive score of each candidate node for the reservoir. Indicates the first The annotation score for the final location coordinates corresponding to each candidate node of the reservoir. Indicates the first The predicted volume of each candidate node in the reservoir. Indicates the first The labeled planned volume value corresponding to each candidate node of the reservoir. Preferably, the site selection regression loss coefficient can be 1, the volume regression loss coefficient can be 0.000001, the loss threshold of the first preset loss function can be 0.001, and the loss threshold of the second preset loss function can be 0.01.

[0054] This invention provides a planning method for rainwater harvesting and supplemental irrigation systems on sloping farmland in arid areas. First, a multi-layer fully connected regression network is used to output the predicted comprehensive score of each candidate node for a water storage tank. Then, the spatial coordinate set of the water storage tanks is automatically verified in conjunction with topographic constraints. This transforms traditional manual qualitative scoring and weight setting into data-driven objective calculation. This process reduces human experience intervention, enabling different personnel to arrive at consistent planning results for the same study area, significantly improving the scientific rigor and repeatability of the planning scheme.

[0055] Then, the terrain-aware attention module in the terrain-aware encoding and decoding feature extraction network is used to extract multi-dimensional target terrain factors, and a spatial attention weight map is generated accordingly. This spatial attention weight map can adaptively apply to the input feature maps of each level of the transposed convolutional branch, enhance the response of key confluence areas, enable the model to accurately capture refined terrain information, thereby improving the prediction accuracy of the confluence potential heat map for surface runoff collection characteristics, and providing an accurate data foundation for subsequent site selection.

[0056] Secondly, the above-mentioned regions are mapped to different nodes through the node mapping module, and a heterogeneous graph structure data containing spatial topological relationships is constructed. Using a graph attention network and pre-constructed physical metapaths, spatial dependencies of this heterogeneous graph structure data are learned, fully exploring the spatial coupling relationships between functional regions. This overcomes the limitations of local optimization in traditional methods and achieves layout coordination and capacity matching of the three major regions from a global perspective.

[0057] Finally, a complete end-to-end neural network architecture is adopted. By integrating data preprocessing, terrain feature extraction, spatial topology analysis, and result verification into a unified framework, the spatial coordinate set of the reservoir is directly output. This end-to-end processing method significantly shortens the planning computation time, enhances the global optimization capability under complex terrain constraints, and can fully meet the actual engineering requirements for efficiency and accuracy in the planning of large-scale rainwater harvesting and supplementary irrigation projects on sloping farmland in arid areas.

[0058] To verify the effectiveness and practical engineering applicability of the rainwater harvesting and supplementary irrigation system planning method for sloping farmland in arid areas proposed in this invention, this embodiment further conducted simulation experiments for verification.

[0059] A. Experimental setup and data sources: This simulation experiment selects a typical arid region of sloping farmland in Northwest my country as the study area. The experimental data sources include Gaofen-1 multispectral remote sensing image data (spatial resolution 2 meters) and SRTM digital elevation model data (spatial resolution 30 meters). According to the specific implementation of this invention, the remote sensing image data is first subjected to radiometric correction, atmospheric correction, and geometric fine correction to obtain preprocessed remote sensing image data; the digital elevation model data is then subjected to hole filling and outlier correction to obtain preprocessed digital elevation model data. These two types of data together constitute the preprocessed data required for the experiment. Based on spatial distribution characteristics, the study area is divided into a training set (70%), a validation set (15%), and a test set (15%). The experiment uses an Intel Xeon Platinum processor and an NVIDIA A100 GPU as the hardware platform, and the software is built based on the PyTorch deep learning framework.

[0060] B. Selection of comparison methods: This simulation experiment selects three methods commonly used in existing technologies for planning rainwater harvesting and supplementary irrigation systems on sloping farmland as comparative examples: 1. GIS multi-factor overlay method: Based on the analytic hierarchy process (AHP), the weights of influencing factors such as slope, aspect, catchment area, and land use type are determined. The GIS raster calculator is used to weight and overlay the layers of each factor to generate a comprehensive suitability evaluation layer to determine the spatial location of the rainwater harvesting facilities; 2. K-means clustering algorithm: Based on the candidate points generated by the GIS multi-factor overlay method, the K-means algorithm is used to optimize the spatial clustering of the candidate points; 3. Genetic algorithm (GA) method: Taking irrigation water demand and engineering investment as objective functions, the genetic algorithm is used to iteratively optimize the capacity configuration and spatial location of the reservoir.

[0061] C. Evaluation Indicators: The following metrics are used to evaluate the model performance in this simulation experiment: 1. Confluence potential regression loss: reflects the pixel-level numerical deviation between the predicted confluence potential heatmap and the labeled confluence potential heatmap; 2. Confluence potential heatmap structural similarity loss: reflects the similarity of the predicted confluence potential heatmap and the labeled confluence potential heatmap in terms of image structure; 3. Site selection success rate: defined as the proportion of nodes in the final output set of reservoir spatial coordinates of the model whose deviation from the actual engineering site selection coordinates is within a preset tolerance radius (set to 30 meters) to the total number of candidate nodes output by the model; 4. Average inference time: refers to the end-to-end average time taken by the model from input preprocessed data to the final output set of reservoir spatial coordinates.

[0062] D. Simulation Results and Analysis: On the test set, the method of this invention was compared with the three existing methods mentioned above. All models were trained and inferred under the same data processing flow and training set conditions. The simulation results of different planning methods are shown in Table 1.

[0063] Table 1

[0064] As can be seen from the experimental results in Table 1, firstly, in terms of the prediction accuracy of the confluence potential heatmap, the confluence potential regression loss of the method of this invention is only 0.0082, and the structural similarity loss of the confluence potential heatmap reaches 0.894, both significantly better than the other three existing methods. This indicates that the spatial attention weight map generated by the terrain-aware encoding and decoding feature extraction network in this invention through the terrain-aware attention module can adaptively enhance the response of key confluence areas, greatly improving the prediction accuracy of the confluence potential heatmap for surface runoff convergence characteristics under complex terrain.

[0065] Secondly, regarding site selection accuracy, the method of this invention achieves a success rate of 95.7%. This is because the invention maps the slope confluence area, reservoir candidate area, and irrigation area to nodes in the heterogeneous graph structure data through a node mapping module. Combined with a graph attention network, it learns spatial dependencies based on pre-constructed physical element paths, fully exploring the spatial topological relationships between functional areas. Furthermore, it verifies the spatial coordinate set of the reservoir by incorporating terrain constraints, eliminating coordinates that do not meet the terrain constraints. This effectively avoids the shortcomings of traditional methods, such as easily getting trapped in local optima or relying on manual experience in weight setting.

[0066] Finally, in terms of operational efficiency, the average inference time of the method of this invention is only 12.4 seconds. Compared with the complex and time-consuming iterative optimization of genetic algorithms, this invention adopts an end-to-end neural network model inference, which significantly reduces the planning computation time. At the same time, since data preprocessing, multi-scale feature extraction, spatial topology analysis and result verification are integrated into a unified framework, the processing flow is significantly simplified and the computational efficiency is greatly improved compared with the traditional GIS multi-factor overlay method.

[0067] In summary, the simulation results verify that the method provided by this invention has significant technical advantages in terms of the accuracy of heat map prediction of runoff potential, the success rate of final site selection of reservoirs, and the processing efficiency of end-to-end planning. It can meet the actual engineering needs of large-scale rainwater harvesting and supplementary irrigation projects for sloping farmland in arid areas.

[0068] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the inventive concept, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A planning method for rainwater harvesting and supplemental irrigation systems on sloping farmland in arid areas, characterized in that, include: Acquire multi-source geospatial data of the study area and perform data preprocessing on the multi-source geospatial data to obtain preprocessed data; The preprocessed data is input into a pre-trained rainwater harvesting and irrigation system planning model for system planning processing, and the spatial coordinate set of the water storage tank is output. The spatial coordinate set of the reservoir is verified in combination with the terrain constraints. Spatial coordinates that do not meet the constraints are eliminated, and the final site selection result that passes the verification is output. The pre-trained rainwater harvesting and irrigation system planning model includes a terrain-aware encoding / decoding feature extraction network, a node mapping module, and a graph attention network connected in sequence. The terrain-aware encoding / decoding feature extraction network has a network structure containing an encoding branch, a terrain-aware attention module, and a transposed convolutional branch. The terrain-aware attention module is connected between the encoding branch and the transposed convolutional branch and is used to generate a spatial attention weight map based on the preprocessed data and apply the spatial attention weight map to the transposed convolutional branch. The node mapping module is used to map the processing results of the terrain-aware encoding / decoding feature extraction network to nodes and generate heterogeneous graph structure data. The graph attention network performs spatial dependency learning processing on the heterogeneous graph structure data to obtain the spatial coordinate set of the reservoir.

2. The planning method for a rainwater harvesting and supplementary irrigation system for sloping farmland in arid areas according to claim 1, characterized in that, The multi-source geospatial data includes: remote sensing image data and digital elevation model data. The process involves acquiring multi-source geospatial data of the study area and preprocessing the multi-source geospatial data to obtain preprocessed data, including: The remote sensing image data is sequentially subjected to radiometric correction, atmospheric correction and geometric fine correction to obtain preprocessed remote sensing image data. The digital elevation model data is subjected to hole filling and outlier correction processes in sequence to obtain preprocessed digital elevation model data. The preprocessed remote sensing image data and the preprocessed digital elevation model data together constitute the preprocessed data.

3. The planning method for a rainwater harvesting and supplementary irrigation system for sloping farmland in arid areas according to claim 1, characterized in that, The process of inputting the preprocessed data into a pre-trained rainwater harvesting and irrigation system planning model for system planning processing, and outputting a set of spatial coordinates of the reservoir, includes: The terrain-aware encoding and decoding feature extraction network is used to perform multi-scale feature extraction and skip connection-based feature reuse processing on the preprocessed data to obtain a heat map of confluence potential. The node mapping module uses the confluence potential heat map to identify the preset planned functional areas within the study area, and maps the preset planned functional areas to different nodes. Different physical and hydrological attribute vectors are assigned to the nodes according to their attribute types to obtain the mapped node information. The mapped node information is then used to construct an adjacency matrix according to the spatial topological relationship between the nodes to generate the heterogeneous graph structure data. The graph attention network uses pre-constructed physical metapaths to learn spatial dependencies in the heterogeneous graph structure data, thereby obtaining the set of spatial coordinates of the reservoir.

4. The planning method for a rainwater harvesting and supplementary irrigation system for sloping farmland in arid areas according to claim 3, characterized in that, The terrain-aware encoding and decoding feature extraction network further includes a spatial transformation network. The encoding branch includes a first residual block, a second residual block, and a third residual block connected in sequence. The transposed convolution branch includes a first transposed convolutional layer, a second transposed convolutional layer, and a third transposed convolutional layer connected in sequence.

5. The planning method for a rainwater harvesting and supplementary irrigation system for sloping farmland in arid areas according to claim 4, characterized in that, The process of using the terrain-aware encoding / decoding feature extraction network to perform multi-scale feature extraction and skip-connection-based feature reuse processing on the preprocessed data to obtain a confluence potential heatmap includes: The spatial transformation network is used to spatially register the preprocessed remote sensing image data and preprocessed digital elevation model data in the preprocessed data, and then the data is stitched together according to the channel dimension to form the input feature tensor. The input feature tensor is passed sequentially through the first residual block, the second residual block, and the third residual block for downsampling and deep feature extraction, resulting in the output of shallow coding feature maps, mid-level coding feature maps, and deep coding feature maps. The preprocessed digital elevation model data is input into the terrain-aware attention module to extract target terrain factors, and the spatial attention weight map is generated based on the target terrain factors. The deep coding feature map and the spatial attention weight map are weighted element-wise to generate a deep weighted feature map. The deep weighted feature map is then input into the first transposed convolutional layer for upsampling to obtain a deep upsampled feature map. The deep upsampled feature map and the mid-layer coding feature map are concatenated along the channel dimension to obtain a deep-mid-layer concatenated feature map. The deep-to-middle layer concatenated feature map is convolved to obtain the deep-to-middle layer decoded feature map; The deep-to-middle layer decoding feature map and the spatial attention weight map are weighted element-wise to generate a middle layer weighted feature map. The middle layer weighted feature map is then input into the second transposed convolutional layer for upsampling to obtain a middle layer upsampled feature map. The middle layer upsampled feature map and the shallow layer encoding feature map are concatenated along the channel dimension to obtain a middle-to-shallow concatenated feature map. The convolutional processing of the middle-shallow layer spliced ​​feature map yields the middle-shallow layer decoded feature map; The middle-shallow decoding feature map and the spatial attention weight map are weighted element-wise to generate a shallow weighted feature map. The shallow weighted feature map is then input into the third transposed convolutional layer for upsampling to obtain a shallow upsampled feature map. After mapping the shallow upsampled feature map using an activation function layer, the confluence potential heatmap is generated.

6. The planning method for a rainwater harvesting and supplementary irrigation system for sloping farmland in arid areas according to claim 3, characterized in that, The graph attention network employs the GATv2 model. The spatial dependency learning process on the heterogeneous graph structure data, using pre-constructed physical meta-paths, yields the spatial coordinate set of the reservoir, including: According to each of the physical meta-paths, the corresponding node sequence and edge sequence are extracted from the heterogeneous graph structure data to generate the isomorphic subgraph corresponding to each of the physical meta-paths; Input the physical hydrological attribute vectors of each node in each of the isomorphic subgraphs into the GATv2 model to obtain the path-level encoding feature vectors of each node under each of the physical metapaths; The path-level encoded feature vectors belonging to the same node are fused to generate a fused feature vector for each node. The fused feature vectors of each candidate node of the reservoir are input into a multilayer fully connected regression network, which outputs the comprehensive score and predicted volume value of each candidate node of the reservoir. The spatial coordinates of the candidate nodes of the reservoir whose comprehensive scores meet the preset score threshold are selected to form the spatial coordinate set of the reservoir.

7. The planning method for a rainwater harvesting and supplementary irrigation system for sloping farmland in arid areas according to claim 3, characterized in that, The pre-planned functional areas include: slope confluence area, reservoir candidate area, and irrigation area.

8. The planning method for a rainwater harvesting and supplementary irrigation system for sloping farmland in arid areas according to claim 1, characterized in that, The training process of the pre-trained rainwater harvesting and supplemental irrigation system planning model includes: Acquire multi-source geospatial sample data, as well as labeled heatmaps of confluence potential, labeled final site selection coordinates, and labeled planned volume values ​​corresponding to the multi-source geospatial sample data; The multi-source geospatial sample data is preprocessed to obtain preprocessed sample data; The preprocessed sample data is input into the initial rainwater harvesting and irrigation system planning model for multi-round iterative first-stage training. During the first-stage training, the labeled runoff potential heat map is used as a supervision signal, and the terrain perception encoding and decoding feature extraction network in the initial rainwater harvesting and irrigation system planning model is iteratively trained using the first preset loss function. The network parameters of the trained terrain-aware encoding and decoding feature extraction network are frozen to obtain the current rainwater harvesting and irrigation system planning model. The preprocessed sample data is then input into the current rainwater harvesting and irrigation system planning model for a second stage of training with multiple iterations. The predicted comprehensive score and predicted volume value of each candidate node of the reservoir are output. The predicted comprehensive score and the predicted volume value, along with the labeled final site selection coordinates and the labeled planned volume value, are substituted into the second preset loss function for calculation. The network parameters of the current rainwater harvesting and irrigation system planning model are iteratively adjusted according to the calculation results. Repeat the above iterative training process until the preset iteration stop condition is met. The current rainwater harvesting and irrigation system planning model that meets the preset iteration stop condition is used as the pre-trained rainwater harvesting and irrigation system planning model. The preset iteration stopping condition is that the number of iterations meets the iteration number threshold or the values ​​of the first preset loss function and the second preset loss function are continuously less than the corresponding loss threshold.

9. The planning method for a rainwater harvesting and supplementary irrigation system for sloping farmland in arid areas according to claim 8, characterized in that, The first preset loss function is expressed as: ； in, This represents the value of the first preset loss function. This indicates the potential for convergence loss. This indicates the loss of structural similarity in the heatmap of confluence potential. Indicates the regularization loss. This represents the regression loss coefficient of the confluence potential. The coefficient representing the structural similarity loss in the heatmap of confluence potential. This represents the regularization loss coefficient; The second preset loss function is a weighted combination of location regression loss and volume regression loss; both the location regression loss and the volume regression loss adopt the mean squared error loss function.

10. A planning method for a rainwater harvesting and supplementary irrigation system for sloping farmland in arid areas according to claim 9, characterized in that, The regression loss of the convergence potential is expressed as: ； in, This represents the total number of pixels in the heatmap indicating confluence potential. The heatmap indicating the potential of confluence points is shown below. The annotation value of each pixel. The first element in the heatmap representing the predicted confluence potential is... The predicted value of each pixel; The structural similarity loss of the confluence potential heatmap is expressed as: ； ； in, The total number of image patches in the heatmap of confluence potential. To mark the first in the heat map of confluence potential Image blocks, To predict the confluence potential heat map, the first Image blocks, express and Structural similarity index between them and They represent and The average pixel value, and They represent and pixel variance for and covariance, Represents the first numerical stability constant. This represents the second numerical stability constant; The regularization loss is expressed as: ； in, This represents the total number of network layers to be regularized in the initial rainwater harvesting and supplemental irrigation system planning model. For the first The weight parameter matrix of the layer to be regularized network, This represents the L2 norm.

Images