A local crop yield per mu prediction method and system based on historical data analysis

Abstract

The present application belongs to the technical field of agricultural data analysis, and relates to a local crop yield per mu prediction method and system based on historical data analysis, comprising: decoupling multiple source original monitoring data to generate an environment time series tensor and a local farming record input vector; inputting the time series tensor into a model to generate an agricultural climate scenario identifier representing meteorological combination characteristics; using the identifier as an index to query, extract a hierarchical sequence, a topological edge set and a data aggregation operator to generate a topological configuration file; parsing the topological edge set to construct a running network framework; extracting a calculation instruction set according to the sequence and binding it to a node to dynamically assemble a causal calculation graph that senses climate constraints; extracting a scalar feature and injecting the farming record into the calculation graph to drive forward propagation; and converging end numerical output to obtain a final result. The present application solves the problem that a traditional prediction model has a fixed structure and is difficult to dynamically adjust internal calculation logic according to different climate scenarios to accurately reflect causal relationships.

Description

Technical Field

[0001] This invention belongs to the technical field of agricultural data analysis, and relates to a method and system for predicting local crop yield per acre based on historical data analysis. Background Technology

[0002] In modern agricultural production management, accurate forecasting of crop yield per acre is a crucial link in providing data support for agricultural policy formulation and market analysis. The core challenge in achieving high-precision yield forecasting lies in how to effectively integrate and analyze multi-dimensional heterogeneous information affecting crop growth. This information includes both large-scale climatic environmental factors such as temperature and precipitation, and micro-level agricultural management measures such as fertilization and irrigation. These two types of factors interact in a complex nonlinear manner and jointly determine the final yield.

[0003] To address the technical challenges of multi-dimensional information fusion, existing technologies have primarily developed two processing paths. The first is process-based crop growth simulation models, which rely on mathematical equations to quantitatively describe key physiological and biochemical processes such as photosynthesis, providing a strong mechanistic basis. The second is data-driven methods based on statistics and machine learning. These approaches mainly utilize historical meteorological and agricultural management records, employing algorithms such as deep learning to construct fixed mapping relationships between multimodal inputs and yield outputs, offering advantages in handling nonlinear data mapping.

[0004] The aforementioned solutions in existing technologies still have some technical limitations in practical applications. On the one hand, the high sensitivity of process-based models to localized parameters restricts their widespread application in areas with sparse data collection. On the other hand, the network structures of most machine learning methods are set as static architectures before training. This fixed network computation logic lacks sufficient dynamic adaptability when faced with significantly different climate scenarios across different years. When the external stressors that dominate yields change, the pre-set and fixed feature extraction network struggles to autonomously reconstruct its internal causal transmission paths and computational priorities, thereby reducing the model's effectiveness in targeted quantitative analysis of specific anomalous climate characteristics. Summary of the Invention

[0005] In a first aspect, the present invention provides a method for predicting local crop yield per acre based on historical data analysis, comprising the following steps:

[0006] S1. Obtain multi-source raw monitoring data of the target area, and after decoupling processing, generate independent and parallel environmental time series tensors and local agricultural record input vectors respectively.

[0007] S2. Input the environmental time series tensor into the pre-set model to generate agricultural climate scenario labels that characterize the meteorological combination at the spatiotemporal joint scale.

[0008] S3. Using agricultural climate scenario identifiers as indexes, query the preset mapping lookup table, extract the target node hierarchy sequence, topology connection edge set, and data aggregation operator set that match the current scenario, and generate a network topology configuration file.

[0009] S4. Parse the set of topology connection edges in the network topology configuration file and construct a running network framework that matches the current scenario. The running network framework contains computing nodes to be configured.

[0010] S5. Based on the target node hierarchy sequence, extract the corresponding computation instruction set from the pre-configured program library, load and bind it to the computation node to be configured, and dynamically assemble it to form a causal computation graph.

[0011] S6. Extract scalar environmental features based on the data aggregation operator set, and inject the scalar environmental features and the local agricultural record input vector into the causal computation graph to drive forward propagation to calculate the stage-based yield prediction values ​​corresponding to different computation nodes.

[0012] S7. Aggregate the predicted total value output from the end nodes of the causal computation graph and output the final predicted yield per acre.

[0013] A further aspect of the present invention, step S1, includes the following steps:

[0014] Acquire daily grid meteorological elements covering the target area at a preset cycle, and multispectral satellite remote sensing images at a preset time step, and integrate them into a continuous environmental data stream;

[0015] Extract unstructured micro-level agricultural operation history information corresponding to the target plot and transform it into a standardized local agricultural record input vector based on natural language processing algorithms;

[0016] The transformed sparse vectors are segmented and aggregated according to the key growth period or preset time window, and then normalized by dividing by the physical area of ​​the target plot to generate local agricultural record input vectors that independently represent agricultural input per unit area. At the same time, the extracted environmental data are organized into independent environmental time series tensors.

[0017] A further aspect of the present invention, step S2, includes the following steps:

[0018] Directly call the extracted environment time series tensor;

[0019] A pre-built spatiotemporal coding and decoding network is used to map the environmental temporal tensor to the latent feature space and extract the spatiotemporal joint feature vector;

[0020] Density clustering algorithm is used to perform unsupervised classification of spatiotemporal joint feature vectors to output agricultural climate scenario labels.

[0021] A further aspect of the present invention, step S3, includes the following steps:

[0022] Obtain the call address labels of each physiological response calculation instruction set in the pre-configured program library;

[0023] Based on the operation of the query mapping lookup table, extract the target node hierarchy sequence containing the call address label;

[0024] The target node hierarchical sequence, topology connection edge set, and data aggregation operator set are encapsulated and serialized, and packaged to generate a network topology configuration file.

[0025] A further aspect of the present invention, step S4, includes the following steps:

[0026] Receive network topology configuration file;

[0027] Disassemble the network topology configuration file, extract the set of topology connection edges, and thereby determine the data flow direction and combination level;

[0028] Based on the set of topological connecting edges, the system requests address space, establishes connections between nodes, and builds a mesh-structured operating network framework.

[0029] A further aspect of the present invention, step S5, includes the following steps:

[0030] Traverse and parse the call address labels in the hierarchical sequence of the target node;

[0031] The specific physiological response calculation instruction set is extracted from the pre-configured program library according to the call address number and assigned to the function pointer of the corresponding computing node to be configured in the running network framework.

[0032] Connecting all the computing nodes that have loaded the instruction set forms a complete directed acyclic execution loop from local input to top-level prediction output, thus constituting a causal computation graph.

[0033] A further aspect of the present invention, step S6, includes the following steps:

[0034] The system retrieves the temporarily stored local agricultural record input vector from the system storage area, and extracts scalar environmental features from the environmental time series tensor based on preset operator rules.

[0035] The local agricultural record input vector and scalar environmental features are respectively assigned to the bottom initial nodes of the corresponding functions in the causal computation graph as start-up variables;

[0036] Following the topological order of the directed acyclic execution loop, the data flow is directed to penetrate each layer of computing nodes in sequence, outputting phased production forecast values ​​carrying climate constraints.

[0037] A further aspect of the present invention, step S7, includes the following steps:

[0038] Locate the convergence node of the causal computation graph and collect the total predicted value after calculation by all nodes;

[0039] Based on the preset unit conversion factor, the unit conversion operation is performed on the predicted total value;

[0040] Output the final yield prediction result with completed unit conversion, and release the memory space occupied by the causal computation graph.

[0041] A further aspect of this invention involves extracting unstructured historical information on micro-level agricultural operations corresponding to the target plot and converting it into a standardized local agricultural record input vector based on a natural language processing algorithm, including the following steps:

[0042] The named entity recognition task extracts the types and specific amounts of agricultural activities from the historical information of micro-level agricultural operations.

[0043] Each agricultural activity with a specific usage is encoded and mapped to a specific component position of a preset feature dimension, and the remaining components are filled with zero values ​​to generate a sparse vector with timestamps.

[0044] Secondly, the present invention provides a local crop yield prediction system based on historical data analysis, comprising the following modules:

[0045] The multi-source data processing module is used to acquire multi-source raw monitoring data of the target area, and after decoupling processing, it generates independent and parallel environmental time series tensors and local agricultural record input vectors respectively.

[0046] The climate scenario identification module is used to input environmental time series tensors into a pre-set model and generate agricultural climate scenario labels that characterize meteorological combination features at a spatiotemporal joint scale.

[0047] The network topology configuration module is used to use agricultural climate scenario identifiers as indexes to query a preset mapping lookup table, extract the target node hierarchy sequence, topology connection edge set, and data aggregation operator set that match the current scenario, and generate a network topology configuration file.

[0048] The network framework building module is used to parse the set of topology connection edges in the network topology configuration file and build a running network framework that matches the current scenario. The running network framework contains computing nodes to be configured.

[0049] The causal computation graph assembly module is used to extract the corresponding computation instruction set from the pre-configured program library according to the target node hierarchy sequence, load and bind it to the computing node to be configured, and dynamically assemble it into a causal computation graph.

[0050] The forward propagation computation module is used to extract scalar environmental features based on the data aggregation operator set, and inject the scalar environmental features and the local agricultural record input vector into the causal computation graph to drive forward propagation to calculate the stage-based yield prediction values ​​corresponding to different computation nodes.

[0051] The yield prediction output module is used to aggregate the total predicted values ​​output by the end nodes of the causal computation graph and output the final yield prediction result per acre.

[0052] In summary, the present invention has the following beneficial technical effects:

[0053] 1. This invention identifies agricultural climate scenario markers for specific growing seasons using a pre-built model and uses these as indexes to dynamically extract target node hierarchical sequences, overcoming the application limitations of traditional prediction systems with static and fixed network structures. This mechanism enables the system to adaptively retrieve and assemble corresponding physiological response calculation instruction sets from a pre-configured program library based on the dominant constraints of the external environment, constructing a causal computation graph. This site-specific network assembly method couples the model's underlying computational logic with the dominant environmental factors of a specific year, improving the model's computational flexibility and environmental adaptability in dealing with complex and abnormal climate changes.

[0054] 2. This invention decouples multi-source raw monitoring data, separating the input channels of environmental time-series tensors and micro-level local agricultural records in the model's operating mechanism. The system uses climate characteristics as prior boundary conditions to determine the topology and functional node types of the entire operating network framework; then, it treats micro-level agricultural characteristics as optimization variables operating within these constraints, injecting them into the pre-defined computational graph to drive forward propagation. This feature architecture defines the weight hierarchy of environmental constraints and human management in the crop growth model, helping to more objectively simulate the response mechanism of agricultural operations under specific meteorological backgrounds, and providing a quantitative path for evaluating the effectiveness of different management measures under the same climatic conditions.

[0055] 3. This invention employs discrete agricultural climate scenario identifiers and independent physiological response mathematical function modules to construct a directed acyclic execution loop with a certain mechanistic basis. The system decomposes the complex crop response process into multiple clearly defined computational nodes, and the data flow path of the causal computation graph directly maps the correlation chain of "management-environment-yield". While obtaining the final predicted value, it allows technicians to trace the specific functional nodes through which the data flows, such as the water stress loss node or nutrient use efficiency node, and then extract and analyze the impact coefficients of specific environmental variables on yield separately. This design alleviates the technical defects of traditional black-box models where the internal logic is not visible, and enhances the logical coherence of the deduction process and the credibility of the prediction results. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings are used to provide a further understanding of the present invention.

[0057] Figure 1 A flowchart illustrating an embodiment of this application is disclosed.

[0058] Figure 2 Structural schematic diagrams of embodiments of this application are disclosed. Detailed Implementation

[0059] The following is in conjunction with the appendix Figures 1-2 A preferred description of the present invention is provided below.

[0060] See attached document Figure 1 This invention proposes a method for predicting local crop yield per acre based on historical data analysis, comprising the following steps:

[0061] S1. Obtain multi-source raw monitoring data of the target area, and after decoupling processing, generate independent and parallel environmental time series tensors and local agricultural record input vectors respectively.

[0062] S2. Input the environmental time series tensor into the pre-set model to generate agricultural climate scenario labels that characterize the meteorological combination at the spatiotemporal joint scale.

[0063] S3. Using agricultural climate scenario identifiers as indexes, query the preset mapping lookup table, extract the target node hierarchy sequence, topology connection edge set, and data aggregation operator set that match the current scenario, and generate a network topology configuration file.

[0064] S4. Parse the set of topology connection edges in the network topology configuration file and construct a running network framework that matches the current scenario. The running network framework contains computing nodes to be configured.

[0065] S5. Based on the target node hierarchy sequence, extract the corresponding computation instruction set from the pre-configured program library, load and bind it to the computation node to be configured, and dynamically assemble it to form a causal computation graph.

[0066] S6. Extract scalar environmental features based on the data aggregation operator set, and inject the scalar environmental features and the local agricultural record input vector into the causal computation graph to drive forward propagation to calculate the stage-based yield prediction values ​​corresponding to different computation nodes.

[0067] S7. Aggregate the predicted total value output from the end nodes of the causal computation graph and output the final predicted yield per acre.

[0068] In one embodiment of the present invention, step S1 includes the following steps:

[0069] Acquire daily grid meteorological elements covering the target area at a preset cycle, and multispectral satellite remote sensing images at a preset time step, and integrate them into a continuous environmental data stream;

[0070] Extract unstructured micro-level agricultural operation history information corresponding to the target plot and transform it into a standardized local agricultural record input vector based on natural language processing algorithms;

[0071] The transformed sparse vectors are segmented and aggregated according to the key growth period or preset time window, and then normalized by dividing by the physical area of ​​the target plot to generate local agricultural record input vectors that independently represent agricultural input per unit area. At the same time, the extracted environmental data are organized into independent environmental time series tensors.

[0072] The process involves extracting unstructured historical information on micro-level agricultural operations corresponding to the target plot and converting it into a standardized local agricultural record input vector based on natural language processing algorithms. This includes the following steps:

[0073] The named entity recognition task extracts the types and specific amounts of agricultural activities from the historical information of micro-level agricultural operations.

[0074] Each agricultural activity with a specific usage is encoded and mapped to a specific component position of a preset feature dimension, and the remaining components are filled with zero values ​​to generate a sparse vector with timestamps.

[0075] Specifically, this is performed by a data processing module deployed on a cloud server or local computing terminal. This data processing module periodically accesses public meteorological databases and commercial or public remote sensing data platforms via network interfaces to acquire environmental monitoring data covering the geographical area of ​​the target site according to a preset cycle.

[0076] For meteorological elements, the system automatically acquires and downloads gridded meteorological reanalysis data within a specified latitude and longitude range on a daily basis. This data includes key variables such as daily average temperature, cumulative precipitation, sunshine hours, and relative humidity.

[0077] For remote sensing images, the system uses ten-day periods as the time unit to filter and download multispectral satellite remote sensing images covering the target area that are cloudless or have low cloud cover.

[0078] The system preprocesses these raw data, including spatial interpolation of meteorological data to match the center point coordinates of the target plot, radiometric calibration and atmospheric correction of remote sensing imagery, and extraction of spectral indices related to crop growth. All preprocessed data are organized into a multidimensional time series matrix with a daily time step, forming a continuous environmental data stream.

[0079] The data processing module reads historical information about micro-level agricultural operations on the target plot from local storage. This information is typically in unstructured formats such as text logs and spreadsheets. The system calls a natural language processing algorithm module based on a pre-trained language model to parse these text records. This algorithm module performs named entity recognition, automatically extracting key information such as the type of agricultural activity, execution time, material name, and specific dosage. For example, it identifies time entities, agricultural activity type entities, and material entities from records of foliar fertilizer application in early June.

[0080] Next, the system standardizes this extracted structured information, for example, converting vague time descriptions into precise dates, unifying material usage into international standard units, and finally encoding each agricultural operation into a fixed-dimensional numerical vector. The value of this vector in the corresponding agricultural activity dimension is the specific usage, while other dimensions are zero. All agricultural records are processed to form a set of sparse vectors with timestamps.

[0081] Finally, the data processing module directly outputs the continuous environmental data stream as an environmental time-series tensor. Simultaneously, the system aggregates and summarizes a set of timestamped sparse vectors generated by the system according to the key growth stages of the crop or preset time windows, segmenting them along the time dimension. For example, it sums the amounts of fertilizer and irrigation applied in different stages, obtains the actual measured physical area of ​​the target plot, and normalizes the scale by dividing the absolute total amount of each agricultural activity after segmental summation by this physical area, generating a local agricultural record input vector representing the agricultural input per unit area at different stages. The environmental time-series tensor and the local agricultural record input vector are output in parallel as two independent data channels, providing a multi-dimensional and decoupled data foundation for subsequent prediction models. It should be noted that this decoupling process refers to separating uncontrollable macro-meteorological factors and controllable micro-agricultural operations into two independent parallel input paths in the data flow architecture, rather than using statistical algorithms to remove their physical correlation. This aims to prevent the model from confusing the causal boundary between environmental constraints and human intervention in the initial feature extraction stage.

[0082] The data source for daily gridded meteorological elements can be a dataset from a regional weather forecasting center. Multispectral satellite remote sensing imagery is preferably selected from satellite red and near-infrared band data for calculating the Normalized Difference Vegetation Index (NDVI), which is calculated as follows: ,in Represents reflectivity in the near-infrared band. Represents the reflectivity of the red light band. Natural language processing algorithms can employ named entity recognition models based on the BERT architecture and fine-tuned using agricultural corpora. The dimension of the local agricultural record input vector is pre-defined according to the types of agricultural activities covered; for example, a vector of length... The vectors can represent irrigation amount, nitrogen fertilizer application amount, phosphate fertilizer application amount, etc., in sequence. Different agricultural operations. The environmental time series tensor is mathematically represented as a matrix. ,in The total number of days in the growing season. This represents the number of feature dimensions of the environment. The local agricultural record input vector is mathematically represented as a one-dimensional vector. ,in The preset number of dimensions for agricultural activities to be covered.

[0083] In one example of this invention, the system first processes data for the selected date, June 15th. The system retrieves grid data covering the coordinate point from the database, interpreting it to show an average temperature of 26.5 degrees Celsius and cumulative precipitation of 2.0 millimeters. Simultaneously, the system searches remote sensing data archives and finds that the most recent cloudless satellite image was taken on June 12th. The system then extracts the pixel reflectance corresponding to the plot, calculates the Normalized Difference Vegetation Index (NDVI) value to be 0.72, and uses this value as the remote sensing feature for June 15th. Thus, the environmental data components are [26.5, 2.0, 0.72]. At the same time, the system reads the agricultural log file and finds a record stating "June 15th, irrigation 20 cubic meters". The natural language processing module parses this record and outputs structured information {Time: 2023-06-15, Agricultural activity: Irrigation, Amount: 20 cubic meters}. The preset sparse vector dimension is 3, corresponding to [absolute irrigation amount, absolute nitrogen fertilizer amount, absolute herbicide amount], so the sparse vector for that day is encoded as [20, 0, 0]. The system will repeat the data acquisition process for each day during the growing season, ultimately forming a 122-dimensional vector. 3. Environmental temporal tensor. Simultaneously, the system traverses the agricultural logs throughout the entire growing season, aggregating and summing all extracted micro-agricultural operations in segments according to preset time windows. Simultaneously, it obtains the actual measured physical area of ​​the target plot and divides the summed absolute total by this area. The system ultimately outputs an independent one-dimensional local agricultural record input vector, assumed to have values ​​of [0.075, 0.009, 0], representing the total irrigation amount per unit area during the vegetative growth period of the plot (0.075 cubic meters / m²), the total nitrogen application per unit area (0.009 kg / m²), and the herbicide application amount (0 ml / m²), respectively.

[0084] In one embodiment of the present invention, step S2 includes the following steps:

[0085] Directly call the extracted environment time series tensor;

[0086] A pre-built spatiotemporal coding and decoding network is used to map the environmental temporal tensor to the latent feature space and extract the spatiotemporal joint feature vector;

[0087] Density clustering algorithm is used to perform unsupervised classification of spatiotemporal joint feature vectors to output agricultural climate scenario labels.

[0088] Calculate the historical sample set -Nearest neighbor distance distribution is used to determine the corresponding inflection point as the neighborhood radius, and clustering operation is performed in combination with the preset minimum number of samples of core objects;

[0089] Each cluster obtained from the calculation is assigned a unique discrete integer value as the output representation of the comprehensive classification label.

[0090] Specifically, upon receiving the decoupled data stream output in parallel from the previous step, the model's core computational engine directly calls the environmental time-series tensor. The system inputs this tensor into a pre-defined spatiotemporal encoding and decoding network. This network typically employs a convolutional long short-term memory (LSTM) encoder structure, designed to learn and capture the complex interaction patterns of different meteorological variables in the temporal and spatial dimensions from unlabeled historical data.

[0091] The encoder compresses and maps the environmental data stream throughout the growing season, projecting it into a latent feature space and ultimately outputting a fixed-length vector. This vector is the spatiotemporal joint feature, whose internal dimensions comprehensively represent the climate rhythm, extreme event occurrences, and crop growth conditions throughout the growing season.

[0092] The system inputs the extracted spatiotemporal joint feature vector into a pre-trained density clustering algorithm model. For example, the system can employ a density-based spatial clustering algorithm, which performs unsupervised classification based on the distribution density of spatiotemporal joint features in the feature space. The spatiotemporal joint feature vector is assigned to the nearest cluster that meets the density requirements, and each cluster is pre-labeled with a unique discrete identifier. The algorithm ultimately outputs the identifier of the cluster to which the vector belongs; this identifier is the agricultural climate scenario identifier, serving as a comprehensive classification label for the environment and providing a quantified scenario definition for subsequent steps.

[0093] Spatiotemporal encoder-decoder networks are autoencoder models that have been unsupervised pre-trained on a large amount of historical environmental data. The encoder part consists of multiple layers of ConvLSTM units stacked together, while the decoder part has a structure symmetrical to the encoder. The training objective is to minimize the difference between the original input and the decoder reconstructed output, thereby forcing the encoder to learn the most representative and compact representation of the data.

[0094] The spatiotemporal joint features are the embedding vectors generated by the network encoder in the latent space, with dimensions typically set between 64 and 256, such as 128-dimensional. The density clustering algorithm used is DBSCAN, with its key hyperparameter being the neighborhood radius. and the minimum number of samples for the core object It is determined based on the statistical distribution of the historical feature vector set. The value can be calculated from historical samples. - The inflection point of the nearest neighbor distance distribution is used to set the value. It is usually set to an integer between 3 and 10 based on experience to balance the stability of clustering and its robustness to noise.

[0095] Agricultural climate scenario identifiers are integer or enumeration values. For example, identifier 0 may represent a standard bumper year, identifier 1 may represent a year of spring drought and summer flooding, and identifier 2 may represent a year of sustained high temperature and drought.

[0096] In one example of this invention, the system acquires an environmental time-series tensor covering a certain period, processed in a previous step. This tensor is a matrix with a time length of 122 days and a feature dimension of 3, containing [daily average temperature, daily cumulative precipitation, and normalized difference vegetation index]. This matrix is ​​input into a pre-set ConvLSTM encoder. After forward propagation computation, the encoder compresses the entire time-series data and outputs a 128-dimensional spatiotemporal joint feature vector, assuming its value is... Subsequently, this vector The data is fed into the DBSCAN clustering model, whose parameters are set to neighborhood radius. Minimum number of samples for core objects The model is calculated internally. The distance to the cluster core points formed with other historical data. Based on the calculation results, determine... Those that fall into cluster 2 Within the neighborhood. Therefore, the final agricultural climate scenario identifier output by the system is an integer 2, which represents that the overall environmental characteristics of the 2023 growing season are mid-term high-temperature stress type.

[0097] In one embodiment of the present invention, step S3 includes the following steps:

[0098] Obtain the call address labels of each physiological response calculation instruction set in the pre-configured program library;

[0099] Based on the operation of the query mapping lookup table, extract the target node hierarchy sequence containing the call address label;

[0100] The target node hierarchical sequence, topology connection edge set, and data aggregation operator set are encapsulated and serialized, and packaged to generate a network topology configuration file.

[0101] Specifically, upon receiving the agricultural climate scenario identifier generated in the previous step, the system's network topology configuration module executes a static matching mechanism to generate a set of instructions for governing the construction of downstream networks.

[0102] The network topology configuration module obtains the call address labels of each physiological response calculation instruction set in the pre-configured program library; the pre-configured program library contains basic mathematical function modules for simulating crop physiological response processes.

[0103] Using agroclimate scenario identifiers as unique retrieval index pointers, queries are performed on a pre-built and fixed mapping lookup table. This lookup table associates each possible agroclimate scenario identifier with a specific set of network construction schemes.

[0104] Through this query, the system can accurately extract the target node hierarchy sequence, topological connection edge set, and data aggregation operator set that match the current scenario identifier. The data aggregation operator set defines how to extract the scalar features required for specific nodes from a long sequence of environments, such as extracting the average daily maximum temperature at a specific growth stage. These elements collectively define the data flow and structural blueprint of the computational network.

[0105] To ensure the atomicity and integrity of data transmission, the system encapsulates the retrieved target node hierarchy sequence, topology connection edge set, and data aggregation operator set into a single communication frame. After serialization, the frame is packaged to generate a network topology configuration file, which is then sent to the downstream dynamic network building module.

[0106] The pre-configured library is a software library that stores compiled functions or interpretable scripts. The functions include, for example, the Michaelis-Menten equation model describing the change of photosynthetic rate with light intensity, and nonlinear functions simulating the relationship between root water absorption and soil moisture content.

[0107] The call address label is a string or hexadecimal address code that uniquely identifies and calls a function in the library. The mapping lookup table is a static data structure implemented based on a hash table or key-value database, where the key is an integer value representing an agricultural climate scenario, and the value is a structure containing node and edge definitions. The target node hierarchy sequence is an ordered list defining all functional nodes in the computation graph. Each element in the list contains at least one node ID and the call address label of the function it needs to load.

[0108] The set of topological connecting edges defines the directed paths through which data flows between these nodes; each element in the list typically consists of a source node ID and a destination node ID. Data aggregation subsets are a series of pre-defined data slices and statistical algorithms tailored to specific input interfaces, such as calculating the average, cumulative, or maximum value for a particular month.

[0109] A network topology configuration file is a self-contained data format, such as a JSON object or a Protocol Buffers message body, that ensures the integrity and consistency of the network topology definition.

[0110] In one example of this invention, the agricultural climate scenario identifier received by the network topology configuration module is an integer 2, representing a mid-term high-temperature stress type. The module first scans its local pre-configured library to confirm the existence of functions with identifiers such as F_TempStress_YieldLoss (for calculating yield loss rate under high-temperature stress) and F_BaseYield_Calc (for calculating base yield based on other normal parameters), along with their call address labels. The module uses identifier 2 as the key to perform a search in the mapping lookup table. Suppose the lookup table is: {..., 1:{...}, 2:{nodes:[{id:'N1', label:'F_BaseYield_Calc'}, {id:'N2', label:'F_TempStress_YieldLoss'}, {id:'N_out', label:'F_Combine'}], edges:[{from:'Input_Micro', to:'N1'}, {from:'Input_Macro_T', to:'N2'}, {from:'N1', to:'N_out'}, {from:'N2', to:'N_out'}]}, 3:{...}, ...}. Based on this, the system extracts the matching target node hierarchy sequence, the set of topological connecting edges, and a data aggregation subset containing the rule "extract the average temperature from early July to mid-August for the Input_Macro_T interface". Finally, the system encapsulates these collections into a unified JSON object, namely the network topology configuration file, such as: "{"scene_id":2,"nodes":[...],"edges":[...],"operators":{"Input_Macro_T":"AVG_TEMP_JUL_AUG"}}", and sends it to the subsequent processing module.

[0111] In one embodiment of the present invention, step S4 includes the following steps:

[0112] Receive network topology configuration file;

[0113] Disassemble the network topology configuration file, extract the set of topology connection edges, and thereby determine the data flow direction and combination level;

[0114] Based on the set of topological connecting edges, the system requests address space, establishes connections between nodes, and builds a mesh-structured operating network framework.

[0115] After the network topology configuration file is generated in the preceding steps, the system's network framework building module is awakened from its initial standby state and receives the configuration file. Its internal parser module immediately disassembles the configuration file. The parser deserializes the data packet content according to a predefined data format, such as JSON or XML, and extracts key fields containing network topology information.

[0116] Specifically, the topology edge set is located and retrieved first. This set details the dependencies and data flow between all computing nodes. By analyzing these connections, the network framework building module can determine the overall data flow direction and the hierarchical combination of different computing nodes. Based on the network depth and breadth requirements implied by the parsed topology edge set, such as the number of parallel nodes and the length of the longest computation path, the network framework building module allocates the corresponding memory address space.

[0117] After obtaining the address space, the network framework building module begins instantiating network node objects in this memory area. Based on the instructions of the topology connection edge set, it sets reference pointers or link relationships between nodes, thereby constructing a running network framework with a complete mesh structure and full connectivity logic. When this framework is completed, its nodes are all empty placeholders without loaded specific computation instructions, but the connection relationships between nodes are fully determined, preparing for subsequent dynamic function assembly.

[0118] The network framework building block is a dedicated runtime engine designed according to an event-driven architecture, ensuring low-latency responses to incoming configuration commands. The initial standby state is a low-resource-consumption polling or blocking wait state. The set of topology edges is typically represented in data structures as a list of multiple tuples, each in the form of (source node ID, destination node ID), clearly defining a directed edge. Network depth refers to the number of nodes along the longest path from any input node to the final output node, while network breadth refers to the maximum number of parallel nodes in the network topology. The running network framework is represented in memory as a directed acyclic graph data structure, consisting of a series of node objects and pointers to these objects, ensuring unidirectional data flow during processing.

[0119] It should be noted that the running network framework is essentially the structural skeleton of the subsequent causal computation graph before the specific computation functions are assembled. Both refer to the state of the same physical object entity at different configuration stages at the system's underlying level, and are not two independent processing systems.

[0120] In one example of this invention, the network framework building module receives a network topology configuration file. The parser of the network framework building module first disassembles the JSON string, extracting two arrays: `nodes` and `edges`. Based on the `nodes` array, the network framework building module identifies the need to create three core computing nodes N1, N2, and N_out, as well as two input interfaces, `Input_Micro` and `Input_Macro_T`. Subsequently, the network framework building module analyzes the `edges` array, which is the set of topology connection edges. It processes the connection instructions one by one: the first instruction, "{from:'Input_Micro', to:'N1'}", establishes a data flow channel in memory from the input interface `Input_Micro` to node N1; the second instruction, "{from:'Input_Macro_T', to:'N2'}", establishes a channel from `Input_Macro_T` to N2; subsequent instructions further connect the outputs of nodes N1 and N2 to the input of node N_out. During this process, the system requests sufficient memory space from the operating system to accommodate the five node objects and their associated pointers, and completes the setting of all pointers. Finally, the running network framework, instantiated in memory, with a defined topology but node functionality yet to be filled, is completed. Its structure is presented as a computation graph with two parallel branches converging at a point.

[0121] In one embodiment of the present invention, step S5 includes the following steps:

[0122] Traverse and parse the call address labels in the hierarchical sequence of the target node;

[0123] The specific physiological response calculation instruction set is extracted from the pre-configured program library according to the call address number and assigned to the function pointer of the corresponding computing node to be configured in the running network framework.

[0124] Connecting all the computing nodes that have loaded the instruction set forms a complete directed acyclic execution loop from local input to top-level prediction output, thus constituting a causal computation graph.

[0125] Specifically, once the network framework is built, the system will dynamically assemble the functions of the framework according to the node instructions in the network topology configuration file. This process is executed by the causal computation graph assembly module in the network framework building module.

[0126] First, the module initiates a traversal process, parsing the target node hierarchy sequence embedded in the network topology configuration file one by one. During this traversal, for each node definition in the sequence, the module extracts its unique node ID and the associated call address label.

[0127] Subsequently, the module uses the extracted call address label as the query key to initiate a call request to the pre-configured library. The pre-configured library retrieves and returns the memory entry address or function handle of the corresponding specific physiological response calculation instruction set based on the label.

[0128] Upon acquiring the instruction set, the causal computation graph assembly module immediately uses the previously extracted node IDs to locate the corresponding computation node to be configured within the constructed running network framework. Next, the module performs a binding operation, which essentially assigns the entry address of the acquired physiological response computation instruction set to a function pointer member variable within the computation node to be configured, thereby associating the abstract computational logic with the concrete framework node entity. This assembly process is repeated for all nodes in the target node hierarchy until all computation nodes to be configured have been loaded and bound with their dedicated computational instructions.

[0129] Once the last node completes instruction binding, the mesh network framework is fully activated. Since the structural connections between nodes were established during the framework construction phase, the activated data flow logic path also becomes fully effective, forming a complete directed acyclic execution loop from local input nodes to top-level prediction output nodes. This final, fully functional execution entity is the causal computation graph capable of sensing climate constraints.

[0130] The compute node to be configured is an object instantiated in memory, containing attributes such as a node ID, an input data buffer, an output data buffer, and function pointers (initially null) pointing to specific computational logic. The physiological response computation instruction set is a compiled, independently callable function module whose input / output parameter format matches the buffer design of the compute node to be configured. Loading and binding is a software-level dynamic linking process that associates function memory addresses with object methods at runtime, rather than static linking at compile time. The personalization of the causal computation graph lies in its topology and node functionality being entirely determined by specific agricultural climate scenarios, while causality is reflected in the directed connections between nodes directly simulating the causal chain of how agricultural inputs and environmental factors influence the final yield step by step.

[0131] In one example of this invention, following the example from the previous step, the causal computation graph assembly module begins processing a network topology configuration file containing information such as {id:'N1', label:'F_BaseYield_Calc'}, {id:'N2', label:'F_TempStress_YieldLoss'}, and {id:'N_out', label:'F_Combine'}. First, the module processes the first node, extracting the ID N1 and the call address label F_BaseYield_Calc. The module uses F_BaseYield_Calc to query the pre-configured library, which returns the memory address of the function, for example, 0x00A10F01. Subsequently, the module finds the compute node to be configured with ID N1 in the running network framework and sets its internal function pointer to 0x00A10F01. Next, the module processes the second node, extracting N2 and the label F_TempStress_YieldLoss, obtaining its memory address 0x00B21C05 from the library, and binding it to the node with ID N2. Finally, the module processes the third node, extracting N_out and the label F_Combine, obtaining its memory address 0x00C32D08, and binding it to the node with ID N_out. At this point, the functionality of all nodes in the framework is configured. The original structural framework automatically forms a causal computation graph because all nodes have been assigned explicit computation instructions and are connected according to a preset topology. This graph is now ready to receive input data and perform computations along paths flowing from Input_Micro and Input_Macro_T to N1 and N2 respectively, ultimately converging at N_out.

[0132] In one embodiment of the present invention, step S6 includes the following steps:

[0133] The system retrieves the temporarily stored local agricultural record input vector from the system storage area, and extracts scalar environmental features from the environmental time series tensor based on preset operator rules.

[0134] The local agricultural record input vector and scalar environmental features are respectively assigned to the bottom initial nodes of the corresponding functions in the causal computation graph as start-up variables;

[0135] Following the topological order of the directed acyclic execution loop, the data flow is directed to penetrate each layer of computing nodes in sequence, outputting phased production forecast values ​​carrying climate constraints.

[0136] Specifically, after the causal computation graph is fully assembled and instantiated, the system's forward propagation computation module reads multi-dimensional agricultural prediction input sources from the system storage area:

[0137] For micro-branches, the system directly reads the previously generated independent and aggregated local agricultural record input vectors and assigns them to the lowest-level micro-initial nodes of the causal computation graph as first-class initiation variables.

[0138] For branches, the system parses the set of data aggregation operators carried in the network topology configuration file, performs dynamic slicing and statistical calculations on the long-time environmental time series tensor according to the operator instructions, extracts the scalar environmental features required by specific computing nodes, and assigns them to the corresponding initial nodes as the second type of start-up variables.

[0139] Next, the forward propagation computation module formally triggers the computation process within the graph. It directs the data flow from bottom to top according to the established topology. The data flow first enters the first-layer computation nodes, where the functions encapsulated within these nodes use the input micro-management variables, or combine them with preset local annual climate benchmark parameters, to calculate preliminary physiological effects or basic outputs.

[0140] Once the first-layer node completes its calculation, its output is passed as input to the next layer node. As the data flow sequentially traverses each layer of computational nodes along a predetermined route, it interacts with specific functions determined by the climate scenario—that is, computational nodes carrying climate-constrained characteristics. These nodes, based on their bound physiological response models, generate phased weighted or suppressed computational results on the data flow. For example, a high-temperature stress node might generate a non-linear reduction factor for the underlying yield data based on the input average temperature. This process repeats until the data flow reaches the top-level convergence node. In the convergence node, a combination function finally integrates the computational results of all upstream branches, for example, through a formula... Calculations are performed to determine the phased yield forecasts for agricultural activities under the current climate conditions.

[0141] The computation performed at the terminal convergence node of the causal computation graph can be represented by the following exemplary formula:

[0142]

[0143] in, This represents the final revised expected output. It is the basic yield potential calculated by one or more nodes based on the local agricultural record input vector and the annual environmental baseline parameters. It reflects the effect of management measures under ideal or average conditions. The annual environmental baseline parameters are pre-stored in the system and represent the average meteorological constant values ​​of the target plot over many years during this growth period, such as the historical average temperature, which are used to establish a baseline without extreme climate fluctuations. The yield loss coefficient is calculated by one or more other nodes based on the climatic characteristics of the growing season. This coefficient is a dimensionless scalar between 0 and 1, reflecting the inhibitory effect of climatic constraints. In a specific example, the yield loss coefficient... The calculation model is a linear penalty function based on a temperature threshold: ,in This refers to the actual scalar environmental characteristics extracted so far, such as the average daily temperature. This is the preset critical threshold for high-temperature stress. is the crop sensitivity coefficient constant.

[0144] This formula ensures that calculation results from different sources can be coupled in a physically meaningful way.

[0145] It should be noted that for agricultural climate scenarios with suitable or favorable growth conditions, the nodal function can also be a positive gain model, for example... ,in This is the climate gain coefficient. Similarly, the climate gain coefficient... Calculations can be made based on the positive deviation of suitable accumulated temperature or precipitation, for example. The introduction of this coefficient aims to quantify the actual increase in production brought about by favorable climates such as years with abundant water.

[0146] The forward propagation computation module is a core runtime component of the network framework building module. It is responsible for scheduling node execution according to the topological order of the graph and managing data transfer between nodes. The initiation variable refers to the data value injected into the initial node of the computation graph that has no upstream input nodes. The phased weighted or suppressed computation result is the output of any intermediate node in the graph, representing a transformation of the data stream. This transformation logic is determined by the agricultural climate scenario identifier and therefore carries climate constraint information.

[0147] It should be further noted that in the scalar environment feature extraction method, the data aggregation operator set includes at least one pre-defined statistical function rule. For example, for the input environment time series tensor... ,in In terms of time dimension, For environmental features, a typical operator rule AVG_TEMP_JUL_AUG can be defined as:

[0148]

[0149] in, and These represent time indices, such as the time indices for July 1st and August 31st. Indicates time The system extracts the daily average temperature characteristic values. By executing these operator rules, the system can accurately extract the corresponding scalar environmental features from the original high-dimensional time series tensor and input them into the downstream computational model.

[0150] In one example of this invention, the execution flow of the causal computation graph is as follows: First, the forward propagation computation module directly obtains the local agricultural record input vector [0.075, 0.009, 0] prepared in stage T1 from the system and assigns it to the micro-input node Input_Micro of the graph. Simultaneously, the system parses the operator "AVG_TEMP_JUL_AUG" for the environmental input and automatically extracts the data segment from early July to mid-August from the 122-day environmental time-series tensor according to this rule. The system calculates the true average temperature for this period to be 34 degrees Celsius and then assigns this exact scalar value of 34 to the environmental input node Input_Macro_T. The data stream begins to flow, and the aggregated micro-management feature vector enters the node with ID N1. The F_BaseYield_Calc function bound to this node calculates the basic yield potential per unit standard area based on these unit area input parameters. for Meanwhile, the temperature data of 34 degrees Celsius is sent to the node with ID N2, and its bound function F_TempStress_YieldLoss calculates a yield loss coefficient. It is 0.15. Finally, from N1 The two interim calculation results, 0.15 from N2, are simultaneously input into the pool node with ID N_out. This node executes the F_Combine function, which performs the calculation. Therefore, the final result of the forward propagation of the causal computation graph is: This value represents the expected output per unit area of ​​agricultural management for the plot after taking into account the effects of medium-term high-temperature stress.

[0151] In one embodiment of the present invention, step S7 includes the following steps:

[0152] Locate the convergence node of the causal computation graph and collect the total predicted value after calculation by all nodes;

[0153] Based on the preset unit conversion factor, the unit conversion operation is performed on the predicted total value;

[0154] Output the final yield prediction result with completed unit conversion, and release the memory space occupied by the causal computation graph.

[0155] After the forward propagation calculation is completed and the total output forecast is generated, the system executes the final output forecast output and resource cleanup procedure.

[0156] First, the system's yield prediction output module acquires the output terminal node data of the causal computation graph, which is the last-level vertex of the causal computation graph. This vertex is the only node in the graph with an out-degree of zero, and it is responsible for aggregating the data streams from all upstream computation branches. The system collects the total predicted value after all climate adaptability reduction procedures from the output buffer of this vertex. This total predicted value represents the overall expected yield of the target plot.

[0157] The system performs a unit conversion operation. Since the collected total forecast value represents the expected yield per unit standard area of ​​the target plot, the system directly converts it to the local standard yield per unit area, such as yield per mu (a Chinese unit of land area), using commonly used agricultural units of measurement and a preset conversion factor. The converted value is then used as the final yield per mu prediction result and is output to the user interface or stored in the results database via the application programming interface. After output, the system releases the memory space occupied by the causal computation graph.

[0158] Unit conversion operations follow the mathematical formulas below:

[0159]

[0160] In the formula, The final output is the predicted yield per acre, expressed in units of the target unit area. . It is the total predicted value collected from the last-level vertices of the causal computation graph, representing the expected yield per unit standard area, for example... . It is a constant used to convert standard area units to commonly used agricultural units. For example, when converting from... Converted to mu-hour The value is approximately .

[0161] The last vertex of the causal computation graph corresponds to the sink in graph theory, and is the endpoint of all computational paths. The predicted total value is the final value obtained after the combined action of all node functions within the causal computation graph, including the basic output function and the environmental stress loss function.

[0162] The unit conversion factor is a static configuration item, pre-stored in the system's basic data layer to ensure that the output results conform to the agricultural production habits of the target region. The cache clearing operation is a memory management process; the system explicitly releases all memory objects dynamically allocated for building the causal computation graph, including node objects, edge objects, and intermediate computation results, preventing memory leaks and ensuring module reusability.

[0163] In one example of the invention, the system captures its output, namely the total predicted value per unit standard area, at the last-level vertex N_out of the causal computation graph it constructs. for The system then initiates the unit conversion process. To output yield per acre in a way that conforms to user habits, the system uses a unit conversion factor. After conversion, the final predicted yield per acre can be obtained by directly substituting the values ​​into the formula: The value 155.83 was officially output by the system. After the output was completed, the system's memory manager was immediately invoked. It traversed and destroyed all memory objects representing nodes N1, N2, N_out and the connections between them, completely releasing the agreed memory space and restoring the network framework building blocks to their initial standby state.

[0164] See appendix Figure 2 This invention also proposes a local crop yield prediction system based on historical data analysis, comprising the following modules:

[0165] The multi-source data processing module is used to acquire multi-source raw monitoring data of the target area, and after decoupling processing, it generates independent and parallel environmental time series tensors and local agricultural record input vectors respectively.

[0166] The climate scenario identification module is used to input environmental time series tensors into a pre-set model and generate agricultural climate scenario labels that characterize meteorological combination features at a spatiotemporal joint scale.

[0167] The network topology configuration module is used to use agricultural climate scenario identifiers as indexes to query a preset mapping lookup table, extract the target node hierarchy sequence, topology connection edge set, and data aggregation operator set that match the current scenario, and generate a network topology configuration file.

[0168] The network framework building module is used to parse the set of topology connection edges in the network topology configuration file and build a running network framework that matches the current scenario. The running network framework contains computing nodes to be configured.

[0169] The causal computation graph assembly module is used to extract the corresponding computation instruction set from the pre-configured program library according to the target node hierarchy sequence, load and bind it to the computing node to be configured, and dynamically assemble it into a causal computation graph.

[0170] The forward propagation computation module is used to extract scalar environmental features based on the data aggregation operator set, and inject the scalar environmental features and the local agricultural record input vector into the causal computation graph to drive forward propagation to calculate the stage-based yield prediction values ​​corresponding to different computation nodes.

[0171] The yield prediction output module is used to aggregate the total predicted values ​​output by the end nodes of the causal computation graph and output the final yield prediction result per acre.

[0172] Each of the modules can be implemented in whole or in part through software, hardware, or a combination thereof. It supports hardware embedded in or independent of the processor in the computer device, and also supports software stored in the memory of the computer device, so that the processor can call and execute the operations corresponding to each of the above modules.

[0173] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for predicting local crop yield per mu based on historical data analysis, characterized in that, Includes the following steps: S1. Obtain multi-source raw monitoring data of the target area, and after decoupling processing, generate independent and parallel environmental time series tensors and local agricultural record input vectors respectively. S2. Input the environmental time series tensor into the pre-set model to generate agricultural climate scenario labels that characterize the meteorological combination at the spatiotemporal joint scale. S3. Using agricultural climate scenario identifiers as indexes, query the preset mapping lookup table, extract the target node hierarchy sequence, topology connection edge set, and data aggregation operator set that match the current scenario, and generate a network topology configuration file. S4. Parse the set of topology connection edges in the network topology configuration file and construct a running network framework that matches the current scenario. The running network framework contains computing nodes to be configured. S5. Based on the target node hierarchy sequence, extract the corresponding computation instruction set from the pre-configured program library, load and bind it to the computation node to be configured, and dynamically assemble it to form a causal computation graph. S6. Extract scalar environmental features based on the data aggregation operator set, and inject the scalar environmental features and the local agricultural record input vector into the causal computation graph to drive forward propagation to calculate the stage-based yield prediction values ​​corresponding to different computation nodes. S7. Aggregate the predicted total value output from the end nodes of the causal computation graph and output the final predicted yield per acre.

2. The method for predicting local crop yield per mu based on historical data analysis according to claim 1, characterized in that, Step S1 includes the following steps: Acquire daily grid meteorological elements covering the target area at a preset cycle, and multispectral satellite remote sensing images at a preset time step, and integrate them into a continuous environmental data stream; Extract unstructured micro-level agricultural operation history information corresponding to the target plot and transform it into a standardized local agricultural record input vector based on natural language processing algorithms; The transformed sparse vectors are segmented and aggregated according to the key growth period or preset time window, and then normalized by dividing by the physical area of ​​the target plot to generate local agricultural record input vectors that independently represent agricultural input per unit area. At the same time, the extracted environmental data are organized into independent environmental time series tensors.

3. The method for predicting local crop yield per mu based on historical data analysis according to claim 1, characterized in that, Step S2 includes the following steps: Directly call the extracted environment time series tensor; A pre-built spatiotemporal coding and decoding network is used to map the environmental temporal tensor to the latent feature space and extract the spatiotemporal joint feature vector; Density clustering algorithm is used to perform unsupervised classification of spatiotemporal joint feature vectors to output agricultural climate scenario labels.

4. The method for predicting local crop yield per mu based on historical data analysis according to claim 1, characterized in that, Step S3 includes the following steps: Obtain the call address labels of each physiological response calculation instruction set in the pre-configured program library; Based on the operation of the query mapping lookup table, extract the target node hierarchy sequence containing the call address label; The target node hierarchical sequence, topology connection edge set, and data aggregation operator set are encapsulated and serialized, and packaged to generate a network topology configuration file.

5. The method for predicting local crop yield per mu based on historical data analysis according to claim 4, characterized in that, Step S4 includes the following steps: Receive network topology configuration file; Disassemble the network topology configuration file, extract the set of topology connection edges, and thereby determine the data flow direction and combination level; Based on the set of topological connecting edges, the system requests address space, establishes connections between nodes, and builds a mesh-structured operating network framework.

6. The method for predicting local crop yield per mu based on historical data analysis according to claim 1, characterized in that, Step S5 includes the following steps: Traverse and parse the call address labels in the hierarchical sequence of the target node; The specific physiological response calculation instruction set is extracted from the pre-configured program library according to the call address number and assigned to the function pointer of the corresponding computing node to be configured in the running network framework. Connecting all the computing nodes that have loaded the instruction set forms a complete directed acyclic execution loop from local input to top-level prediction output, thus constituting a causal computation graph.

7. The method for predicting local crop yield per mu based on historical data analysis according to claim 1, characterized in that, Step S6 includes the following steps: The system retrieves the temporarily stored local agricultural record input vector from the system storage area, and extracts scalar environmental features from the environmental time series tensor based on preset operator rules. The local agricultural record input vector and scalar environmental features are respectively assigned to the bottom initial nodes of the corresponding functions in the causal computation graph as start-up variables; Following the topological order of the directed acyclic execution loop, the data flow is directed to penetrate each layer of computing nodes in sequence, outputting staged production forecast values ​​carrying climate constraints.

8. The method for predicting local crop yield per mu based on historical data analysis according to claim 1, characterized in that, Step S7 includes the following steps: Locate the convergence node of the causal computation graph and collect the total predicted value after calculation by all nodes; Based on the preset unit conversion factor, the unit conversion operation is performed on the predicted total value; Output the final yield prediction result with completed unit conversion, and release the memory space occupied by the causal computation graph.

9. A method for predicting local crop yield per mu based on historical data analysis according to claim 2, characterized in that, Extracting unstructured historical information on micro-level agricultural operations corresponding to the target plot and converting it into a standardized local agricultural record input vector based on natural language processing algorithms includes the following steps: The named entity recognition task extracts the types and specific amounts of agricultural activities from the historical information of micro-level agricultural operations. Each agricultural activity with a specific usage is encoded and mapped to a specific component position of a preset feature dimension, and the remaining components are filled with zero values ​​to generate a sparse vector with timestamps.

10. A local crop yield prediction system based on historical data analysis, characterized in that, Includes the following modules: The multi-source data processing module is used to acquire multi-source raw monitoring data of the target area, and after decoupling processing, it generates independent and parallel environmental time series tensors and local agricultural record input vectors respectively. The climate scenario identification module is used to input environmental time series tensors into a pre-set model and generate agricultural climate scenario labels that characterize meteorological combination features at a spatiotemporal joint scale. The network topology configuration module is used to use agricultural climate scenario identifiers as indexes to query a preset mapping lookup table, extract the target node hierarchy sequence, topology connection edge set, and data aggregation operator set that match the current scenario, and generate a network topology configuration file. The network framework building module is used to parse the set of topology connection edges in the network topology configuration file and build a running network framework that matches the current scenario. The running network framework contains computing nodes to be configured. The causal computation graph assembly module is used to extract the corresponding computation instruction set from the pre-configured program library according to the target node hierarchy sequence, load and bind it to the computing node to be configured, and dynamically assemble it into a causal computation graph. The forward propagation computation module is used to extract scalar environmental features based on the data aggregation operator set, and inject the scalar environmental features and the local agricultural record input vector into the causal computation graph to drive forward propagation to calculate the stage-based yield prediction values ​​corresponding to different computation nodes. The yield prediction output module is used to aggregate the total predicted values ​​output by the end nodes of the causal computation graph and output the final yield prediction result per acre.

Images