AI-based agricultural wisdom planting precision agricultural operation management system

Abstract

The application discloses an AI-based agricultural intelligent planting precision agricultural operation management system and belongs to the technical field of intelligent agriculture. The system comprises five modules, namely, a data acquisition module, a space-time fusion module, a growth prediction module, a scheduling decision module and an execution monitoring module. The data acquisition module acquires satellite remote sensing, ground sensors, weather and field management record signals; the space-time fusion module completes space-time alignment and feature fusion of heterogeneous data through a multi-modal fusion model; the growth prediction module generates a crop growth state prediction signal by relying on a time-series space dual attention model; the scheduling decision module outputs an agricultural scheduling scheme through a multi-objective optimization algorithm; and the execution monitoring module receives the agricultural operation scheduling scheme signal and transmits operation execution feedback signals to the scheduling decision module to trigger dynamic rescheduling. The application solves the problems of insufficient crop growth prediction accuracy and lack of multi-objective optimization mechanism in agricultural scheduling, and improves the precision of agricultural operation.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture and artificial intelligence, specifically to an AI-based smart agricultural planting precision farming operation management system. Background Technology

[0002] The core concept of precision agriculture lies in implementing differentiated management based on the spatial heterogeneity within farmland and the temporal differences in crop growth, aiming to achieve optimal yield and quality output with minimal resource input. In recent years, the rapid development of satellite remote sensing technology, ground-based Internet of Things sensor networks, and weather forecasting technology has provided abundant data sources for real-time perception of farmland environment and crop status. Breakthroughs in artificial intelligence technology, especially deep learning methods, have provided powerful computing tools for the analysis and mining of massive amounts of agricultural data.

[0003] However, existing smart agriculture management systems generally lack effective spatiotemporal alignment mechanisms at the data fusion level. Significant spatial scale mismatches and inconsistent temporal sampling frequencies exist between the pixel-level spatial information of satellite remote sensing imagery and the point-to-point temporal data from ground sensors. Most systems merely perform simple format unification and stitching of heterogeneous data, failing to establish a deep fusion representation under a unified spatiotemporal coordinate system. This results in the inability to interpret and complement the large-area crop growth information in remote sensing data and the in-situ soil environmental information in sensor data within the same semantic space. At the crop growth prediction level, existing models either rely on a single remote sensing vegetation index for growth monitoring or use statistical regression methods to establish empirical relationships between environmental factors and yield. The former only reflects canopy spectral characteristics and cannot analyze the nutrient accumulation and organ differentiation processes within the crop, while the latter ignores the temporal continuity of the crop growth process, treating the prediction of each growth stage as an independent task and failing to capture the lagged effects of early environmental stress on later yield formation. At the agricultural operation scheduling level, existing systems primarily output single irrigation or fertilization prescriptions, lacking the ability to make systematic decisions under multi-dimensional objective constraints. Complex competition and trade-offs exist between multiple objectives such as yield improvement, quality enhancement, resource conservation, and environmental emission control. Single-objective-driven scheduling schemes often sacrifice other objectives. Furthermore, prediction models and scheduling decisions are typically operated as independent modules in existing systems. Prediction results are only used for information display, while scheduling decisions are generated independently based on current observation data. Trend indicators and uncertainty estimates in the prediction information cannot be effectively transmitted to the scheduling decision-making stage, resulting in a lack of foresight and adaptive adjustment capabilities when facing abnormal situations such as sudden weather changes. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this invention is to provide an AI-based intelligent agricultural planting precision farming operation management system to solve the problems of insufficient crop growth prediction accuracy and lack of multi-objective optimization mechanism in agricultural scheduling. The system uses a spatiotemporal fusion module to align and fuse multi-source farmland data from satellite remote sensing, ground sensors, and meteorology in a unified spatiotemporal grid, forming a multimodal fusion representation signal. The growth prediction module utilizes a dual attention prediction model of temporal and spatial data, combining historical fusion features with future meteorological prediction signals to generate a crop growth status prediction signal containing multi-dimensional indicators such as leaf area index, biomass, and yield. The scheduling decision module, based on a multi-objective optimization dynamic scheduling algorithm, uses the prediction signal as input to solve for the Pareto front among yield, quality, resource efficiency, and environmental emission objectives, generating an agricultural operation scheduling scheme signal. The execution monitoring module outputs irrigation, fertilization, and pesticide application instructions and triggers dynamic rescheduling based on operational feedback.

[0005] This invention provides an AI-based intelligent agricultural planting precision farming operation management system, comprising: The data acquisition module collects satellite remote sensing image signals, ground sensor signals, meteorological data signals and field management record signals of the target plot in real time, forming multi-source farmland data signals; The spatiotemporal fusion module receives multi-source farmland data signals and aligns and fuses the heterogeneous data in a unified spatiotemporal grid through a preset multimodal fusion model to form a multimodal fusion representation signal. The growth prediction module receives multimodal fusion representation signals and future weather prediction signals, and generates crop growth status prediction signals through a preset temporal and spatial dual attention prediction model. The scheduling decision module receives crop growth status prediction signals and generates agricultural operation scheduling scheme signals through a dynamic scheduling algorithm based on multi-objective optimization. The execution monitoring module receives agricultural operation scheduling plan signals, outputs irrigation command signals, fertilization command signals, and pesticide application command signals to the operation terminal, and receives operation execution feedback signals returned by the operation terminal. The operation execution feedback signals are then transmitted to the scheduling decision module to trigger dynamic rescheduling.

[0006] In one embodiment of the present invention, the data acquisition module includes a satellite remote sensing image acquisition unit, a ground sensor acquisition unit, a meteorological data acquisition unit, and a field management record entry unit. The satellite remote sensing image acquisition unit acquires multispectral image sequences of the target plot in various bands by accessing a satellite remote sensing data platform. The ground sensor acquisition unit acquires soil temperature, soil moisture, soil conductivity, light intensity, air temperature, and air humidity values ​​respectively by using soil temperature sensors, soil moisture sensors, soil conductivity sensors, light intensity sensors, air temperature sensors, and air humidity sensors deployed within the target plot. The meteorological data acquisition unit acquires precipitation, wind speed, atmospheric pressure, and solar radiation flux of the area where the target plot is located by accessing a meteorological data service platform. The field management record entry unit acquires irrigation records, fertilizer variety and application rate records, tillage time records, and agricultural machinery operation trajectory records of the target plot. The data acquisition module adds spatial coordinate labels and timestamp labels to the data from the above data sources and assembles them into multi-source farmland data signals.

[0007] In one embodiment of the present invention, the preset multimodal fusion model includes a spatiotemporal alignment layer, a channel-specific feature extraction layer, and a cross-modal attention fusion layer. The spatiotemporal alignment layer maps satellite remote sensing image data, ground sensor data, meteorological data, and field management record data from multi-source farmland data signals to a spatiotemporal reference system with the target plot boundary as the spatial range and the crop sowing date as the time origin. In the spatial dimension, it transforms the spatial coordinates of each data source to a unified spatial reference system and aggregates them using grid cells of a preset resolution. In the temporal dimension, it resamples the time series of each data source according to a preset time step to align them to a unified time axis, forming multidimensional tensor data. The channel-specific feature extraction layer divides the multidimensional tensor data into remote sensing channels, ground sensor channels, and meteorological channels. Each channel is input to a corresponding feature extraction sub-network. The remote sensing channel uses a 3D convolutional neural network to perform spatiotemporal joint feature extraction along the spatial grid and time step. The ground sensor channel uses a gated recurrent unit network to process the multivariate time series of each sensor. The meteorological channel uses a gated recurrent unit network to process the multivariate time series of the meteorological grid cell. The feature vectors output by each channel are concatenated into a multi-channel feature matrix after layer normalization. The cross-modal attention fusion layer calculates the attention weight matrix between remote sensing features and meteorological features, as well as the attention weight matrix between ground sensor features and meteorological features. The cross-modal representation vector is obtained by weighted summation. The cross-modal representation vector is concatenated with the multi-channel feature matrix and then dimensionality is reduced by a fully connected layer to output the multi-modal fused representation signal.

[0008] In one embodiment of the present invention, the specific process of calculating the attention weight matrix between remote sensing features and meteorological features in the cross-modal attention fusion layer is as follows: the remote sensing feature vector is transformed into a remote sensing query vector through a first linear projection matrix; the meteorological feature vector is transformed into a meteorological key vector and a meteorological value vector through a second linear projection matrix; the dot product similarity between the remote sensing query vector and the meteorological key vector is calculated; the dot product similarity is scaled by dividing by the square root of the dimension of the meteorological key vector; and a normalized exponential function is applied to the scaled similarity matrix to obtain the normalized attention weight. The normalized attention weight represents the relationship between the remote sensing observation features at each spatial location of the target plot and the meteorological conditions of each meteorological grid unit. To determine the degree of correlation, the meteorological value vector is weighted and summed using normalized attention weights to generate a cross-modal representation vector of remote sensing and meteorology. Simultaneously, when calculating the attention weight matrix between ground sensor features and meteorological features, the ground sensor feature vector is transformed into a sensor query vector using a third linear projection matrix. Using the meteorological key vector and meteorological value vector as shared key-value pairs, the dot product similarity between the sensor query vector and the meteorological key vector is calculated and mapped using scaling and a normalized exponential function to obtain the normalized attention weight between the ground sensor and meteorology. This weight is then used to weight and sum the meteorological value vector to generate the cross-modal representation vector of the ground sensor and meteorology. The formula for calculating the attention weight matrix is ​​as follows:

[0009] in, This is the remote sensing-meteorological attention weight matrix. For normalized exponential functions, To query the projection matrix, This is the remote sensing feature matrix. The key projection matrix, The meteorological feature matrix, For matrix transpose, The dimension of the key vector. This is the bias matrix.

[0010] In one embodiment of the present invention, the preset temporal and spatial dual attention prediction model adopts a network structure of stacked encoder and decoder. The encoder receives the multimodal fusion representation signal and unfolds the multimodal fusion representation signal into a feature sequence along the time dimension. The encoder is composed of multiple layers of temporal self-attention layers and multiple layers of feedforward network layers stacked alternately. Each temporal self-attention layer generates a query vector, a key vector, and a value vector for each time step position in the feature sequence through linear projection. The dot product similarity between the query vector and the key vector between any two time step positions in the sequence is calculated and mapped to the temporal self-attention weight through a normalized exponential function. The temporal self-attention weight represents the relative importance of each time step in the historical feature sequence to the prediction of the growth state of the current time step. The value vector is weighted and summed using the temporal self-attention weight. The encoder hidden state is output after residual connection and layer normalization. The decoder receives the encoder hidden state and the future weather forecast signal. The decoder consists of multiple cross-attention layers. Each cross-attention layer uses the query vector generated by linear projection of the future weather forecast signal as the query and the key vector and value vector generated by linear projection of the encoder hidden state as the key-value pair. It calculates the cross-attention weight between the future weather conditions and historical growth characteristics. The value vector of the encoder hidden state is weighted and summed with the cross-attention weights and then processed by residual connection and layer normalization to output the decoder hidden state. The decoder hidden state is mapped by the multi-task prediction head network to output the predicted values ​​of leaf area index, aboveground biomass accumulation, soil moisture content, nitrogen accumulation, and yield estimate, which together constitute the crop growth status prediction signal.

[0011] In one embodiment of the present invention, the multi-task prediction head network includes a shared feature transformation layer and multiple task-specific output layers. The shared feature transformation layer receives the decoder hidden state and maps the decoder hidden state to a shared feature representation through a multi-layer fully connected network and a nonlinear activation function. The multiple task-specific output layers correspond to the leaf area index prediction task, aboveground biomass accumulation prediction task, soil moisture content prediction task, nitrogen accumulation prediction task, and yield estimation task, respectively. Each task-specific output layer contains an independent linear transformation matrix that maps the shared feature representation to the numerical output of the corresponding prediction target. The loss function of the preset temporal and spatial dual attention prediction model during the training phase is the weighted sum of the mean square errors between the predicted and measured values ​​for each task. The loss weights for each task are automatically learned through an uncertainty weighting method. The calculation formula of the loss function is as follows:

[0012] in, Here, T represents the joint loss function for multiple tasks, and T is the total number of prediction tasks. Let Variance be the uncertainty of the t-th task. Let be the mean square error function. Let t be the predicted value for the t-th task. Let t be the measured value of the t-th task.

[0013] In one embodiment of the present invention, the dynamic scheduling algorithm based on multi-objective optimization includes a decision variable encoding step, a multi-objective function construction step, a multi-objective optimization solution step, and a scheme selection step. The decision variable encoding step combines the irrigation amount, nitrogen application amount, phosphorus application amount, potassium application amount, and pesticide application amount for each operation time step into a decision variable vector. A feasible solution space is constructed based on constraints such as the single irrigation amount not exceeding the difference between the soil field capacity and the current water content, the single nitrogen application amount not exceeding the safety threshold of the crop's current growth stage, the time interval between two adjacent operations of the same type not being less than a preset minimum interval, and the upper limit of available agricultural machinery resources. The function construction steps involve constructing a yield objective function, a quality objective function, a resource utilization efficiency objective function, and an environmental emission objective function. The yield objective function is based on the yield estimate in the crop growth status prediction signal and mapped to a standardized interval through a piecewise linear function. The quality objective function is calculated based on the deviation of the predicted quality index value from the standard value of the target variety. The resource utilization efficiency objective function is calculated based on the amount of irrigation water and total fertilizer consumed per unit yield. The environmental emission objective function is calculated based on the weighted sum of nitrogen surplus and deep seepage. Each objective function is normalized to a unified standardized interval from minimum to maximum.

[0014] In one embodiment of the present invention, the multi-objective optimization solution step employs a decomposition-based multi-objective evolutionary optimization method. The multiple objective functions constructed in the multi-objective function construction step are decomposed into multiple single-objective sub-problems using a Chebyshev aggregation function. Each sub-problem corresponds to a set of weight vectors covering different directions of the objective space. Decision variable vectors are used as population individual codes. Simulated binary crossover and polynomial mutation operators are used to generate offspring individuals. Penalty functions are applied to offspring individuals that violate constraints. After each iteration, the Chebyshev aggregation value of each sub-problem is calculated, and the current optimal solution for each sub-problem is updated. Simultaneously, an external archive is maintained to store non-dominated features discovered during the iteration process. When the preset iteration termination condition is met, all non-dominated solutions in the external archive are output as the Pareto front solution set. The solution selection step selects a satisfactory solution from the Pareto front solution set. The selection method is to calculate the weighted Euclidean distance between the multidimensional objective function vector of each non-dominated solution in the Pareto front and the ideal point. The weighting weight is determined by the user through the preset preference vector via the decision preference configuration interface. The non-dominated solution with the smallest weighted Euclidean distance is selected. The decision variable vector of this solution is decoded into irrigation command signals, fertilization command signals, and pesticide application command signals for each operation time step. These signals are organized in time step order as agricultural operation scheduling scheme signals. The formula for calculating the weighted Euclidean distance is as follows:

[0015] in, The weighted Euclidean distance is K, where K is the number of targets to optimize. The preference weight for the k-th objective is... To solve for the k-th objective value of x, For the k-th target ideal value, This is the regularization coefficient.

[0016] In one embodiment of the present invention, after transmitting the agricultural operation scheduling scheme signal to the execution monitoring module, the scheduling decision module continuously receives the operation execution feedback signal returned by the execution monitoring module. The operation execution feedback signal includes irrigation execution amount, fertilization execution amount, pesticide application execution amount, and operation timestamp. At the same time, in each preset reassessment cycle, the scheduling decision module obtains the soil moisture content observation value and crop growth status observation value at the current time step from the data acquisition module, calculates the deviation vector between the soil moisture content observation value and crop growth status observation value and the soil moisture content prediction value and crop growth status prediction value at the corresponding time step generated by the growth prediction module in the previous scheduling cycle. If the absolute value of any element in the deviation vector exceeds the preset tolerance threshold corresponding to that element, dynamic rescheduling is triggered. Dynamic rescheduling starts with the soil moisture content observation value and crop growth status observation value at the current time step as the initial state, and re-executes the decision variable encoding step, multi-objective function construction step, multi-objective optimization solution step, and scheme selection step to generate a corrected agricultural operation scheduling scheme signal to replace the unexecuted part of the original scheme.

[0017] In one embodiment of the present invention, the execution monitoring module includes an instruction distribution unit and a feedback receiving unit. The instruction distribution unit receives agricultural operation scheduling scheme signals, parses the irrigation amount, fertilizer amount, pesticide amount, agricultural machinery allocation and operation mode instructions corresponding to the current time step in the agricultural operation scheduling scheme signals, converts the irrigation amount into an irrigation instruction signal to control the opening and closing time of irrigation valves or the flow rate of irrigation pumps, converts the fertilizer amount into a fertilizer instruction signal to control the discharge rate or discharge amount of fertilizer application device, and converts the pesticide amount into a pesticide instruction signal to control the spraying rate or spraying amount of pesticide application device. Each instruction signal is accompanied by a target plot identifier and operation time window information. The feedback receiving unit receives the operation execution feedback signal returned by the operation terminal after the operation is performed. The operation execution feedback signal includes the actual irrigation amount, actual fertilizer amount, actual pesticide amount, actual agricultural machinery operation trajectory and the completion timestamp of each operation. The feedback receiving unit transmits the operation execution feedback signal to the scheduling decision module.

[0018] This invention provides an AI-based intelligent agricultural precision farming operation management system. The system uses a spatiotemporal fusion module to align and fuse multi-source farmland data (satellite remote sensing, ground sensors, meteorology, etc.) within a unified spatiotemporal grid, forming a multimodal fusion representation signal. A growth prediction module utilizes a dual-attention prediction model (temporal and spatial), combining historical fusion features with future meteorological prediction signals to generate a crop growth status prediction signal containing multi-dimensional indicators such as leaf area index, biomass, and yield. A scheduling decision module, based on a multi-objective optimization dynamic scheduling algorithm, uses the prediction signal as input to solve for the Pareto front among yield, quality, resource efficiency, and environmental emission objectives, generating an agricultural operation scheduling scheme signal. The execution monitoring module outputs irrigation, fertilization, and pesticide application instructions and triggers dynamic rescheduling based on operation feedback. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a system architecture diagram of an AI-based smart agricultural planting precision farming operation management system. Figure 2 This is the overall main flowchart of the system; Figure 3 Internal flowchart for spatiotemporal fusion + growth prediction; Figure 4 This is a flowchart illustrating the closed-loop process of scheduling decision-making and execution monitoring. Detailed Implementation

[0021] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0022] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0023] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0024] Please see Figure 1-4 The diagram illustrates the AI-based precision agricultural operation management system of the present invention, comprising: a data acquisition module that collects satellite remote sensing image signals, ground sensor signals, meteorological data signals, and field management record signals of the target plot in real time, forming multi-source farmland data signals; a spatiotemporal fusion module that receives multi-source farmland data signals and aligns and fuses heterogeneous data in a unified spatiotemporal grid using a preset multimodal fusion model, forming a multimodal fusion representation signal; a growth prediction module that receives the multimodal fusion representation signal and future meteorological prediction signals, and generates crop growth status prediction signals using a preset temporal and spatial dual attention prediction model; a scheduling decision module that receives the crop growth status prediction signal and generates an agricultural operation scheduling scheme signal using a dynamic scheduling algorithm based on multi-objective optimization; and an execution monitoring module that receives the agricultural operation scheduling scheme signal, outputs irrigation command signals, fertilization command signals, and pesticide application command signals to the operation terminal, and receives the operation execution feedback signal returned by the operation terminal, transmitting the operation execution feedback signal to the scheduling decision module to trigger dynamic rescheduling.

[0025] Figure 1 As shown, the AI-based smart agricultural precision farming management system comprises a data acquisition module, a spatiotemporal fusion module, a growth prediction module, a scheduling and decision-making module, and an execution monitoring module. The fundamental problem this system aims to solve is how to achieve deep fusion of multi-source heterogeneous data, accurate prediction of crop growth status, and multi-objective optimized scheduling of agricultural operations in agricultural planting scenarios, enabling the three stages of data perception, status prediction, and operational decision-making to operate in a coordinated manner within the same technical framework.

[0026] The data acquisition module undertakes the sensing tasks at the system's front end, responsible for comprehensively acquiring information related to the farmland environment and crop growth of the target plot from multiple dimensions. This module comprises four data acquisition units: a satellite remote sensing image acquisition unit, a ground sensor acquisition unit, a meteorological data acquisition unit, and a field management record entry unit. The satellite remote sensing image acquisition unit, by connecting to a satellite remote sensing data platform, acquires remote sensing image sequences of the target plot in various multispectral bands. These image sequences provide pixel reflectance information of the target plot in multiple bands, including visible light, near-infrared, and shortwave infrared, providing fundamental data for monitoring the spatial distribution of large-area crop canopy structure, chlorophyll content, and vegetation cover. The ground-based sensor acquisition unit measures in-situ environmental parameters through multiple sensor nodes deployed within the target plot. These sensors include soil temperature, soil moisture, soil conductivity, light intensity, air temperature, and air humidity sensors, which collect values ​​for soil temperature, soil moisture, soil conductivity, light intensity, air temperature, and air humidity, respectively. These parameters reflect the water, heat, and salinity of the soil in the crop root zone and the near-surface microclimate conditions, providing direct evidence for analyzing environmental stresses and water and nutrient supply capacity for crop growth. The meteorological data acquisition unit connects to a meteorological data service platform to obtain meteorological elements such as precipitation, wind speed, atmospheric pressure, and solar radiation flux in the target plot area. The meteorological data includes both historical measured values ​​and future weather forecasts, providing prior information on medium-term meteorological conditions for growth prediction models. The field management record entry unit acquires irrigation records, fertilizer variety and application rate records, tillage time records, and agricultural machinery operation trajectory records for the target plot. These records reflect the type, quantity, and spatiotemporal distribution of agricultural operations performed, and are an essential component in constructing a complete historical archive of the crop growth process. The data acquisition module adds spatial coordinate labels and timestamp labels to the raw data acquired from the above data sources. The spatial coordinate labels associate each data point with its corresponding geographic location, and the timestamp labels record the precise time of data acquisition. The data from each source after being labeled are uniformly assembled into a multi-source farmland data signal. This signal contains four components: remote sensing image data sub-signal, ground sensor data sub-signal, meteorological data sub-signal, and field management record data sub-signal, and is transmitted to the spatiotemporal fusion module.

[0027] like Figure 2As shown, this flowchart presents the entire workflow of the system from data acquisition to execution feedback. It primarily uses linear flow supplemented by feedback loops, clearly demonstrating the connection logic of the five core modules. It covers the complete process from farmland data acquisition, feature fusion, growth prediction, agricultural decision-making to operation execution, forming a closed-loop control system to ensure precise and controllable agricultural operations throughout the entire process. The data acquisition module, as the system's front-end sensing unit, undertakes the task of real-time acquisition and standardized output of multi-source farmland data. It simultaneously acquires four types of core signals: satellite remote sensing imagery, ground sensor monitoring data, meteorological monitoring information, and field management records. It uniformly adds spatial coordinates and timestamps to data from different sources, standardizing them into multi-source farmland data signals with a unified format. This solves the problems of chaotic data formats and inconsistent spatiotemporal benchmarks from different sources, providing standardized and complete raw data support for subsequent data processing and serving as the data foundation for the entire system. The spatiotemporal fusion module follows the data acquisition module. Its core function is to achieve spatiotemporal alignment and deep feature fusion of heterogeneous data. After receiving multi-source farmland data signals, it maps scattered remote sensing, sensor, and meteorological data to a unified spatiotemporal grid through a preset multimodal fusion model. This eliminates the differences in spatial resolution and time step between different data sources, extracts the core features of each modality, and completes cross-modal fusion to generate a multimodal fusion representation signal. This significantly improves the correlation and effectiveness of the data, eliminates redundant information, strengthens key features, and lays a high-quality data foundation for accurate crop growth prediction.

[0028] The growth prediction module, as the core analysis unit of the system, receives the multimodal fusion representation signal output by the spatiotemporal fusion module and simultaneously accesses future weather forecast signals. Relying on a dual-attention prediction model combining temporal and spatial data, it fully explores the correlation between historical growth characteristics and future weather conditions, accurately outputting multi-dimensional crop growth status prediction signals such as leaf area index, biomass, soil moisture content, nitrogen accumulation, and yield. This comprehensively predicts crop growth status and needs, providing a scientific and accurate core basis for agricultural scheduling decisions. The scheduling decision module, the system's decision-making center, receives the crop growth status prediction signal and, through a dynamic scheduling algorithm based on multi-objective optimization, integrates four objectives—yield, quality, resource utilization efficiency, and environmental emissions—to generate an agricultural operation scheduling plan signal that takes into account the needs of all parties. This transforms crop growth needs into specific and executable irrigation, fertilization, and pesticide application plans, balancing the economy, environmental protection, and yield of agricultural operations, and avoiding resource waste or quality degradation caused by a single-objective approach. The execution monitoring module, acting as the link between decision-making and field operations, receives agricultural operation scheduling signals from the scheduling decision module. It parses and converts these signals into irrigation, fertilization, and pesticide application instructions that can be recognized by the field terminals, then sends these instructions to the field terminals. Simultaneously, it receives real-time operational data from the field terminals, generating an execution feedback signal. This ensures the accurate implementation of scheduling instructions and provides real-world data support for subsequent dynamic adjustments. The execution monitoring module then sends this feedback signal back to the scheduling decision module, triggering a dynamic rescheduling mechanism. This forms a closed-loop feedback cycle, ensuring the system can optimize agricultural scheduling plans in real-time based on actual field conditions and crop growth changes, adapting to complex and changing field environments and continuously improving the accuracy and adaptability of agricultural operations.

[0029] The spatiotemporal fusion module receives multi-source farmland data signals from the data acquisition module. Through a pre-defined multimodal fusion model, it aligns and fuses the heterogeneous data within a unified spatiotemporal grid, outputting a multimodal fusion representation signal. The pre-defined multimodal fusion model structurally comprises three sequentially connected processing layers. The first layer is the spatiotemporal alignment layer, which maps data from different sensors in the multi-source farmland data signals to a spatiotemporal reference system with the target plot boundary as the spatial range and the crop sowing date as the time origin. Spatially, the spatiotemporal alignment layer transforms the geographic coordinates of satellite remote sensing imagery, the installation location coordinates of ground sensors, and the center coordinates of meteorological grid cells to the same geospatial reference system. It spatially discretizes the target plot using grid cells of a pre-defined resolution. When multiple observations fall within the same grid cell, the arithmetic mean is taken as the value of that grid cell. When a grid cell has no direct observation, spatial interpolation is used to estimate the missing value. In the temporal dimension, the spatiotemporal alignment layer uses the crop sowing date of the target plot as the time origin and a preset time step as the basic time unit to establish a uniform time axis extending from the sowing date to the current date. The time series data from each data source are resampled to match each time step on this uniform time axis. During resampling, the average of multiple samples falling within the same time step is taken, and missing time steps are filled using linear interpolation. After spatiotemporal alignment, the multi-source data is reconstructed into a unified form of multidimensional tensor data. The dimensions of this tensor correspond to spatial grid rows, spatial grid columns, time steps, and variable channels, respectively. The second layer is a channel-specific feature extraction layer. This layer inputs the multidimensional tensor data into three independent feature extraction sub-networks according to remote sensing channels, ground sensor channels, and meteorological channels. The feature extraction sub-network for the remote sensing channel uses a three-dimensional convolutional neural network. The convolution kernel performs convolution operations simultaneously in two dimensions of the spatial grid and the time step dimension, extracting the spatial distribution pattern and evolution trend of vegetation indices in remote sensing images along the time axis, capturing the spatiotemporal correlation between crop canopy spectral characteristics and growth processes. The feature extraction subnetwork of the ground sensor channel employs a gated recurrent unit network. The multivariate time series data from each sensor node is expanded along time steps and input into this recurrent network. The gated recurrent unit adaptively determines the retention and forgetting ratio of historical information through its internal reset and update gate mechanisms, thereby effectively capturing the long-term temporal dependencies of in-situ environmental variables such as soil temperature and humidity. The feature extraction subnetwork of the meteorological channel also employs a gated recurrent unit network, taking the multivariate time series data from the meteorological grid cells as input and outputting the meteorological context representation vector for each time step.

[0030] The feature vectors output from the three channels are each processed by layer normalization to eliminate differences in numerical scale between channels, and then concatenated along the feature dimension to form a multi-channel feature matrix. The third layer is a cross-modal attention fusion layer. This layer calculates the attention weight matrix between remote sensing features and meteorological features, as well as the attention weight matrix between ground sensor features and meteorological features. It automatically learns the semantic association strength between different modal data through an attention mechanism. When calculating the attention weight between remote sensing features and meteorological features, the remote sensing feature vector is transformed into a remote sensing query vector through a first linear projection matrix, and the meteorological feature vector is transformed into a meteorological key vector and a meteorological value vector through a second linear projection matrix. The dot product similarity between the remote sensing query vector and the meteorological key vector is calculated and scaled by dividing by the square root of the dimension of the meteorological key vector. The normalized attention weight is obtained by applying a normalized exponential function to the scaled similarity matrix. This weight represents the degree of association between the remote sensing observation features at each spatial location of the target plot and the meteorological conditions of each meteorological grid unit. The meteorological value vector is weighted and summed with this weight to generate a cross-modal representation vector of remote sensing and meteorology. Simultaneously, when calculating the attention weights between ground sensor features and meteorological features, the ground sensor feature vector is transformed into a sensor query vector through a third linear projection matrix. Using the aforementioned meteorological key vector and meteorological value vector as shared key-value pairs, the dot product similarity between the sensor query vector and the meteorological key vector is calculated and mapped using scaling and a normalized exponential function to obtain the normalized attention weights for ground sensor and meteorological features. These weights are then used to weight and sum the meteorological value vectors to generate a cross-modal representation vector for ground sensor and meteorological features. The two cross-modal representation vectors are then concatenated with the multi-channel feature matrix output from the multi-channel feature extraction layer along the feature dimension. The resulting extended feature matrix is ​​then compressed through a fully connected layer to finally output a multi-modal fusion representation signal. This signal, in tensor form, represents the comprehensive characteristics of the target plot at different time steps, integrating remote sensing, ground sensor, and meteorological multi-modal information. The formula for calculating the attention weight matrix is ​​as follows:

[0031] in This is the normalized attention weight matrix between remote sensing features and meteorological features. Each element in the matrix corresponds to the correlation strength between the remote sensing features of a single spatial location of the target plot and the meteorological conditions of a single meteorological grid unit. The value ranges from 0 to 1, and the closer the value is to 1, the stronger the correlation between the two. The normalization exponential function maps the original similarity to weights in probabilistic form, ensuring that the sum of the elements in each row of the weight matrix is ​​1, thus achieving weight normalization. This is the remote sensing feature query projection matrix, used to map the original remote sensing features to the query space, with the dimensions matching the feature dimensions; It is a remote sensing feature matrix, which is composed of spatiotemporal joint features extracted by a three-dimensional convolutional neural network, and contains remote sensing feature information of each spatial location and time step of the target plot; This is the meteorological feature key projection matrix, used to map the original meteorological features to the key space, with the same dimensions as the query projection matrix; The meteorological feature matrix is ​​composed of meteorological time-series features extracted by the gated cyclic unit network, and includes the time-series feature information of each meteorological grid unit. For matrix transpose operations, the key matrix dimension is adapted to complete the dot product calculation; denoted by , B is the square root of the key vector dimension, used to scale the dot product similarity to avoid the problem of excessively large dot product results leading to softmax function saturation and reduced weight discrimination under high-dimensional features; B is the bias matrix, used to fine-tune the similarity distribution to adapt to the feature association characteristics of different plots and crop types. This algorithm achieves heterogeneous feature space alignment through linear projection, calculates feature similarity using dot product, and obtains accurate association weights through scaling and normalization. It effectively captures the dynamic dependence between crop growth and meteorological conditions reflected in remote sensing images, solving the shortcomings of traditional stitching and fusion methods that cannot distinguish feature importance and have high feature redundancy. This makes the fused features more closely reflect the actual impact on crop growth, providing high-quality input features for subsequent temporal spatial dual attention prediction models, improving the accuracy of crop growth status prediction from the source, laying a reliable foundation for precise agricultural scheduling, adapting to the multi-source data fusion needs of different climate zones and planting plots, and significantly improving the system's environmental adaptability and feature extraction effectiveness.

[0032] like Figure 3As shown in the flowchart, this diagram illustrates the detailed internal workflow of the spatiotemporal fusion module and the growth prediction module. It clearly presents the complete logic of feature extraction, fusion, and prediction from raw multi-source data to crop growth prediction results. Each sub-module performs its own function and progresses layer by layer, ensuring the effectiveness of feature extraction and the accuracy of growth prediction. The spatiotemporal alignment layer, as the starting unit of the internal process, plays a crucial role in unifying the spatiotemporal reference of multi-source data. After receiving multi-source farmland data signals output by the data acquisition module, it maps satellite remote sensing, ground sensor, meteorological, and field record data to a spatiotemporal reference system with the target plot as the spatial range and the sowing date as the time origin. The spatial dimension aggregates grid cells according to a preset resolution, and the time dimension resamples time-series data at a fixed step size, completely eliminating the deviations of heterogeneous data in spatial coordinates and time scales. This results in standardized multi-dimensional tensor data with a unified structure, laying a standardized data foundation for subsequent channel-specific feature extraction and serving as a prerequisite for heterogeneous data fusion. The channel-specific feature extraction layer follows the spatiotemporal alignment layer and employs a modality-independent extraction strategy. Multidimensional tensor data is divided into three modalities: remote sensing, ground sensor, and meteorological. Each channel is matched with a dedicated feature extraction network: the remote sensing channel uses a 3D convolutional neural network to simultaneously capture spatial grid distribution features and temporal series variation features, achieving spatiotemporal joint feature extraction; the ground sensor and meteorological channels both use gated recurrent unit networks to efficiently process multivariate time-series data and capture the dynamic variation patterns of various parameters. The feature vectors extracted from each channel are layer-normalized and then concatenated into a multi-channel feature matrix. This preserves the independence of each modality's features while achieving preliminary feature integration, avoiding intermodal feature interference, and improving the targeting of feature extraction.

[0033] The cross-modal attention fusion layer, as the core unit of feature fusion, receives the multi-channel feature matrix output by the channel-specific feature extraction layer. It employs a dual-branch parallel fusion structure to complete the feature fusion of remote sensing and meteorological features, and ground sensor and meteorological features respectively: The first branch calculates the attention weight matrix of remote sensing and meteorological features, generates query, key, and value vectors through linear projection, obtains the association weights through dot product similarity calculation, scaling, and normalization, and generates a remote sensing-meteorological cross-modal representation vector through weighted summation; The second branch uses meteorological features as shared key-value pairs, similarly calculates the attention weights of ground sensor features and meteorological features, and generates a sensor-meteorological cross-modal representation vector; The cross features of the two branches are concatenated with the multi-channel feature matrix to reduce dimensionality, outputting a multi-modal fusion representation signal, accurately capturing the dynamic correlation between heterogeneous features, and solving the problem of weak correlation in traditional feature concatenation and fusion. The encoder, acting as the temporal coding unit for growth prediction, receives the multimodal fusion representation signal output from the cross-modal attention fusion layer. It unfolds the signal along the time dimension into a feature sequence, consisting of alternating stacks of multiple temporal self-attention layers and feedforward network layers. Each self-attention layer generates query, key, and value vectors through linear projection, calculates the dependencies between temporal features, generates temporal self-attention weights, and after weighted summation, performs residual connection and layer normalization processing to output the encoder hidden state. This fully explores the historical temporal patterns of crop growth, providing historical feature support for subsequent growth trend decoding. The decoder, acting as the trend decoding unit for growth prediction, receives the encoder hidden state and future weather prediction signals. It consists of multiple cross-attention layers. Each cross-attention layer uses future weather features as the query vector and the encoder hidden state as the key-value pair, calculates the correlation weight between future weather and historical growth features, and after weighted summation, performs residual connection and layer normalization processing to output the decoder hidden state. This accurately decodes the crop growth trend under future weather conditions, achieving joint decoding of temporal and spatial features. The multi-task prediction head network, as the prediction result output unit, receives the hidden state of the decoder, generates unified shared features through a shared feature transformation layer, and then passes through multiple task-specific output layers to map and output five core prediction results: leaf area index, aboveground biomass accumulation, soil moisture content, nitrogen accumulation, and yield. This comprehensively and accurately presents the crop growth status, providing multi-dimensional and all-round growth data support for agricultural scheduling decisions and ensuring that the scheduling plan fits the actual growth needs of the crop.

[0034] Specifically, the growth prediction module receives the multimodal fusion representation signal output by the spatiotemporal fusion module, and simultaneously receives the future weather forecast signal from the meteorological data acquisition unit. It then generates a crop growth status prediction signal through a pre-defined temporal and spatial dual-attention prediction model. This model employs a network structure with stacked encoders and decoders. The encoder receives the multimodal fusion representation signal and expands it along the time dimension into a feature sequence arranged in time step order. The encoder consists of alternating stacks of multiple temporal self-attention layers and multiple feedforward network layers. For each time step position in the input feature sequence, each temporal self-attention layer generates a query vector, key vector, and value vector at that position using three independent linear projection matrices. For any two time step positions in the sequence, the dot product of the query vector at one position and the key vector at the other is calculated. This dot product reflects the similarity between the features at the two time steps. The dot product result is scaled by dividing by the square root of the key vector dimension to prevent excessive values. A normalized exponential function is applied to the scaled dot products of all position pairs in the sequence to obtain the temporal self-attention weight matrix. Each element in the temporal self-attention weight matrix represents the relative importance of a certain time step in the historical feature sequence to the prediction of the growth state at another time step. For example, the features of the early water stress time step may significantly influence the prediction of the later yield formation time step through a higher attention weight. The value vector is weighted and summed using the temporal self-attention weights to obtain the attention output at each time step. This output is added to the input via a residual connection and then normalized through layer normalization to stabilize the numerical distribution. It is then further transformed nonlinearly by a feedforward network layer, which consists of two fully connected layers and an intermediate nonlinear activation function. Its output is also processed through residual connections and layer normalization to finally obtain the encoder hidden state. The decoder receives the encoder hidden state output by the encoder and the future weather forecast signal. The future weather forecast signal is the temporal features of meteorological variables at each time step within the future time period extracted from the weather forecast data. The decoder consists of multiple cross-attention layers. Each cross-attention layer uses the query vector generated by the linear projection of the future weather forecast signal as the query and the key vector and value vector generated by the linear projection of the encoder hidden state as key-value pairs to calculate the cross-attention weights between future weather conditions and historical growth features. This mechanism uses future weather conditions as query conditions, retrieving the best-matching growth response pattern from the encoded memory of historical growth characteristics. For example, when a sustained high-temperature and drought condition is predicted in future weather forecasts, cross-attention weights will assign higher weights to feature vectors of crop growth performance under similar weather conditions in historical sequences, thus integrating future weather prediction information into the growth state prediction process. The encoder's hidden state value vector is weighted and summed using cross-attention weights, and after residual connection and layer normalization, the decoder's hidden state is output. The decoder's hidden state is mapped to multiple specific prediction target values ​​via a multi-task prediction head network.The multi-task prediction head network consists of a shared feature transformation layer and multiple task-specific output layers. The shared feature transformation layer receives the decoder's hidden state and further abstracts it into a shared feature representation suitable for multiple tasks through a multi-layer fully connected network and a non-linear activation function. Each task-specific output layer contains an independent linear transformation matrix that maps the shared feature representation to the numerical output of the corresponding prediction target. These prediction targets include leaf area index prediction, aboveground biomass accumulation prediction, soil moisture content prediction, nitrogen accumulation prediction, and yield estimate, which together constitute the crop growth status prediction signal. This signal is transmitted to the scheduling decision module. The loss function is calculated as follows:

[0035] in The multi-task joint loss function is used to measure the overall deviation between the model's predicted values ​​and the measured values. The smaller the loss value, the higher the model's prediction accuracy. T is the total number of prediction tasks, corresponding to the prediction tasks of five core crop growth indicators: leaf area index, aboveground biomass accumulation, soil moisture content, nitrogen accumulation, and yield. Let V be the uncertainty variance of the t-th prediction task, which is automatically learned by the model during training. It reflects the confidence of the model in the prediction results of the task. The larger the variance, the higher the difficulty and uncertainty of the prediction task. Here, the mean squared error function is used to calculate the squared mean deviation between the predicted and measured values ​​for the t-th task. Let t be the predicted value of the t-th task output by the model. This represents the actual value of the t-th task measured in the field. To perform the summation operation, the losses from the five tasks are integrated; The mean squared error term is weighted with uncertainty. The weight of each task loss is automatically adjusted by the uncertainty variance. The weight of tasks with high prediction difficulty is reduced, and the weight of tasks with low difficulty is increased, so as to avoid the model overfitting simple tasks and ignoring complex tasks. The regularization term is used to prevent the uncertainty variance from increasing infinitely, which could lead to model training instability. It constrains the variance value to a reasonable range, ensuring the convergence of the loss function. This algorithm abandons the traditional fixed-weight weighted summation method and introduces task uncertainty to achieve dynamic adaptive adjustment of the loss weights. It can effectively solve the problems of large differences in the prediction difficulty of different growth indicators and the underfitting of some tasks due to fixed weights. It enables the model to simultaneously optimize all prediction tasks during training, balancing the prediction accuracy of various indicators. It ensures the prediction accuracy of core indicators such as yield and biomass, while also taking into account the reliability of auxiliary indicators such as soil moisture content and nitrogen accumulation. The output crop growth status prediction signal can comprehensively and accurately reflect the crop growth dynamics, providing complete and reliable growth data support for multi-objective agricultural scheduling. It adapts to the growth patterns of different crop growth stages and improves the prediction stability and practicality of the system in complex planting scenarios.

[0036] like Figure 4As shown, this flowchart vertically presents the closed-loop workflow of the scheduling decision-making module and the execution monitoring module, integrating the entire process of decision optimization, instruction issuance, operation execution, and dynamic feedback. It clearly demonstrates the complete logic of agricultural scheduling from plan generation to implementation and dynamic optimization, ensuring the scientific, flexible, and accurate nature of agricultural scheduling. Decision variable encoding, as the starting unit of the scheduling process, has the core function of transforming agricultural operation requirements into standardized decision variables. After receiving crop growth status prediction signals, it combines operational parameters such as irrigation volume, nitrogen application, phosphorus application, potassium application, and pesticide application at each time step into a decision variable vector. Simultaneously, it incorporates constraints such as soil water holding capacity, crop growth period thresholds, operation intervals, and upper limits of agricultural machinery resources to construct a compliant and feasible solution space. This ensures that subsequent optimized solutions conform to actual field operation constraints, avoiding the generation of ineffective scheduling plans that cannot be implemented, and providing standardized decision inputs for multi-objective optimization. The multi-objective function construction unit is immediately followed by the decision variable encoding, focusing on the multi-dimensional optimization needs of agricultural operations. Based on crop growth status prediction signals, it constructs four core objective functions: yield, quality, resource utilization efficiency, and environmental emissions. The yield objective is mapped to a standardized interval based on the predicted yield; the quality objective is calculated with the deviation of quality indicators as the core; the resource efficiency objective is based on the resource consumption per unit of yield; and the environmental emissions objective is based on nitrogen surplus and deep seepage. Each objective function is normalized to eliminate the influence of dimensional differences, so that multi-objective optimization has a unified evaluation standard, balancing the needs of high yield, high quality, water and fertilizer conservation, and environmental protection, and avoiding the imbalance of agricultural operations caused by a single objective orientation. The multi-objective optimization solution unit, as the core generation unit of the scheduling scheme, adopts a decomposition-based multi-objective evolutionary optimization method. It receives the normalized multi-objective function and decomposes the multi-objective problem into multiple single-objective sub-problems through Chebyshev aggregation functions. Each sub-problem is matched with a unique weight vector. Using the decision variable vector as the population individuals, it generates offspring individuals using simulated binary crossover and polynomial mutation operators, and applies penalty functions to correct individuals that violate the constraints. After multiple rounds of iteration, it calculates the optimal solution of each sub-problem, maintains an external archive to store non-dominated solutions, and outputs the Pareto front solution set after the iteration terminates. It provides multiple sets of optimal alternative schemes that take into account multiple objectives, providing a sufficient and high-quality selection space for scheme selection. The scheme selection unit serves as the optimal scheme screening unit. It receives the Pareto front solution set and, combined with the user's preset decision preferences, calculates the weighted Euclidean distance between each non-dominated solution and the ideal point. The solution with the smallest distance is the optimal scheduling scheme that meets the user's needs. The decision variable vector of the optimal solution is decoded into irrigation, fertilization, and pesticide application instruction signals for each time step. These signals are then organized into a complete agricultural operation scheduling scheme signal according to the operation time window, taking into account both the customer's objective optimality and the user's subjective preferences, to ensure that the scheduling scheme is practical and meets the actual planting needs.

[0037] The instruction distribution unit, acting as the link between the scheduling and operation terminals, receives agricultural operation scheduling signals and accurately analyzes the operation type, workload, machinery allocation, and operation method instructions corresponding to each time step. It transforms abstract scheduling instructions into specific control signals recognizable by the terminals: irrigation instructions correspond to valve opening / closing duration or pump flow control signals; fertilization instructions correspond to discharge rate or discharge volume control signals; and pesticide application instructions correspond to spraying rate or spraying volume control signals. It also adds target plot identifiers and operation time window information to each instruction to ensure accurate delivery to the corresponding operation terminals, avoiding misdelivery or erroneous execution and guaranteeing the accurate implementation of the scheduling plan. The operation terminal execution unit, as the main body executing field operations, receives irrigation, fertilization, and pesticide application instruction signals from the instruction distribution unit. It drives the corresponding irrigation equipment, fertilization devices, and pesticide application machinery to complete field operations, strictly adhering to the instruction parameters. Transforming the scheduling plan into actual agricultural actions, it is the final implementation link in the entire scheduling process and directly determines the execution effect of agricultural operations. The feedback receiving unit, as the data transmission unit for execution, collects real-time actual operation data from the operation terminal, including actual irrigation volume, actual fertilization volume, actual pesticide application volume, agricultural machinery operation trajectory, and operation completion timestamp. It integrates these data into operation execution feedback signals and transmits them back to the deviation judgment unit. This accurately records the operation execution status, providing real and reliable actual execution data support for the dynamic optimization of the scheduling plan, and opening up the data link from the execution end to the decision-making end. As the core control unit of the closed-loop system, the deviation judgment unit receives the operation execution feedback signal from the feedback receiving unit, synchronously retrieves the crop growth prediction data for the corresponding time step output by the growth prediction module, calculates the deviation between the measured and predicted values ​​of key indicators such as soil moisture content and crop growth status, and forms a deviation vector. Through branch judgment logic, when any indicator in the deviation vector exceeds the preset tolerance threshold, it is determined that the current scheduling plan does not match the actual field conditions, triggers dynamic rescheduling, and restarts the entire process of encoding decision variables to plan selection. When the deviation is within a reasonable range, it is determined that the current plan is suitable for the field conditions, and the remaining scheduling plan is used. Through this branch loop mechanism, the real-time dynamic optimization of agricultural scheduling plans is realized, adapting to the dynamic changes in the field environment and crop growth, forming a closed-loop control of "decision-execution-feedback-optimization", continuously improving the accuracy, adaptability, and scientific nature of agricultural operations, and meeting the planting needs of refined smart agriculture.

[0038] The scheduling decision module receives crop growth status prediction signals from the growth prediction module and generates agricultural operation scheduling scheme signals through a dynamic scheduling algorithm based on multi-objective optimization. The execution flow of this dynamic scheduling algorithm includes decision variable encoding, multi-objective function construction, multi-objective optimization solution, and scheme selection. The decision variable encoding step combines irrigation, nitrogen, phosphorus, potassium, and pesticide application amounts for each operation time step into a decision variable vector. Each component of this vector corresponds to a specific agricultural operation at a particular time step, and the vector's dimension is determined by the total number of time steps and operation types within the operation cycle. The feasible value space of the decision variables is restricted by multiple constraints, including: a single irrigation amount not exceeding the difference between the soil field capacity and the current water content to avoid surface runoff and deep seepage; a single nitrogen application amount not exceeding the safe threshold for the current crop growth stage to prevent seedling burn and excessive absorption; a time interval between two adjacent operations of the same type not less than a preset minimum interval to meet agronomic requirements; and the allocation of agricultural machinery resources at any time step not exceeding the total number of available agricultural machinery. The multi-objective function construction steps involve building four objective functions to comprehensively evaluate the merits of the scheduling scheme. The yield objective function is based on the yield estimate from the crop growth status prediction signal. A piecewise linear function maps the yield value to a standardized interval; the higher the yield, the closer the function value is to the lower limit of the interval, indicating a better yield target. The quality objective function is calculated based on the deviation of the predicted quality index value from the standard value of the target variety. Quality indices can be indicators related to crop variety and use, such as protein content, rice yield, or oil content. A smaller deviation results in a better function value. The resource utilization efficiency objective function is calculated based on the amount of irrigation water consumed per unit of yield and the total amount of fertilizer consumed per unit of yield. Less consumption results in a better function value. The environmental emission objective function is calculated based on the weighted sum of nitrogen surplus and deep seepage. Nitrogen surplus is the difference between the total nitrogen applied and the total nitrogen absorbed by the crop. Deep seepage is the portion of the difference between the total irrigation and the total crop evapotranspiration exceeding the soil's water-holding capacity. Less emission results in a better function value. Each objective function is normalized to a unified standardized interval from minimum to maximum, ensuring comparability between objectives with different dimensions and numerical ranges. The multi-objective optimization solution employs a decomposition-based multi-objective evolutionary optimization method, decomposing the multiple objective functions into multiple single-objective subproblems using Chebyshev aggregation functions. Each subproblem corresponds to a set of weight vectors covering different directions of the objective space, with decision variable vectors serving as the encoding form for individuals in the population. During the iterative evolution of the population, a simulated binary crossover operator is used to generate offspring from parent individuals. A polynomial mutation operator is used to apply small perturbations to offspring individuals to maintain population diversity. A penalty function is applied to offspring individuals that violate constraints; the penalty function adds a penalty term proportional to the degree of violation to the objective function value, causing individuals that violate constraints to be naturally eliminated in the competition.After each iteration, the Chebyshev aggregate value of each subproblem under the current solution is calculated. If the aggregate value is better than the previously recorded best value for that subproblem, the current best solution for that subproblem is updated. Simultaneously, an external archive is maintained to store all non-dominated solutions discovered during the iteration process. A non-dominated solution is one where no other solution is inferior to it across all objectives and is strictly superior to it at least on one objective. When the number of iterations reaches a preset termination condition, all non-dominated solutions in the external archive constitute the Pareto front solution set. The scheme selection step selects a satisfactory solution from the Pareto front solution set as the execution scheme for the current work cycle. The selection method involves calculating the weighted Euclidean distance between the multidimensional objective function vector of each non-dominated solution in the Pareto front and the ideal point. The weighting weights are determined by the user's preset preference vector through the decision preference configuration interface. The preference vector reflects the user's subjective tendency regarding the relative importance of each objective—output, quality, resource efficiency, and environmental emissions. The non-dominated solution with the smallest weighted Euclidean distance is selected as the satisfactory solution. The decision variable vector of the satisfactory solution is decoded into specific operational quantities for each operation time step, generating irrigation command signals, fertilization command signals, and pesticide application command signals. These are then organized in time step order into an agricultural operation scheduling scheme signal, which is transmitted to the execution monitoring module. The formula for calculating the weighted Euclidean distance is as follows:

[0039] in is the weighted Euclidean distance between the nondominated solution and the ideal point. The smaller the distance, the better the scheduling scheme fits the user's preferences and the better the overall performance. K is the total number of optimization objectives, corresponding to the four core agricultural scheduling objectives: yield, quality, resource utilization efficiency, and environmental emissions. The preference weight for the k-th target is preset by the user through the decision preference configuration interface, reflecting the user's emphasis on different targets. For example, pursuing high yield can increase the weight of the yield target, and focusing on environmental protection can increase the weight of the environmental emission target. Let x be the k-th objective function value corresponding to the x-th scheduling scheme in the Pareto solution set, which takes a value ranging from 0 to 1 after normalization. Let be the ideal value of the k-th objective, that is, the theoretically achievable optimal value, which is usually set as the upper or lower limit of the normalization interval; For summation, a weighted integration of multi-target deviations is achieved; ⋅ is for square root operation, restoring the sum of squared deviations to distance dimensions; This is the regularization coefficient, used to balance the weights of target deviation and scheme stability. Its value is preset based on the planting scale and agricultural machinery constraints. The 2-norm of the scheduling scheme vector and the historical best scheme vector represents the degree of difference between the current scheme and the historical best scheme. Introducing this term can avoid frequent and large-scale adjustments to the scheme, ensuring the continuity and stability of agricultural operations. This algorithm combines user subjective preferences and objective goal deviations, and quantifies the comprehensive advantages and disadvantages of schemes through weighted distance. It respects the user's priority settings for core needs such as yield and quality, while ensuring that the scheme achieves an optimal balance in terms of resource conservation and environmental protection. It solves the problems of one-sided single-objective screening and low efficiency of manual screening for multiple objectives. It quickly locates agricultural scheduling schemes that take into account the optimality of multiple objectives and user preferences from the Pareto front solution set. The generated irrigation, fertilization, and pesticide application instructions are accurately adapted to crop growth needs and actual field conditions, reducing resource waste and environmental emissions, improving the accuracy, economy, and environmental protection of agricultural operations, and adapting to the scheduling needs of large-scale and refined agricultural planting. It provides an efficient and flexible scheme selection method for smart agriculture precision agricultural management.

[0040] The monitoring module receives agricultural operation scheduling plan signals output by the scheduling decision module and is responsible for converting these signals into equipment-level control commands that can be distributed to field operation terminals. This module includes a command distribution unit and a feedback receiving unit. After receiving the agricultural operation scheduling plan signals, the command distribution unit parses the irrigation, fertilization, pesticide application, machinery allocation, and operation mode commands corresponding to the current time step. It converts the irrigation amount into irrigation command signals controlling the opening and closing duration of irrigation valves or the flow rate of irrigation pumps; the fertilization amount into fertilization command signals controlling the discharge rate or amount of fertilizer from the fertilizer application device; and the pesticide application amount into pesticide application command signals controlling the spraying rate or amount of pesticide application. Each command signal is appended with a target plot identifier and operation time window information to ensure that the commands are executed at the correct time and place. After the operation is executed at the operation terminal, the feedback receiving unit receives the operation execution feedback signal returned by the terminal. This feedback signal includes the actual irrigation amount, actual fertilization amount, actual pesticide application amount, actual machinery operation trajectory, and completion timestamps for each operation, accurately recording the deviation between the actual execution and the planned scheme. The feedback receiving unit transmits the job execution feedback signal to the scheduling decision module. The scheduling decision module compares the actual execution quantity in the feedback signal with the planned quantity. Simultaneously, within each preset reassessment cycle, it acquires the observed soil moisture content and crop growth status from the data acquisition module for the current time step. It calculates the deviation vector between the observed values ​​and the predicted soil moisture content and crop growth status generated by the growth prediction module in the previous scheduling cycle. If the absolute value of any element in the deviation vector exceeds the preset tolerance threshold corresponding to that element, it indicates a significant deviation between the actual growth status and the predicted status, and the applicable conditions of the original scheduling scheme have changed, triggering the dynamic rescheduling process. Dynamic rescheduling uses the actual observed status of the current time step as the initial condition, re-executing the decision variable encoding step, multi-objective function construction step, multi-objective optimization solution step, and scheme selection step. This generates a corrected agricultural operation scheduling scheme signal to replace the unexecuted parts of the original scheme, achieving continuous adaptive adjustment of the scheduling scheme to real-time environmental changes and deviations in crop growth status.

[0041] The AI-based smart agricultural precision farming operation management system of this invention aligns and fuses multi-source farmland data, such as satellite remote sensing, ground sensors, and meteorological data, in a unified spatiotemporal grid through a spatiotemporal fusion module, forming a multimodal fusion representation signal. The growth prediction module utilizes a dual attention prediction model of temporal and spatial data, combining historical fusion features with future meteorological prediction signals, to generate a crop growth status prediction signal containing multi-dimensional indicators such as leaf area index, biomass, and yield. The scheduling decision module, based on a multi-objective optimization dynamic scheduling algorithm, uses the prediction signal as input to solve the Pareto front among yield, quality, resource efficiency, and environmental emission objectives, and generates an agricultural operation scheduling scheme signal. The execution monitoring module outputs irrigation, fertilization, and pesticide application instructions and triggers dynamic rescheduling based on operation feedback.

[0042] Therefore, the AI-based smart agricultural planting precision farming operation management system of the present invention can solve the problems of insufficient crop growth prediction accuracy and lack of multi-objective optimization mechanism for agricultural scheduling.

[0043] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. An AI-based intelligent agricultural planting precision farming operation management system, characterized in that: include: The data acquisition module collects satellite remote sensing image signals, ground sensor signals, meteorological data signals and field management record signals of the target plot in real time, forming multi-source farmland data signals; The spatiotemporal fusion module receives the multi-source farmland data signals and aligns and fuses the heterogeneous data in a unified spatiotemporal grid through a preset multimodal fusion model to form a multimodal fusion representation signal. The growth prediction module receives the multimodal fusion representation signal and the future weather prediction signal, and generates a crop growth status prediction signal through a preset temporal and spatial dual attention prediction model. The scheduling decision module receives the crop growth status prediction signal and generates an agricultural operation scheduling scheme signal through a dynamic scheduling algorithm based on multi-objective optimization. The execution monitoring module receives the agricultural operation scheduling plan signal, outputs irrigation command signal, fertilization command signal and pesticide application command signal to the operation terminal, and receives the operation execution feedback signal returned by the operation terminal. The operation execution feedback signal is then transmitted to the scheduling decision module to trigger dynamic rescheduling.

2. The AI-based intelligent agricultural planting precision farming operation management system according to claim 1, characterized in that, The data acquisition module includes a satellite remote sensing image acquisition unit, a ground sensor acquisition unit, a meteorological data acquisition unit, and a field management record entry unit. The satellite remote sensing image acquisition unit acquires multispectral image sequences of the target plot in various bands by accessing a satellite remote sensing data platform. The ground sensor acquisition unit acquires soil temperature, soil moisture, soil conductivity, light intensity, air temperature, and air humidity values ​​by using soil temperature sensors, soil moisture sensors, soil conductivity sensors, light intensity sensors, air temperature sensors, and air humidity sensors deployed within the target plot, respectively. The meteorological data acquisition unit acquires precipitation, wind speed, atmospheric pressure, and solar radiation flux of the target plot area by accessing a meteorological data service platform. The field management record entry unit acquires irrigation records, fertilizer type and amount records, tillage time records, and agricultural machinery operation trajectory records of the target plot. The data acquisition module adds spatial coordinate labels and timestamp labels to the data from the above data sources and assembles them into the multi-source farmland data signal.

3. The AI-based intelligent agricultural planting precision farming operation management system according to claim 2, characterized in that, The preset multimodal fusion model includes a spatiotemporal alignment layer, a channel-specific feature extraction layer, and a cross-modal attention fusion layer. The spatiotemporal alignment layer maps satellite remote sensing imagery, ground sensor data, meteorological data, and field management records from the multi-source farmland data signals to a spatiotemporal reference system with the target plot boundary as the spatial range and the crop sowing date as the time origin. Spatially, it transforms the spatial coordinates of each data source to a unified spatial reference system and aggregates them using grid cells of a preset resolution. Temporally, it resamples the time series data from each data source according to a preset time step to align them to a unified time axis, forming multidimensional tensor data. The channel-specific feature extraction layer outputs the multidimensional tensor data separately for remote sensing channels, ground sensor channels, and meteorological channels. The corresponding feature extraction subnetworks are fed into the remote sensing channel, which uses a 3D convolutional neural network to perform spatiotemporal joint feature extraction along the spatial grid and time step. The ground sensor channel uses a gated recurrent unit network to process the multivariate time series of each sensor. The meteorological channel uses a gated recurrent unit network to process the multivariate time series of the meteorological grid unit. The feature vectors output by each channel are concatenated into a multi-channel feature matrix after layer normalization. The cross-modal attention fusion layer calculates the attention weight matrix between remote sensing features and meteorological features, as well as the attention weight matrix between ground sensor features and meteorological features. The cross-modal representation vector is obtained by weighted summation. The cross-modal representation vector is concatenated with the multi-channel feature matrix and then dimensionality is reduced by a fully connected layer to output the multi-modal fused representation signal.

4. The AI-based intelligent agricultural planting precision farming operation management system according to claim 3, characterized in that, The specific process for calculating the attention weight matrix between remote sensing features and meteorological features in the cross-modal attention fusion layer is as follows: The remote sensing feature vector is transformed into a remote sensing query vector using a first linear projection matrix; the meteorological feature vector is transformed into a meteorological key vector and a meteorological value vector using a second linear projection matrix; the dot product similarity between the remote sensing query vector and the meteorological key vector is calculated; the dot product similarity is scaled by dividing by the square root of the dimension of the meteorological key vector; and a normalized exponential function is applied to the scaled similarity matrix to obtain the normalized attention weight. The normalized attention weight represents the degree of correlation between the remote sensing observation features at each spatial location of the target plot and the meteorological conditions of each meteorological grid unit. The normalized attention weights are used to weight and sum the meteorological value vectors to generate a cross-modal representation vector of remote sensing and meteorology. Simultaneously, when calculating the attention weight matrix between ground sensor features and meteorological features, the ground sensor feature vectors are transformed into sensor query vectors using a third linear projection matrix. Using the meteorological key vector and meteorological value vector as shared key-value pairs, the dot product similarity between the sensor query vector and the meteorological key vector is calculated and mapped using scaling and a normalized exponential function to obtain the normalized attention weights for ground sensor and meteorology. These weights are then used to weight and sum the meteorological value vectors to generate the cross-modal representation vector of ground sensor and meteorology. The calculation formula for the attention weight matrix is ​​as follows: ； in, This is the remote sensing-meteorological attention weight matrix. For normalized exponential functions, To query the projection matrix, This is the remote sensing feature matrix. The key projection matrix, The meteorological feature matrix, For matrix transpose, The dimension of the key vector. This is the bias matrix.

5. The AI-based intelligent agricultural planting precision farming operation management system according to claim 1, characterized in that, The preset temporal and spatial dual attention prediction model employs a network structure of stacked encoder and decoder. The encoder receives the multimodal fusion representation signal and expands it into a feature sequence along the time dimension. The encoder is composed of alternating stacked layers of temporal self-attention layers and layers of feedforward network layers. Each temporal self-attention layer generates a query vector, key vector, and value vector for each time step position in the feature sequence through linear projection. The dot product similarity between the query vector and the key vector between any two time steps in the sequence is calculated and mapped to temporal self-attention weights through a normalized exponential function. The temporal self-attention weights represent the relative importance of each time step in the historical feature sequence to the prediction of the growth state of the current time step. The value vectors are weighted and summed using the temporal self-attention weights, and then processed through residual connections and layer normalization. The encoder hidden state is processed and output. The decoder receives the encoder hidden state and the future weather prediction signal. The decoder consists of multiple cross-attention layers. Each cross-attention layer uses the query vector generated by linear projection of the future weather prediction signal as the query and the key vector and value vector generated by linear projection of the encoder hidden state as key-value pairs. It calculates the cross-attention weight between future weather conditions and historical growth characteristics. The value vector of the encoder hidden state is weighted and summed with the cross-attention weight, and after residual connection and layer normalization processing, the decoder hidden state is output. The decoder hidden state is mapped and output by a multi-task prediction head network to the predicted values ​​of leaf area index, aboveground biomass accumulation, soil moisture content, nitrogen accumulation, and yield estimate, which together constitute the crop growth status prediction signal.

6. The AI-based intelligent agricultural planting precision farming operation management system according to claim 5, characterized in that, The multi-task prediction head network includes a shared feature transformation layer and multiple task-specific output layers. The shared feature transformation layer receives the decoder's hidden state and maps it to a shared feature representation through a multi-layer fully connected network and a non-linear activation function. The multiple task-specific output layers correspond to leaf area index prediction, aboveground biomass accumulation prediction, soil moisture prediction, nitrogen accumulation prediction, and yield estimation tasks, respectively. Each task-specific output layer contains an independent linear transformation matrix that maps the shared feature representation to the numerical output of the corresponding prediction target. The loss function of the preset temporal and spatial dual attention prediction model during the training phase is the weighted sum of the mean square errors between the predicted and measured values ​​for each task. The loss weights for each task are automatically learned through an uncertainty weighting method. The calculation formula for the loss function is as follows: ； in, Here, T represents the joint loss function for multiple tasks, and T is the total number of prediction tasks. Let Variance be the uncertainty of the t-th task. Let be the mean square error function. Let t be the predicted value for the t-th task. Let t be the measured value of the t-th task.

7. The AI-based intelligent agricultural planting precision farming operation management system according to claim 1, characterized in that, The dynamic scheduling algorithm based on multi-objective optimization includes a decision variable encoding step, a multi-objective function construction step, a multi-objective optimization solution step, and a scheme selection step. The decision variable encoding step combines the irrigation amount, nitrogen application amount, phosphorus application amount, potassium application amount, and pesticide application amount for each operation time step into a decision variable vector. A feasible solution space is constructed based on constraints such as the single irrigation amount not exceeding the difference between the soil field capacity and the current water content, the single nitrogen application amount not exceeding the safe threshold of the crop's current growth stage, the time interval between two adjacent operations of the same type not being less than a preset minimum interval, and the upper limit of available agricultural machinery resources. The multi-objective function construction... The steps involve constructing a yield objective function, a quality objective function, a resource utilization efficiency objective function, and an environmental emission objective function. The yield objective function is based on the yield estimate in the crop growth status prediction signal and mapped to a standardized interval using a piecewise linear function. The quality objective function is calculated based on the deviation of the predicted quality index value from the standard value of the target variety. The resource utilization efficiency objective function is calculated based on the amount of irrigation water and total fertilizer consumed per unit yield. The environmental emission objective function is calculated based on the weighted sum of nitrogen surplus and deep seepage. Each objective function is normalized to a unified standardized interval from minimum to maximum.

8. The AI-based intelligent agricultural planting precision farming operation management system according to claim 7, characterized in that, The multi-objective optimization solution step employs a decomposition-based multi-objective evolutionary optimization method. The multiple objective functions constructed in the multi-objective function construction step are decomposed into multiple single-objective sub-problems using a Chebyshev aggregation function. Each sub-problem corresponds to a set of weight vectors covering different directions of the objective space. Decision variable vectors are used as population individual codes. Simulated binary crossover and polynomial mutation operators are used to generate offspring individuals. Penalty functions are applied to offspring individuals that violate the constraints. After each iteration, the Chebyshev aggregation value of each sub-problem is calculated, and the current optimal solution for each sub-problem is updated. Simultaneously, an external archive is maintained to store non-dominated solutions discovered during the iteration process. When a preset iteration condition is met... When the termination condition is met, all non-dominated solutions in the external archive are output as the Pareto front solution set. The scheme selection step selects a satisfactory solution from the Pareto front solution set. The selection method is to calculate the weighted Euclidean distance between the multidimensional objective function vector of each non-dominated solution in the Pareto front and the ideal point. The weighting weight is determined by the user through the preference vector preset by the decision preference configuration interface. The non-dominated solution with the smallest weighted Euclidean distance is selected. The decision variable vector of this solution is decoded into irrigation command signal, fertilization command signal and pesticide application command signal for each operation time step. These are organized in time step order to form the agricultural operation scheduling scheme signal. The formula for calculating the weighted Euclidean distance is as follows: ； in, The weighted Euclidean distance is K, where K is the number of targets to optimize. The preference weight for the k-th objective is... To solve for the k-th objective value of x, For the k-th target ideal value, This is the regularization coefficient.

9. The AI-based intelligent agricultural planting precision farming operation management system according to claim 8, characterized in that, After transmitting the agricultural operation scheduling scheme signal to the execution monitoring module, the scheduling decision module continuously receives the operation execution feedback signal returned by the execution monitoring module. The operation execution feedback signal includes irrigation execution amount, fertilization execution amount, pesticide application execution amount, and operation timestamp. At the same time, in each preset reassessment cycle, the scheduling decision module obtains the soil moisture content observation value and crop growth status observation value at the current time step from the data acquisition module, calculates the deviation vector between the soil moisture content observation value and crop growth status observation value and the soil moisture content prediction value and crop growth status prediction value at the corresponding time step generated by the growth prediction module in the previous scheduling cycle. If the absolute value of any element in the deviation vector exceeds the preset tolerance threshold corresponding to that element, the dynamic rescheduling is triggered. The dynamic rescheduling starts with the soil moisture content observation value and crop growth status observation value at the current time step as the initial state, and re-executes the decision variable encoding step, multi-objective function construction step, multi-objective optimization solution step, and scheme selection step to generate a corrected agricultural operation scheduling scheme signal to replace the unexecuted part of the original scheme.

10. The AI-based intelligent agricultural planting precision farming operation management system according to claim 1, characterized in that, The execution monitoring module includes an instruction distribution unit and a feedback receiving unit. The instruction distribution unit receives the agricultural operation scheduling plan signal, parses the irrigation amount, fertilizer amount, pesticide amount, agricultural machinery allocation, and operation mode instructions corresponding to the current time step in the agricultural operation scheduling plan signal, converts the irrigation amount into an irrigation instruction signal to control the opening and closing time of irrigation valves or the flow rate of irrigation pumps, converts the fertilizer amount into a fertilizer instruction signal to control the discharge rate or discharge amount of fertilizer application device, and converts the pesticide amount into a pesticide application instruction signal to control the spraying rate or spraying amount of pesticide application device. Each instruction signal is accompanied by a target plot identifier and operation time window information. The feedback receiving unit receives the operation execution feedback signal returned by the operation terminal after the operation is performed. The operation execution feedback signal includes the actual irrigation amount, actual fertilizer amount, actual pesticide amount, actual agricultural machinery operation trajectory, and completion timestamp of each operation. The feedback receiving unit transmits the operation execution feedback signal to the scheduling decision module.

Images