A deep learning-based crop growth intelligent decision-making method and system

Abstract

This invention discloses a deep learning-based intelligent decision-making method and system for crop growth. It acquires multimodal agricultural data of a target plot over continuous time series, including remote sensing phenotypic data, meteorological and environmental data, soil condition data, and historical agricultural operation logs. The data undergoes spatiotemporal alignment and missing value completion to construct a spatiotemporal graph with plots and time points as nodes and temporal and spatial proximity as edges. Complete time-series data is generated using a graph attention mechanism with causal masking. This complete time-series data is then input into a multi-scale causal perception encoder to extract short-term dynamic features and long-term causal dependency features, which are then fused to obtain causal latent states. Based on these causal latent states, an intervention simulation under a structural causal model is executed through a dynamic counterfactual decision generator to solve an optimization problem with agricultural operation constraints, generating the optimal agricultural operation sequence within a preset number of days and the corresponding counterfactual yield prediction results.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of smart agriculture and artificial intelligence, and relates to a method and system for intelligent decision-making on crop growth based on deep learning. Background Technology

[0002] In modern agricultural production, crop growth management decisions have long relied on empirical judgments or static rule models. Traditional methods struggle to effectively integrate multi-source, heterogeneous information such as remote sensing phenotypic data, meteorological and environmental data, soil condition data, and historical agricultural operation logs. Existing intelligent agriculture systems often employ single-modal deep learning models, relying solely on remote sensing images or meteorological sequences for yield prediction, neglecting the asynchronicity and inherent correlations of different data sources over time. Some solutions attempt to integrate multimodal data through early or late fusion strategies, but fail to consider causal relationships between variables, easily misinterpreting statistical correlations as causal drivers, leading to decision recommendations that deviate from actual agronomic logic.

[0003] Furthermore, mainstream models such as Long Short-Term Memory (LSTM) networks typically employ bidirectional structures or recurrent mechanisms without causal constraints, which may introduce future information leakage in online decision-making scenarios, leading to response delays or logical errors. Current irrigation or fertilization recommendation systems often output black-box instructions, lacking explanatory explanations of "why to execute" and "what the consequences would be if not to execute," resulting in low farmer adoption rates. Although some studies have introduced reinforcement learning or expert rules for optimization, the former requires extensive field trial and error, while the latter struggles to adapt to changes in regional environments.

[0004] Therefore, there is an urgent need for an intelligent decision-making method that can achieve multimodal data collaborative completion, explicit modeling of causal dependencies between the environment and crops, and support counterfactual intervention simulation, in order to solve the technical bottlenecks of data fragmentation, weak model generalization ability, delayed response, and uninterpretable decision-making. Summary of the Invention

[0005] To address the problems existing in the background technology, this invention proposes a deep learning-based intelligent decision-making method and system for crop growth.

[0006] The first aspect of this application provides a deep learning-based intelligent decision-making method for crop growth, including: Acquire multimodal agricultural data of the target plot over a continuous time series, including remote sensing phenotypic data, meteorological and environmental data, soil condition data, and historical agricultural operation logs; The multimodal agricultural data is spatiotemporally aligned and missing values ​​are filled in to construct a spatiotemporal graph with plots and time points as nodes and temporal and spatial proximity as edges. Based on this spatiotemporal graph, complete time-series data is generated through a graph attention mechanism with causal masking. The complete time-series data is input into a multi-scale causal sensing encoder, which includes a causal convolutional submodule and a Granger causal guided long short-term memory network branch, to extract short-term dynamic features and long-term causal dependency features respectively, and then fuse them to obtain the causal hidden state. Based on the aforementioned causal hidden state, an intervention simulation under the structural causal model is executed through a dynamic counterfactual decision generator to solve an optimization problem with agricultural operation constraints, generating the optimal agricultural operation sequence and corresponding counterfactual yield prediction results within a preset number of days in the future.

[0007] Optionally, an edge exists between any two nodes in the spatiotemporal graph if and only if the corresponding time difference does not exceed seven days and the corresponding spatial distance does not exceed five hundred meters.

[0008] Optionally, the graph attention mechanism with causal masking, when calculating the attention weight of a node, forces the attention weight of neighboring nodes whose timestamps are later than the current node to be reset to zero.

[0009] Optionally, the causal convolution submodule uses a one-dimensional causal convolution kernel to perform convolution operations on the complete time series data, and the convolution kernel only covers the data at the current time and previous time.

[0010] Optionally, the Granger causality-guided long short-term memory network branch introduces a Granger causality strength coefficient in the input gating of the long short-term memory network, and the Granger causality strength coefficient is calculated in real time through a sliding window vector autoregressive model.

[0011] Optionally, the dynamic counterfactual decision generator eliminates the interference of confounding factors on the estimation of causal effects through a dual machine learning method, and calculates the counterfactual yield prediction results after performing specific agricultural operations based on a differentiable causal model.

[0012] Optionally, the dynamic counterfactual decision generator eliminates the interference of confounding factors on the estimation of causal effects through a dual machine learning method, and calculates the counterfactual yield prediction results after performing specific agricultural operations based on a differentiable causal model.

[0013] A second aspect of this application provides a deep learning-based intelligent decision-making system for crop growth, comprising: The acquisition module is used to acquire remote sensing phenotypic data, meteorological and environmental data, soil condition data, and historical agricultural operation logs of the target plot over continuous time series. The completion module is used to construct a spatiotemporal graph with land parcels and time points as nodes and temporal and spatial proximity as edges, and to generate complete time series data based on the spatiotemporal graph through a graph attention mechanism with causal masking. The encoding module includes a causal convolution submodule and a Granger causal guided long short-term memory network submodule, which are used to extract short-term dynamic features and long-term causal dependency features respectively, and fuse them to obtain causal hidden states; The generation module is used to perform intervention simulations under the structural causal model based on the causal hidden state, solve optimization problems with agricultural operation constraints, and generate the optimal agricultural operation sequence and corresponding counterfactual yield prediction results within a preset number of days in the future.

[0014] Optionally, in the spatiotemporal graph constructed by the completion module, there is a connecting edge between any two nodes if and only if the corresponding time difference does not exceed seven days and the corresponding spatial distance does not exceed five hundred meters.

[0015] Optionally, the Granger causal guided long short-term memory network submodule in the encoding module has its input gating weights modulated by the Granger causal strength coefficient, which is calculated online by a sliding window vector autoregressive model.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a deep learning-based intelligent decision-making method and system for crop growth. By constructing a spatiotemporal graph and introducing a graph attention mechanism with causal masking, it effectively solves the problem of missing data in multimodal agricultural data caused by asynchronous sampling, equipment failure, or weather interference, ensuring that the data completion process strictly follows the temporal causality law and avoiding future information leakage. A multi-scale causal perception encoder integrates a causal convolutional submodule with a Granger causal-guided long short-term memory network branch, simultaneously capturing short-term dynamic responses under sudden environmental disturbances and the true causal dependencies in long-term agronomic processes, significantly improving the accuracy and physical interpretability of feature representation. A dynamic counterfactual decision generator performs intervention simulations based on a structural causal model, combining dual machine learning to eliminate the influence of confounding factors, ensuring that agricultural recommendations are based on reliable causal inference. The optimization problem with agricultural operation constraints models yield targets and resource costs in a unified manner, embedding actual agronomic constraints such as the upper limit of single irrigation volume, the safe interval for fertilization, and the feasibility set of operation types, ensuring that the output plan is executable and implementable. The system not only generates the optimal agricultural operation sequence but also provides counterfactual yield prediction results, intuitively demonstrating the potential consequences of different decision paths, enhancing farmers' understanding and trust in intelligent recommendations. The overall solution achieves a closed loop from data perception to causal understanding and then to explainable decision-making, improving the accuracy, timeliness, and resource utilization efficiency of crop management. Attached Figure Description

[0017] Figure 1 This is a flowchart of a deep learning-based intelligent decision-making method for crop growth in one embodiment of the present invention; Figure 2This is a schematic diagram of a deep learning-based intelligent decision-making system for crop growth, according to one embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] In one embodiment, such as Figure 1 As shown, a deep learning-based intelligent decision-making method for crop growth is provided, which is then applied to... Figure 1 Taking China as an example, the following specific steps will be used: S10: Obtain multimodal agricultural data of the target plot over a continuous time series. The multimodal agricultural data includes remote sensing phenotypic data, meteorological and environmental data, soil condition data, and historical agricultural operation logs.

[0020] Specifically, in this invention, acquiring multimodal agricultural data of a target plot over a continuous time series refers to continuously collecting observational information related to crop growth status from multiple sources and organizing this information according to a unified time benchmark. The multimodal agricultural data includes four categories: remote sensing phenotypic data, meteorological and environmental data, soil condition data, and historical agricultural operation logs.

[0021] Remote sensing phenotypic data is image data obtained through periodic observations of target land parcels using multispectral imaging equipment mounted on satellite platforms or drones. This data records the reflectance values ​​of each pixel in multiple spectral bands in a two-dimensional spatial grid format. Typical spectral bands include the red, near-infrared, and green bands. Based on these bands, phenotypic indices such as the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EDI), and Leaf Area Index (LAI) can be calculated. The temporal resolution of remote sensing phenotypic data is typically every five days, and the spatial resolution is ten meters.

[0022] Meteorological data is provided by automatic weather stations deployed within a five-kilometer radius of the target site. This data is collected daily and includes daily average temperature, daily maximum temperature, daily minimum temperature, daily cumulative precipitation, relative humidity, sunshine duration, and wind speed. All meteorological parameters are expressed in the International System of Units (SI), such as temperature in degrees Celsius, precipitation in millimeters, and wind speed in meters per second.

[0023] Soil condition data were obtained from a network of soil sensors embedded in the topsoil of the target plot. The sensors were deployed at a density of at least three monitoring points per hectare, at a depth of 20 to 30 centimeters. The collected parameters included volumetric water content, soil temperature, pH, electrical conductivity, and the concentrations of nitrogen, phosphorus, and potassium. Data was sampled every two hours, and the system calculated the arithmetic mean of the twenty-four samples taken daily as the soil condition data for that day.

[0024] Historical agricultural operation logs are manually entered by farmers via a mobile application or automatically uploaded by the intelligent agricultural machinery operation control system. The log records, in Boolean vector form, whether a specific type of agricultural operation was performed each day. Operation types include irrigation, fertilization, spraying, tilling, and harvesting. If irrigation was performed on a given day, the corresponding irrigation dimension is marked as 1; otherwise, it is marked as 0. All operation records are bound to specific dates and plot identifiers to ensure strict alignment with other modal data on the timeline.

[0025] In practical applications, the above four types of data are synchronously collected to the central data processing platform. The platform uses the Gregorian calendar date as an index to construct daily data records. For dates where a certain type of data is missing, the system retains the empty field and hands it over to the subsequent completion module for processing. In this way, the system can establish a structured multimodal time-series database covering the entire crop growth period, providing a complete input foundation for subsequent causal perception coding and intelligent decision-making. This data acquisition mechanism effectively integrates space-air-ground integrated observation resources, overcoming the technical obstacles of isolated data sources, inconsistent formats, and asynchronous time in traditional agricultural information systems, thus laying the data foundation for achieving accurate, dynamic, and interpretable crop growth management decisions.

[0026] S20: Perform spatiotemporal alignment and missing value completion on the multimodal agricultural data, construct a spatiotemporal graph with plots and time points as nodes and temporal and spatial proximity as edges, and generate complete time-series data based on the spatiotemporal graph through a graph attention mechanism with causal masking.

[0027] Specifically, spatiotemporal alignment and missing value completion of the multimodal agricultural data are prerequisite steps for achieving high-quality decision-making in this invention. This step first maps remote sensing phenotypic data, meteorological environmental data, soil condition data, and historical agricultural operation logs from different sources onto a timeline based on the Gregorian calendar. Each plot constitutes an independent data record unit for each day, referred to as the plot and time point. Even if some data is missing on a certain day, the plot and time point are still retained, and the corresponding fields are marked as empty.

[0028] Based on this, the system constructs a spatiotemporal map. The node set of this spatiotemporal map consists of every target plot and time point within the observation period. An edge is established between any two nodes if and only if the following two conditions are met: the time difference between the two nodes does not exceed seven days; and the Euclidean distance between the spatial locations of the two nodes does not exceed 500 meters. Spatial locations are represented by the latitude and longitude coordinates of the plot's center point and converted to metric units in a Cartesian coordinate system for distance calculation.

[0029] After constructing the spatiotemporal graph, the system employs a graph attention mechanism with causal masking to complete missing values. Specifically, for any node to be completed, its missing feature value is generated by weighted aggregation of its neighboring nodes using graph attention weights. The attention weights are calculated based on the dot product of the learnable query vector and the key vector, and are then scaled and normalized. The key is the introduction of a causal mask: if the timestamp of a neighboring node is later than the timestamp of the current node, its corresponding attention weight is forcibly set to zero. This mechanism ensures that the completion process strictly adheres to temporal causality, meaning that future information is not leaked into the estimation of the current or past moments.

[0030] In one embodiment, continuous rain in a rice-growing area prevented drones from acquiring remote sensing phenotypic data, resulting in a three-day missing NDVI value. Simultaneously, some soil sensors experienced transmission interruptions due to malfunctions. The system, using the aforementioned spatiotemporal map structure, combined with valid observation data from neighboring plots during the same period and valid data from previous periods in the local plot, along with continuous meteorological records provided by weather stations, reconstructed a complete multimodal time series. The completed data maintained the continuity of crop growth trends and regional spatial consistency.

[0031] This spatiotemporal alignment and missing value completion method effectively solves the common problems of asynchronous sampling, equipment failure, and weather interference in agricultural data collection. By fusing temporal and spatial proximity, it fully utilizes the similarity between plots and the temporal evolution patterns. The introduction of causal masking avoids non-physical information backflow, ensuring the temporal logical rationality of the completion results. The resulting complete time-series data provides a reliable input foundation for subsequent causal feature extraction and intelligent decision-making.

[0032] S30: Input the complete time series data into a multi-scale causal sensing encoder, which includes a causal convolutional submodule and a Granger causal guided long short-term memory network branch, to extract short-term dynamic features and long-term causal dependency features respectively, and fuse them to obtain causal hidden states.

[0033] Specifically, inputting the complete time-series data into a multi-scale causal perception encoder is the core step in achieving a deep understanding of crop growth status in this invention. This encoder consists of two parallel processing branches: a causal convolutional submodule and a Granger causal guided long short-term memory network branch. The two branches focus on capturing crop response features at different time scales and ultimately fuse them to generate a unified causal hidden state.

[0034] The causal convolution submodule is used to extract short-term dynamic features. This branch uses a one-dimensional causal convolution structure to process the complete time-series data day by day. A key characteristic of causal convolution is that its kernel only covers data points from the current time and previous times, without relying on future information. In this invention, the temporal receptive field of the convolution kernel is set to five days, meaning that the features at each output time are calculated from the input data of the most recent five consecutive days. This design enables the model to quickly respond to sudden environmental disturbances, such as sudden rainfall, high temperatures, or short-term drought events, thereby capturing the short-term physiological stress responses of crops.

[0035] Granger causality guides a branch of the Long Short-Term Memory (LSTM) network to extract long-term causal dependence features. This branch is based on the LTM network architecture and incorporates Granger causality strength coefficients into its input gating mechanism. Granger causality strength coefficients reflect whether the historical value of one variable helps predict the current value of another variable. The system performs online analysis of complete time-series data using a sliding window vector autoregressive (VAR) model, calculating the Granger causality strength between various agricultural variables in real time. For example, the influence of early effective accumulated temperature on spikelet differentiation is quantified and used to modulate the weights of the corresponding input channels in the LTM network. This mechanism allows the model to focus on driving factors with true predictive power and suppresses spurious correlations.

[0036] The output feature vectors of the two branches are aligned in the time dimension and then concatenated. A layer normalization operation is then performed to generate the final causal hidden state. This causal hidden state simultaneously contains short-term perturbation response information and long-term causal evolution patterns, providing a high-dimensional semantic representation for subsequent decision-making modules.

[0037] In this example, a wheat field experienced three consecutive days of high temperatures during the jointing stage. The causal convolutional submodule quickly identified short-term patterns of abnormally high leaf temperature and abrupt changes in transpiration rate. Simultaneously, the Granger causality-guided long short-term memory network branch confirmed a significant Granger causal relationship between previous soil nitrogen content and the current tiller number, thereby strengthening the feature representation of this path. The fused causal latent state accurately reflects the growth inhibition risk and the fundamental supporting role of nitrogen supply under high-temperature stress.

[0038] This multi-scale causal sensing encoding mechanism overcomes the shortcomings of traditional single-time-scale models, such as slow response to sudden disturbances or insufficient modeling of long-term trends. By explicitly introducing causal constraints, the model not only learns the statistical relationships between variables but also identifies driving relationships with directional and predictive significance. The resulting causal latent states possess stronger physical interpretability and generalization ability, laying a feature foundation for generating reliable agricultural decisions.

[0039] S40: Based on the causal hidden state, an intervention simulation under the structural causal model is executed through a dynamic counterfactual decision generator to solve the optimization problem with agricultural operation constraints, and generate the optimal agricultural operation sequence and the corresponding counterfactual yield prediction results within a preset number of days in the future.

[0040] Specifically, the key step in this invention to achieve interpretable and interventionable agricultural decisions is to perform intervention simulations under a structural causal model based on the aforementioned causal latent state through a dynamic counterfactual decision generator. This process first treats the causal latent state as a state representation of the current crop growth system, and then constructs a differentiable structural causal model based on this. This model treats agricultural operations as external intervention variables, using the do operator form to represent human interventions in the system, such as performing irrigation or applying a certain amount of nitrogen fertilizer.

[0041] The dynamic counterfactual decision generator eliminates the interference of confounding factors on causal effect estimation through a dual machine learning approach. Confounding factors include inherent land attributes such as soil texture, historical fertility levels, and varietal genetic characteristics. The system first trains an auxiliary network to predict the dependencies of multimodal agricultural data and crop growth indicators on confounding factors. Then, it uses residual signals in the main causal model to estimate unbiased causal effects. Thus, the model can accurately answer counterfactual questions: how will future crop growth indicators change if a specific agricultural operation is performed in the current state; and what will happen if the operation is not performed?

[0042] Based on this, the system constructs a constrained optimization problem. The optimization objective is to maximize the expected values ​​of crop growth indicators within a predetermined number of days, while minimizing resource consumption costs. Crop growth indicators include biomass accumulation rate, leaf area index, and number of panicles. Resource consumption costs encompass irrigation water consumption, fertilizer application, and operational energy consumption. Constraints include a single irrigation volume not exceeding 50 mm, a fertilization interval of at least seven days, and the operation type must belong to the set of feasible agricultural operations. The set of feasible agricultural operations is jointly defined by local agronomic standards and equipment capabilities.

[0043] After solving the optimization problem, the system outputs the optimal sequence of agricultural operations for the next preset number of days. This sequence specifies the type of operation to be performed each day and its quantitative parameters; for example, irrigating 30 mm on day t+1 and applying 10 kg of urea per hectare on day t+5. Simultaneously, the system generates corresponding counterfactual yield predictions, including the expected growth trajectory under this sequence and comparative trajectories under several key alternative scenarios, such as no irrigation or delayed fertilization.

[0044] In this example, a cornfield enters a critical water demand stage during the large trumpet stage. A dynamic counterfactual decision generator simulates three intervention options based on the current causal latent state: immediate irrigation of 40 mm, delayed irrigation for three days, or no irrigation. The model combines recent precipitation forecasts and soil moisture trends to calculate the leaf area index evolution path for the next seven days under each option. After optimization, immediate irrigation is recommended, and a counterfactual comparison is output showing that delayed or no irrigation will significantly increase the risk of leaf curling.

[0045] This decision-making mechanism overcomes the limitations of traditional black-box recommendation systems. Through structural causal models and counterfactual reasoning, the system not only provides operational suggestions but also verifiable causal logic support. A constrained optimization framework ensures that the suggestions are feasible for actual agricultural production. The resulting decisions are highly interpretable, traceable, and operable, effectively improving farmers' willingness to adopt them and the accuracy of their implementation.

[0046] In one embodiment, in step S40, the objective function of the optimization problem with agricultural operation constraints is the expected value of crop growth indicators within a preset number of days in the future minus the resource cost term. The constraints include the upper limit of a single irrigation volume, the safe interval for fertilization, and the set of feasible operation types.

[0047] Specifically, the objective function of the optimization problem with agricultural operation constraints is the expected value of crop growth indicators within a preset number of days in the future minus the resource cost term. This objective function is used to rationally control agricultural production inputs while ensuring crop growth potential. Crop growth indicators include leaf area index, biomass accumulation rate, and panicle differentiation number. These indicators are predicted by a dynamic counterfactual decision generator based on a structural causal model. The expected value is obtained by averaging Monte Carlo sampling of future environmental uncertainties to reflect the impact of random factors such as weather fluctuations. The resource cost term consists of a weighted sum of irrigation water costs, fertilizer application costs, and machinery operation costs. The cost coefficients are set according to local market prices and energy consumption standards, with units of yuan per cubic meter, yuan per kilogram, and yuan per operation, respectively.

[0048] The constraints include a maximum single irrigation volume, a safe interval for fertilization, and a set of feasible operation types. The maximum single irrigation volume is set at 50 mm, determined based on soil infiltration capacity and surface runoff risk to prevent water waste and nutrient leaching. The safe interval for fertilization requires a minimum of seven days between applications of the same type of fertilizer; this constraint is based on crop nutrient uptake kinetics and fertilizer residue effects to avoid seedling burn or nutrient imbalance. The set of feasible operation types defines the categories of agricultural operations permitted at the current growth stage; for example, intertillage is prohibited during the heading stage, and nitrogen fertilization is prohibited during the maturity stage. This set is jointly determined by an agronomic expert rule base and the current growth stage of the plot.

[0049] In this example, a rice paddy is in the booting stage, and the system needs to decide on the water and fertilizer management plan for the next seven days. The objective function comprehensively evaluates the impact of different irrigation and fertilization combinations on the expected number of grains per panicle, deducting the corresponding water, electricity, and urea costs. During the optimization process, the algorithm excludes plans with a single irrigation exceeding 50 mm, eliminates topdressing options less than seven days after the last fertilization, and disables out-of-season operations such as harvesting and spraying. The final optimal sequence output includes one 30 mm irrigation and one application of 7.5 kg / ha of potash fertilizer, which satisfies both the high-yield requirement and the principle of efficient resource utilization.

[0050] The optimization problem is designed to explicitly model agricultural production goals and resource constraints, ensuring that the decision-making results are both technically sound and economically feasible. By strictly defining the operational boundaries, it avoids suggestions that are theoretically optimal but practically infeasible. The joint optimization of expected value and cost in the objective function reflects the dual orientation of yield and sustainability in precision agriculture. The resulting agricultural operation sequence can be directly integrated with intelligent agricultural machinery execution systems, achieving seamless integration from decision-making to implementation.

[0051] In one embodiment, such as Figure 2 As shown, a deep learning-based intelligent decision-making system for crop growth is provided. This system corresponds one-to-one with the deep learning-based intelligent decision-making method for crop growth in the above embodiments. The deep learning-based intelligent decision-making system for crop growth includes: an acquisition module, a completion module, an encoding module, and a generation module. Detailed descriptions of each functional module are as follows: The acquisition module is used to acquire remote sensing phenotypic data, meteorological and environmental data, soil condition data, and historical agricultural operation logs of the target plot over continuous time series. The completion module is used to construct a spatiotemporal graph with land parcels and time points as nodes and temporal and spatial proximity as edges, and to generate complete time series data based on the spatiotemporal graph through a graph attention mechanism with causal masking. The encoding module includes a causal convolution submodule and a Granger causal guided long short-term memory network submodule, which are used to extract short-term dynamic features and long-term causal dependency features respectively, and fuse them to obtain causal hidden states; The generation module is used to perform intervention simulations under the structural causal model based on the causal hidden state, solve optimization problems with agricultural operation constraints, and generate the optimal agricultural operation sequence and corresponding counterfactual yield prediction results within a preset number of days in the future.

[0052] Optionally, in the spatiotemporal graph constructed by the completion module, there is a connecting edge between any two nodes if and only if the corresponding time difference does not exceed seven days and the corresponding spatial distance does not exceed five hundred meters.

[0053] Optionally, the Granger causal guided long short-term memory network submodule in the encoding module has its input gating weights modulated by the Granger causal strength coefficient, which is calculated online by a sliding window vector autoregressive model.

[0054] Specific limitations regarding the deep learning-based intelligent decision-making system for crop growth can be found in the limitations of the deep learning-based intelligent decision-making method for crop growth described above, and will not be repeated here. Each module in the aforementioned deep learning-based intelligent decision-making system for crop growth can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0055] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based intelligent decision-making method for crop growth, characterized in that, include: Acquire multimodal agricultural data of the target plot over a continuous time series, including remote sensing phenotypic data, meteorological and environmental data, soil condition data, and historical agricultural operation logs; The multimodal agricultural data is spatiotemporally aligned and missing values ​​are filled in to construct a spatiotemporal graph with plots and time points as nodes and temporal and spatial proximity as edges. Based on this spatiotemporal graph, complete time-series data is generated through a graph attention mechanism with causal masking. The complete time-series data is input into a multi-scale causal sensing encoder, which includes a causal convolutional submodule and a Granger causal guided long short-term memory network branch, to extract short-term dynamic features and long-term causal dependency features respectively, and then fuse them to obtain the causal hidden state. Based on the aforementioned causal hidden state, an intervention simulation under the structural causal model is executed through a dynamic counterfactual decision generator to solve an optimization problem with agricultural operation constraints, generating the optimal agricultural operation sequence and corresponding counterfactual yield prediction results within a preset number of days in the future.

2. The intelligent crop growth decision-making method based on deep learning according to claim 1, characterized in that, An edge exists between any two nodes in the spatiotemporal graph if and only if the corresponding time difference does not exceed seven days and the corresponding spatial distance does not exceed five hundred meters.

3. The intelligent crop growth decision-making method based on deep learning according to claim 1, characterized in that, The graph attention mechanism of the causal mask forces the attention weights of neighboring nodes whose timestamps are later than the current node to be reset to zero when calculating the attention weights of nodes.

4. The intelligent crop growth decision-making method based on deep learning according to claim 1, characterized in that, The causal convolution submodule uses a one-dimensional causal convolution kernel to perform convolution operations on the complete time series data. The convolution kernel only covers the data at the current time and previous time steps.

5. The intelligent crop growth decision-making method based on deep learning according to claim 1, characterized in that, The Granger causality-guided long short-term memory network branch introduces the Granger causality strength coefficient into the input gating of the long short-term memory network. The Granger causality strength coefficient is calculated in real time through a sliding window vector autoregressive model.

6. The intelligent crop growth decision-making method based on deep learning according to claim 1, characterized in that, The dynamic counterfactual decision generator eliminates the interference of confounding factors on the estimation of causal effects through a dual machine learning method, and calculates the counterfactual yield prediction results after performing specific agricultural operations based on a differentiable causal model.

7. The intelligent decision-making method for crop growth based on deep learning according to claim 1, characterized in that, The objective function of the optimization problem with agricultural operation constraints is the expected value of crop growth indicators within a preset number of days in the future minus the resource cost term. The constraints include the upper limit of a single irrigation volume, the safe interval for fertilization, and the set of feasible operation types.

8. A deep learning-based intelligent decision-making system for crop growth, characterized in that, include: The acquisition module is used to acquire remote sensing phenotypic data, meteorological and environmental data, soil condition data, and historical agricultural operation logs of the target plot over continuous time series. The completion module is used to construct a spatiotemporal graph with land parcels and time points as nodes and temporal and spatial proximity as edges, and to generate complete time series data based on the spatiotemporal graph through a graph attention mechanism with causal masking. The encoding module includes a causal convolution submodule and a Granger causal guided long short-term memory network submodule, which are used to extract short-term dynamic features and long-term causal dependency features respectively, and fuse them to obtain causal hidden states; The generation module is used to perform intervention simulations under the structural causal model based on the causal hidden state, solve optimization problems with agricultural operation constraints, and generate the optimal agricultural operation sequence and corresponding counterfactual yield prediction results within a preset number of days in the future.

9. The intelligent decision-making system for crop growth based on deep learning according to claim 8, characterized in that, In the spatiotemporal graph constructed by the completion module, there is a connecting edge between any two nodes if and only if the corresponding time difference does not exceed seven days and the corresponding spatial distance does not exceed five hundred meters.

10. The intelligent crop growth decision-making system based on deep learning according to claim 8, characterized in that, The Granger causal guided long short-term memory network submodule in the encoding module has input gating weights modulated by the Granger causal strength coefficient, which is calculated online by a sliding window vector autoregressive model.

Images