A Spatiotemporal Data-Based Agricultural Soil and Water Resources Change Matching and Prediction System

Abstract

This invention discloses an agricultural water and soil resource change matching and prediction system based on spatiotemporal data, belonging to the field of agricultural resource management technology. It includes a water and soil resource management platform, which is communicatively connected to the following modules: a spatiotemporal data acquisition and processing module, used to integrate spatiotemporal data of agricultural water and soil resources from various data sources. This invention integrates multi-source spatiotemporal data and utilizes a reinforcement learning algorithm (DQN) to model and optimize the dynamic changes of agricultural water and soil resources. It can accurately match land, crops, and water resources. Based on real-time data and prediction models, the system dynamically adjusts irrigation volume and planting structure to ensure optimal utilization of water and soil resources under different climatic conditions and resource constraints. This not only improves water resource utilization efficiency and reduces waste but also enhances land productivity, achieving sustainable agricultural development.

Description

Technical Field

[0001] This invention relates to the field of agricultural resource management technology, specifically to an agricultural water and soil resource change matching and prediction system based on spatiotemporal data. Background Technology

[0002] Global warming has led to changes in climate elements, which in turn have caused changes in the water and heat patterns. For example, water evaporation has increased, soil available moisture has decreased, and the efficiency of water and soil resource utilization has declined. These changes have challenged the traditional agricultural water and soil resource matching model. The rational allocation of agricultural water and soil resources is the key to ensuring food security. With population growth and the acceleration of urbanization, the demand for food is constantly increasing. At the same time, the finiteness of land resources and the scarcity of water resources are becoming increasingly prominent.

[0003] Agricultural water and soil resource data are characterized by spatiotemporal heterogeneity, large data volume, and complexity. Traditional machine learning methods assume stable data distribution, making it difficult to adapt to the nonlinear characteristics of resource changes caused by climate change. Therefore, how to utilize DQN to process non-stationary spatiotemporal data and continuously learn the dynamics of the environment through experience replay and stable training of the target network, so as to improve the dynamic environmental adaptability of the system, is the problem to be solved by this invention. To this end, an agricultural water and soil resource change matching and prediction system based on spatiotemporal data is proposed. Summary of the Invention

[0004] The purpose of this invention is to provide an agricultural water and soil resource change matching and prediction system based on spatiotemporal data to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] A system for matching and predicting changes in agricultural water and soil resources based on spatiotemporal data includes a water and soil resources management platform. The water and soil resources management platform is communicatively connected to the following modules:

[0007] The spatiotemporal data acquisition and processing module is used to integrate spatiotemporal data of agricultural water and soil resources from various data sources, unify spatiotemporal benchmarks, and use 3D-CNN to process spatiotemporal sequence data, automatically extract spatiotemporal patterns, and combine autoencoders to reduce high-dimensional features to low-dimensional embeddings, thereby reducing the input dimension of DQN.

[0008] The DQN reinforcement learning module is used to train the agent on the agricultural water and soil resource matching problem using a deep Q-network (DQN) to update the agent's decision-making strategy.

[0009] The dynamic matching prediction module, based on the optimal strategy output by DQN, predicts the trend of water and soil resource changes and generates matching schemes.

[0010] The visualization and interaction module is used to present the system's prediction results and decision-making strategies to users in an intuitive way, and to provide an interaction interface between users and the system, allowing users to input custom parameters and conditions.

[0011] A further improvement to the technical solution of the present invention is that the spatiotemporal data acquisition and processing module includes:

[0012] By accessing multiple data sources such as meteorological stations, soil monitoring stations, and satellite remote sensing, agricultural water and soil resource data including heterogeneous data such as meteorological, sensor, and remote sensing data are obtained. The timestamps are synchronized through dynamic time warping, and the geographic grid resolution is unified by spatial interpolation. A spatiotemporal cube is constructed to map different data sources to the same coordinate system and time granularity.

[0013] By sliding 3D-CNN convolutional kernels on a spatiotemporal cube, the spatiotemporal correlation patterns of soil moisture, precipitation, and crop growth are captured, and a spatiotemporal feature map containing high-dimensional features is output. The key features are enhanced by combining an attention mechanism, and the importance of different regions and time steps is dynamically weighted.

[0014] The high-dimensional features output by 3D-CNN are input into an autoencoder, compressed into low-dimensional embeddings by the encoder, and then normalized to map the values ​​to the DQN input range before being transmitted to the DQN reinforcement learning module.

[0015] A further improvement of the technical solution of the present invention is that the DQN reinforcement learning module includes a state-action modeling unit, an experience replay and target network unit, and a policy update unit;

[0016] The state-action modeling unit is used to define the state-action-reward space of DQN, train it on the agricultural water and soil resource matching problem, match the agricultural water and soil resource management scenario, and then transform the resource matching problem into a Markov decision process to support DQN in learning the optimal strategy.

[0017] The experience replay and target network unit are used to break data correlation through experience replay, so that the target network can be trained stably, adapt to sudden changes in data distribution under climate change, and solve the Q value oscillation caused by non-stationary data.

[0018] The policy update unit is used to update the agent's decision-making policy based on the results of DQN training, and then apply the learned optimal policy to the actual agricultural water and soil resource matching decision.

[0019] A further improvement to the technical solution of the present invention is that the state-action modeling unit includes:

[0020] Define the state variables in the system, quantify the state variables into state vectors that can be input into DQN, construct the state space, and provide DQN with comprehensive system state information so that it can accurately perceive the environment.

[0021] Based on the needs of agricultural water and soil resource management, the feasible operations are determined, and each operation is discretized into a finite set of actions to construct an action space for DQN to select.

[0022] Design a reward mechanism, define a reward function, and give positive or negative rewards based on the resource matching effect to balance short-term gains and long-term benefits, guide DQN to learn the optimal strategy, and enable it to pursue the optimal resource matching effect in a dynamic environment.

[0023] The agricultural water and soil resource matching problem is transformed into a Markov Decision Process (MDP). The state transition probabilities are defined so that the system state transition depends only on the current state and action, satisfying the Markov property. Then, using the predefined state, action, and reward space, the DQN is trained in conjunction with the Markov Decision Process (MDP) to optimize the agent's decision-making strategy. The strategy is dynamically adjusted according to environmental feedback to achieve optimal resource matching.

[0024] A further improvement to the technical solution of the present invention is that the experience playback and target network unit includes:

[0025] An experience replay mechanism is adopted to store historical interaction data containing historical state-action-reward-new state tuples into a replay buffer to form an experience replay pool. The capacity of the replay pool is managed by a FIFO strategy. In each training iteration, a small batch of data is randomly sampled from the experience replay pool to eliminate temporal correlation, avoid policy deviation caused by continuous similar samples in DQN, and improve sample utilization efficiency.

[0026] The design employs a dual-structure architecture consisting of a main network and a target network. It maintains a dual structure of a main network (updated online) and a target network (updated late). Every C steps, the target network parameters are softly copied from the main network. The target network Q-value is then used to calculate the TD target. By delaying the update of the target network parameters, Q-value oscillations are suppressed, training fluctuations are reduced, and the data distribution changes under climate change are adapted to.

[0027] A dual-network collaborative training method is adopted. The main network calculates the loss based on the empirical replay samples and updates the parameters through backpropagation. The target network provides a stable TD target. The parameters of the main network are synchronized every C steps, so that the policy is gradually optimized and the problem of overestimation or underestimation of Q value caused by non-stationary data is avoided.

[0028] A further improvement to the technical solution of the present invention is that the strategy update unit includes:

[0029] Based on the DQN training results, the main network obtains the action value in the current state based on the current state input, and outputs the Q-value vector of each action. The argmax is used to select the action with the largest Q value as the candidate policy, while retaining the ε-greedy exploration mechanism to balance utilization and exploration.

[0030] The relevant actions of the candidate strategy are input into the decision base of the agent to check whether they violate hard constraints. If they do, they are replaced with suboptimal legal actions, and a corrected action probability distribution is generated. The strategy is gradually optimized to ensure that the strategy meets the requirements of sustainable management and is more adaptable to the dynamic environment.

[0031] The optimized strategy, which includes legal actions, is applied to actual agricultural water and soil resource matching decisions. Environmental response data is recorded simultaneously to optimize resource allocation. Actual rewards and new states are fed back to the experience replay pool, forming a continuous iterative cycle of training, deployment, and optimization.

[0032] A further improvement of the technical solution of the present invention is that the dynamic matching prediction module includes a resource change prediction unit and a multi-objective matching optimization unit;

[0033] The resource change prediction unit combines LSTM-DQN to process time series predictions and output management actions, while using graph neural networks (GNN) to model spatial dependencies and predict future spatiotemporal resource status.

[0034] The multi-objective matching optimization unit generates an initial matching scheme based on the DQN strategy, performs local search optimization by combining it with the genetic algorithm (GA), introduces a constraint satisfaction mechanism to ensure the feasibility of the scheme, and then optimizes the land-crop-water resource matching scheme under resource constraints.

[0035] A further improvement to the technical solution of the present invention is that the resource change prediction unit includes:

[0036] Historical resource data is collected to construct a structured spatiotemporal dataset, which includes time-series data, spatial data, and management data. The dataset is then divided into spatiotemporal blocks, and the influence domain is divided based on Thiessen polygons. Each site represents a spatial unit. An adjacency matrix is ​​constructed, and then the temporal features of the sites are extracted using an LSTM network. A GNN is used to aggregate spatial dependencies through the adjacency matrix to generate a spatiotemporal joint embedding vector.

[0037] Based on the spatiotemporal joint embedding vector, the LSTM-DQN decoder is used to predict the resource state of the next T steps time by time, and the spatial distribution is corrected synchronously with GNN to output a spatiotemporal prediction field with uncertainty.

[0038] The predicted states contained in the spatiotemporal prediction field are input into the DQN main network to calculate the Q value of each management action (irrigation amount / crop switching). Combined with ε-greedy selection of the optimal action, a comprehensive prediction of the future spatiotemporal resource state is generated, and constraint verification is used to ensure compliance with water resource limits and ecological requirements.

[0039] A further improvement to the technical solution of the present invention is that the multi-objective matching optimization unit includes:

[0040] Constraints including land type, crop water requirement, and water availability are encoded as state vectors and input into the DQN network. Matching actions are defined as triplets of land plot-crop-irrigation amount. The DQN outputs the Q value of each action, which is the expected return of each triplet of land plot-crop-irrigation amount. At the same time, an ε-greedy strategy is used to select actions and generate initial matching schemes for each land plot to ensure coverage of all land and crop needs.

[0041] Using the initial scheme generated by DQN as a seed, a genetic algorithm (GA) is used for local search optimization. A fitness function is constructed, and three objectives are considered: economic benefit (profit), ecological benefit (water saving rate), and resource utilization rate. The three objectives are weighted and scored to evaluate the overall performance of each scheme. Genetic operations including selection (roulette wheel), crossover (single-point exchange of matching pairs), and mutation (random adjustment of irrigation amount) are then executed. The genetic operations are executed iteratively to gradually improve the quality of the scheme and explore better solutions.

[0042] The output schemes are subjected to hard constraint verification, checking water resource limits (total irrigation volume ≤ available water volume) and land suitability (e.g., rice cultivation is prohibited in saline-alkali land). For schemes that exceed the limits, the irrigation volume of high water-consuming crops is reduced or replaced with low-demand crops according to priority. Then, using the Pareto output, the non-dominated solution set is screened out, providing decision-makers with a library of feasible schemes for multi-objective trade-offs, so that the schemes meet resource constraints and ecological requirements.

[0043] The system outputs an optimized and validated database of agricultural water and soil resource matching solutions. Each solution in the database contains detailed information on land plots, crops, and irrigation amounts, as well as corresponding economic, ecological, and resource utilization scores, providing decision-makers with feasible solutions that balance multiple objectives.

[0044] A further improvement to the technical solution of the present invention is that the visual interaction module includes:

[0045] The agricultural water and soil matching scheme output by the system is presented in the form of dynamic charts and geographic heat maps, including land block level views, time axis interaction and target comparison panel;

[0046] It provides a drag-and-drop control panel and condition input forms, allowing users to dynamically adjust constraints, including resource constraints, target weights, and crop preferences;

[0047] The system responds instantly to user actions, generates interactive reports, and issues constraint conflict warnings. If user input results in no feasible solution, conflict parameters are highlighted, and Pareto solution navigation is performed to locate the user-defined solution in the 3D target space, recommending nearby high-quality solutions. A one-click export function is also available, supporting the export of solution libraries as Excel / PDF, including matching tables, scores, and visualization screenshots to assist in decision-making and reporting.

[0048] Due to the adoption of the above technical solution, the technical progress achieved by this invention compared to the prior art is as follows:

[0049] 1. This invention provides an agricultural water and soil resource change matching and prediction system based on spatiotemporal data. By integrating multi-source spatiotemporal data and using reinforcement learning algorithm (DQN) to model and optimize the dynamic changes of agricultural water and soil resources, it can accurately match land, crops and water resources. Based on real-time data and prediction models, the system dynamically adjusts irrigation volume and planting structure to ensure that water and soil resources are optimally utilized under different climatic conditions and resource constraints. This not only improves the efficiency of water resource utilization and reduces waste, but also enhances land productivity and achieves sustainable development of agricultural production.

[0050] 2. This invention provides an agricultural water and soil resource change matching and prediction system based on spatiotemporal data. Through experience playback and target network mechanisms, it stabilizes the training process and adapts to dynamic environmental changes. When faced with adverse conditions such as extreme weather, water shortages, or land degradation, it can quickly adjust strategies, optimize resource allocation, ensure the stable operation of the agricultural production system, further enhance the resilience of agricultural production, and reduce the risks caused by environmental changes. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0052] Figure 1 This is a schematic diagram of the workflow of the present invention;

[0053] Figure 2 This is a schematic diagram of the functional modules of the system of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0055] Example 1, as Figure 1 , Figure 2 As shown, this invention provides an agricultural water and soil resource change matching and prediction system based on spatiotemporal data, including a water and soil resource management platform. The water and soil resource management platform is communicatively connected to the following modules, wherein:

[0056] The spatiotemporal data acquisition and processing module integrates spatiotemporal data of agricultural water and soil resources from various data sources, unifies the spatiotemporal benchmark, and uses 3D-CNN to process spatiotemporal sequence data. It automatically extracts spatiotemporal patterns, combines an autoencoder to reduce high-dimensional features to low-dimensional embedding, and reduces the input dimension of DQN. It accesses multiple data sources such as meteorological stations, soil monitoring stations, and satellite remote sensing to acquire agricultural water and soil resource data including heterogeneous data from meteorology, sensors, and remote sensing. It synchronizes timestamps through dynamic time warping, uses spatial interpolation to unify geographic grid resolution, constructs a spatiotemporal cube, maps different data sources to the same coordinate system and time granularity, and uses 3D-CNN convolutional kernels to slide on the spatiotemporal cube to capture the spatiotemporal correlation patterns of soil moisture, precipitation, and crop growth, outputting a spatiotemporal feature map containing high-dimensional features. It combines an attention mechanism to strengthen key features, dynamically weights the importance of different regions and time steps, inputs the high-dimensional features output by 3D-CNN into the autoencoder, compresses them to low-dimensional embedding, and combines normalization to map the values ​​to the DQN input range, and then transmits them to the DQN reinforcement learning module.

[0057] The specific tasks of the spatiotemporal data acquisition and processing module are as follows:

[0058] This system integrates meteorological stations, soil monitoring stations, and satellite remote sensing data to acquire meteorological, sensor, and remote sensing data. Meteorological data, including temperature, precipitation, wind speed, and humidity, is obtained from meteorological stations. Soil moisture, soil nutrients, and soil temperature data are obtained from soil monitoring stations. Land use type, vegetation index, and surface temperature data are obtained from satellite remote sensing. Dynamic time warping is performed on the timestamps of different data sources to address inconsistent data sampling frequencies. The DTW algorithm is used to align time-series data, ensuring consistency across different data sources in the temporal dimension. Spatial interpolation (Kriging interpolation) is employed for spatially unevenly distributed data to unify all data to the same geographic grid resolution, constructing a unified geographic grid system. This maps different data sources to the same coordinate system and temporal granularity. The time-synchronized and spatially aligned data are then integrated into a three-dimensional structure, forming a spatiotemporal cube. The three dimensions of the spatiotemporal cube are... The algorithm combines time, space (longitude and latitude), and features. It utilizes 3D-CNN convolutional kernels sliding across a spatiotemporal cube to capture spatiotemporal correlation patterns among multiple variables such as soil moisture, precipitation, and crop growth. The output is a spatiotemporal feature map containing high-dimensional features. Key patterns and structures in the spatiotemporal data are extracted, and an attention mechanism is used to strengthen key features, dynamically weighting the importance of different regions and time steps. The attention mechanism identifies spatiotemporal features valuable for agricultural water and soil resource matching and prediction. The high-dimensional spatiotemporal features output from the 3D-CNN are input into an autoencoder. The encoder part of the autoencoder compresses the high-dimensional features into low-dimensional embeddings, reducing feature dimensionality. The low-dimensional embeddings are then normalized, mapping the values ​​to the DQN input range to ensure a uniform numerical range for the features input to the DQN module. This results in the output of low-dimensional, normalized spatiotemporal feature embeddings, which are then transmitted as input to the DQN reinforcement learning module.

[0059] The DQN reinforcement learning module is used to train the agent on the agricultural water and soil resource matching problem using a deep Q-network (DQN) to update the agent's decision-making strategy and achieve dynamic environmental adaptation and strategy optimization. The DQN reinforcement learning module includes a state-action modeling unit, an experience playback and target network unit, and a policy update unit.

[0060] The state-action modeling unit (SAM) defines the state-action-reward space of the DQN, trains it on the agricultural water and soil resource matching problem, matches it with agricultural water and soil resource management scenarios, and transforms the resource matching problem into a Markov decision process. This supports the DQN in learning the optimal strategy, clarifies the state variables in the system, quantifies the state variables into state vectors that can be input into the DQN, constructs the state space, and provides the DQN with comprehensive system state information, enabling it to accurately perceive the environment. Based on the needs of agricultural water and soil resource management, it determines executable operations and discretizes each operation into a finite set of actions, constructing an action space for the DQN to choose from. The reward mechanism defines a reward function that provides positive or negative rewards based on the resource matching effect, balancing short-term gains and long-term benefits. This guides the DQN to learn the optimal strategy, enabling it to pursue the optimal resource matching effect in a dynamic environment. The agricultural water and soil resource matching problem is transformed into a Markov Decision Process (MDP), with clearly defined state transition probabilities so that the system state transition depends only on the current state and action, satisfying the Markov property. Then, using the defined state, action, and reward space, combined with the Markov Decision Process (MDP), the DQN is trained to optimize the agent's decision-making strategy. The strategy is dynamically adjusted based on environmental feedback to achieve optimal resource matching.

[0061] The specific tasks of the state-action modeling unit are as follows:

[0062] The system's state variables are clearly defined, including soil moisture, crop growth stage, water resource inventory, meteorological conditions, soil nutrients, and vegetation index. Soil moisture reflects the water content in the soil. Crop growth stages include sowing, seedling, flowering, and maturity, each with different water and soil resource requirements. Water resource inventory includes groundwater reserves, reservoir water volume, and available irrigation water. Meteorological conditions include temperature, precipitation, wind speed, and humidity, affecting crop growth and water resource utilization efficiency. Soil nutrients include the content of major nutrients such as nitrogen, phosphorus, and potassium, affecting crop growth and yield. The vegetation index reflects the crop's growth status and can be used to assess crop health. These state variables are then quantified into inputs to the DQN. A state vector is generated, and the quantified state variables are combined into a single state vector to form the system's state space. Executable operations are determined, including irrigation adjustment, planting structure adjustment, fertilization, and drainage. Irrigation adjustment involves increasing or decreasing irrigation water volume; planting structure adjustment involves changing the type of crop or planting area; fertilization involves adjusting the amount or timing of fertilizer application; and drainage involves controlling drainage to prevent waterlogging. These operations are then discretized into a finite set of actions: irrigation adjustment is defined as increasing irrigation volume, maintaining the current irrigation volume, or decreasing irrigation volume; planting structure adjustment is defined as planting crop A, planting crop B, planting crop C, etc.; and fertilization is defined as increasing fertilizer amount, maintaining the current fertilizer amount, or decreasing fertilizer amount. Drainage operations are defined as increasing drainage volume, maintaining the current drainage volume, and decreasing drainage volume, and the discretized set of actions is defined as the action space. A reward function is defined, which provides positive or negative rewards based on the resource matching effect. The reward signal can guide DQN to learn the optimal strategy. The reward function includes the following factors: yield, water resource utilization efficiency, soil health, ecological benefits, and economic costs. Higher crop yield results in a higher reward. Efficient water resource utilization, which involves the rational use of water resources and avoidance of waste, provides a positive reward. Soil health, which maintains soil fertility and structure, provides a positive reward. Ecological benefits, which reduce negative environmental impacts, provide a positive reward. Economic costs, which reduce the costs of irrigation, fertilization, and other operations, provide a positive reward. Taking into account both short-term gains and long-term benefits, DQN adjusts weight coefficients to ensure that it pursues both short-term gains and long-term benefits. It defines state transition probabilities so that system state transitions depend only on the current state and action, satisfying the Markov property. These state transition probabilities are obtained through historical data statistics. A complete Markov decision process is constructed, and then, using a predefined state space, action space, and reward function, combined with a pre-constructed MDP model, the agent's decision-making strategy is optimized through continuous trial and error and learning. The DQN algorithm's training mechanism stabilizes the training process, adapts to dynamic environments, and dynamically adjusts the strategy based on environmental feedback, enabling DQN to select the optimal action based on the current state and achieve optimal resource matching.

[0063] Experience replay and target network units are used to break data correlations through experience replay, enabling stable training of the target network, adapting to abrupt changes in data distribution under climate change, and solving Q-value oscillations caused by non-stationary data. Specifically, experience replay stores historical state-action-reward-new state tuples, randomly samples data for training, and uses the target network's delayed update of main network parameters to reduce training fluctuations. The experience replay mechanism stores historical interaction data containing historical state-action-reward-new state tuples in a replay buffer, forming an experience replay pool. A FIFO strategy is used to manage the replay pool's capacity. Each training iteration randomly samples a small batch of data from the experience replay pool to eliminate temporal correlations and avoid the continuous oscillations caused by DQN. Similar samples can lead to policy bias. To improve sample utilization efficiency, a dual-structure architecture of a main network and a target network is designed. The dual structure of the main network (updated online) and the target network (updated late) is maintained. Every C steps, the parameters of the target network are softly copied from the main network, and then the Q value of the target network is used to calculate the TD target. By delaying the update of the target network parameters, the Q value oscillation is suppressed, training fluctuations are reduced, and the data distribution changes under climate change are adapted to. The dual-network collaborative training is carried out. The main network calculates the loss based on the empirical replay of samples and updates the parameters through backpropagation. The target network provides a stable TD target. Every C steps, the parameters of the main network are synchronized, so that the policy is gradually optimized and the problem of Q value overestimation or underestimation caused by non-stationary data is avoided.

[0064] The specific tasks of experience playback and target network unit are as follows:

[0065] Create a playback buffer with a capacity of N = 10. 5The system stores historical interaction data of historical state-action-reward-new state tuples in the replay buffer, forming an experience replay pool. When the experience replay pool reaches its capacity limit, a first-in-first-out (FIFO) strategy is adopted to remove the oldest stored data to make room for new interaction data. This ensures that the data in the experience replay pool reflects the dynamic changes and historical interaction information of the system. In each training iteration, a small batch of data, such as 64 items, is randomly sampled from the experience replay pool. The purpose of random sampling is to eliminate the temporal correlation of data and avoid policy bias caused by continuous similar samples in DQN. Through random sampling, it is ensured that each sample is used multiple times during training, avoiding overfitting problems caused by data order and enhancing the generalization ability of the model. The design employs a dual-structure architecture consisting of a main network and a target network. Two networks are maintained: the main network is updated online, while the target network is updated with a delay. The main network calculates the loss based on empirical replay samples and updates parameters via backpropagation. The target network provides a stable Temporal Objective (TD). Its parameters are soft-copied from the main network to the target network every C steps, proportionally. This delayed update of the target network parameters suppresses Q-value oscillations, reduces training fluctuations, adapts to abrupt data distribution changes due to climate change, and avoids overestimation or underestimation of Q-values ​​caused by non-stationary data. The two networks are trained collaboratively. The main network calculates the loss based on mini-batch data sampled from the empirical replay pool, and the target network's Q-values ​​are used to calculate the Temporal Objective. The Difference objective is used, where the loss function is the mean squared error (MSE), and the parameters of the main network are updated through backpropagation. This provides a stable TD objective through the target network, avoiding Q-value instability caused by frequent updates of the main network parameters. The main network parameters are synchronized every C steps to ensure that the target network can gradually follow the updates of the main network while maintaining a certain degree of stability. Through dual-network collaborative training, the gradual optimization of the strategy is ensured, avoiding Q-value overestimation or underestimation caused by non-stationary data, and improving the system's dynamic environment adaptability.

[0066] The formula for soft copying is as follows:

[0067] θ target ←τθ online +(1-τ)θ target ;

[0068] In the formula, θ target Let θ be the parameter of the target network. online The parameters of the main network are τ, which is the scaling factor for soft replication. It takes a small value, such as 0.001, which means that the parameters of the target network will move closer to the parameters of the main network by a small step with each update.

[0069] The formula for the loss function is as follows:

[0070]

[0071] In the formula, L is the loss function, representing the mean square error between the predicted Q value and the target Q value. Indicates the replay of experiences from the pool Find the expectation Q of all sampled tuples (s, a, r, s`). online (s,a) is the Q-value predicted by the main network, representing the expected reward of performing action a in the current state s, r is the immediate reward obtained after performing action a; γ is the discount factor, representing the degree of decay of future rewards, with a value between 0 and 1. The maximum Q-value predicted by the target network is represented by the value in the new state s. ` The maximum expected return to be obtained from all possible actions;

[0072] The policy update unit is used to update the agent's decision policy based on the DQN training results, and then apply the learned optimal policy to the actual agricultural water and soil resource matching decision. Based on the DQN training results, the main network obtains the action value of the current state input to output the Q-value vector of each action. The argmax is used to select the action with the largest Q value as the candidate policy, while retaining the ε-greedy exploration mechanism (ε decreases to 0.05 as training progresses) to balance utilization and exploration. The relevant actions of the candidate policy are input into the agent's decision base to check whether they violate hard constraints. If they violate the constraints, they are replaced with the second-best legal action, and a corrected action probability distribution is generated to gradually optimize the policy, ensuring that the policy meets the requirements of sustainable management and is more adaptable to the dynamic environment. The optimized policy containing legal actions is applied to the actual agricultural water and soil resource matching decision, and environmental response data is recorded simultaneously to perform optimal resource allocation. The actual reward and the new state are fed back to the experience replay pool to form a continuous iterative cycle of training-deployment-optimization.

[0073] The specific tasks of the policy update unit are as follows:

[0074] The main network receives the current state as input and outputs an action value vector for that state. Each element in the action value vector represents the expected reward of performing an action in the current state. The argmax function selects the action with the largest Q-value from the action value vector as a candidate policy. An ε-greedy exploration mechanism balances utilization and exploration. An action is randomly selected for exploration with probability ε, and the action with the largest Q-value is selected for utilization with probability 1-ε. ε decreases to 0.05 during training to reduce the frequency of random exploration and increase policy stability. The relevant actions of the candidate policies are input into the agent's decision base to verify whether hard constraints are violated, such as groundwater extraction limits and crop rotation cycles. Hard constraints are rules that the system must follow to ensure resource management. To ensure the sustainability of the strategy, if a candidate strategy violates hard constraints, a suboptimal action is selected from the legal actions as an alternative, generating a corrected action probability distribution to ensure that the strategy meets the requirements of sustainable management. The strategy is then gradually optimized to get as close as possible to the optimal strategy while satisfying hard constraints. The optimized optimal strategy, which includes legal actions, is applied to actual agricultural water and soil resource matching decisions to guide operations such as irrigation and planting structure adjustment. Environmental response data, including rewards and new states after the actions are executed, are recorded simultaneously. The actual rewards and new states are fed back to the experience replay pool, forming a continuous iterative cycle of training-deployment-optimization. By continuously collecting new interaction data and updating the experience replay pool, the DQN model is further optimized to ensure that the strategy continuously adapts and optimizes in the dynamic environment.

[0075] The formula for using the argmax function to select the action with the largest Q-value from the action value vector as a candidate policy is as follows:

[0076]

[0077] In the formula, a * This represents the action with the highest Q-value in the current state s, i.e., the candidate policy. Indicates from action space Choose the action that maximizes the function Q(s,q), where Q(s,q) represents the expected reward (action value) of performing action q in state s, s represents the current state (a state vector), and q represents the action space. One of the actions, The action space represents all possible actions. Through the argmax function, the DQN algorithm can select the action with the highest Q value in the current state as a candidate policy. This action is considered optimal in the current state and is expected to bring the greatest expected reward. To balance utilization and exploration, the DQN algorithm combines an ε-greedy exploration mechanism, which selects non-optimal actions with a certain probability to explore the environment and avoid premature convergence to a local optimum.

[0078] The dynamic matching prediction module, based on the optimal strategy output by DQN, predicts the trend of water and soil resource changes and generates matching schemes.

[0079] The visualization and interaction module is used to present the system's prediction results and decision-making strategies to users in an intuitive way, and to provide an interaction interface between users and the system, allowing users to input custom parameters and conditions.

[0080] Example 2, as Figure 1 , Figure 2 As shown, based on Embodiment 1, the present invention provides a technical solution: preferably, the dynamic matching prediction module includes a resource change prediction unit and a multi-objective matching optimization unit;

[0081] The resource change prediction unit combines LSTM-DQN to process time series forecasts and output management actions. Simultaneously, it uses a graph neural network (GNN) to model spatial dependencies and predict future spatiotemporal resource states. Historical resource data is collected to construct a structured spatiotemporal dataset containing time series, spatial, and management data, which is then divided into spatiotemporal blocks. Influence domains are defined based on Thiessen polygons, with each site representing a spatial unit. An adjacency matrix is ​​constructed, and LSTM networks are used to extract site temporal features. GNNs are then used to aggregate spatial dependencies through the adjacency matrix, generating a spatiotemporal joint embedding vector. Based on this vector, an LSTM-DQN decoder predicts the resource state for the next T steps, simultaneously correcting the spatial distribution with GNNs. This outputs a spatiotemporal prediction field with uncertainty. The predicted states from this field are input into the DQN main network to calculate the Q-values ​​of each management action (irrigation / crop switching). An ε-greedy selection process is used to generate a comprehensive prediction of future spatiotemporal resource states, and constraint verification ensures compliance with water resource limits and ecological requirements.

[0082] The specific tasks of the resource change prediction unit are as follows:

[0083] Historical resource data was collected to construct a structured spatiotemporal dataset, including time-series data, spatial data, and management data. The time-series data includes data on changes in soil moisture, precipitation, and temperature over time; the spatial data includes spatial attributes such as geographical location, topography, and soil type; and the management data includes irrigation records and planting structure adjustment records. The data was divided into spatiotemporal blocks, with each site representing a spatial unit. The influence domain of each site was delineated using Thiessen polygons to ensure a clear neighborhood for each site. An adjacency matrix A was then constructed to represent the spatial relationships between sites. If the distance between site i and site j is less than 3 kilometers, then A... ij =1, otherwise A ij=0. The LSTM network is used to extract the temporal features of each station. LSTM captures long-term dependencies in time-series data and is suitable for processing time-series data such as soil moisture and precipitation. A Graph Neural Network (GNN) is used to aggregate spatial dependencies between stations through an adjacency matrix. The GNN utilizes the spatial information in the adjacency matrix to incorporate the mutual influence between stations into the model. Finally, the temporal features extracted by LSTM and the spatial dependency features aggregated by GNN are combined to generate a spatiotemporal joint embedding vector. This embedding vector contains both temporal and spatial information, comprehensively reflecting the spatiotemporal characteristics of the stations. An LSTM-DQN decoder is used to predict the resource state in the next T steps time-by-time. By combining the time-series processing capabilities of LSTM and the decision-making capabilities of DQN, the LSTM-DQN decoder generates predictions of future resource states, which are then simultaneously modified by the GNN. A positive spatial distribution ensures the consistency and rationality of the prediction results in space. Monte Carlo methods are then introduced to generate M sets of random feature samples, outputting a spatiotemporal prediction field with uncertainty. The predicted states contained in the spatiotemporal prediction field are input into the DQN main network to calculate the Q-values ​​of various management actions, such as irrigation adjustment and crop switching. The Q-value represents the expected return of performing a certain management action under the current prediction state. An ε-greedy strategy is used to select the optimal action. Based on the selected optimal action, a comprehensive prediction of the future spatiotemporal resource state is generated. The comprehensive prediction result includes the predicted resource state and suggested management actions. Simultaneously, the generated comprehensive prediction is constrained and validated to ensure compliance with water resource limits and ecological requirements. If the prediction result violates hard constraints, such as groundwater extraction limits or crop rotation cycles, the prediction result is adjusted, and a second-best legal action is selected. The prediction strategy is gradually optimized to ensure compliance with sustainable management requirements.

[0084] The multi-objective matching optimization unit generates initial matching schemes based on the DQN strategy, combines them with a genetic algorithm (GA) for local search optimization, and introduces a constraint satisfaction mechanism to ensure the feasibility of the schemes. Then, under resource constraints, it optimizes the land-crop-water resource matching scheme. Constraints including land type, crop water requirement, and water availability are encoded as state vectors and input into the DQN network. Matching actions are defined as triplets of land plot-crop-irrigation amount. The DQN outputs the Q-value of each action, i.e., the expected return of each land plot-crop-irrigation amount triplet. Simultaneously, an ε-greedy strategy is used to select actions, generating initial matching schemes plot by plot to ensure coverage of all land and crop needs. Using the initial schemes generated by DQN as seeds, the genetic algorithm (GA) is used for local search optimization, and a fitness function is constructed. The system comprehensively considers three objectives: economic benefit (revenue), ecological benefit (water saving rate), and resource utilization rate. The algorithm performs weighted scoring to evaluate the overall performance of each solution. It then executes genetic operations including selection (roulette wheel), crossover (single-point swapping of matching pairs), and mutation (random adjustment of irrigation amounts). These genetic operations are iteratively performed to improve the quality of the solutions and explore better solutions. Hard constraints are applied to the output solutions to check water resource limits (total irrigation amount ≤ available water amount) and land suitability (e.g., rice cultivation is prohibited in saline-alkali land). For solutions exceeding these limits, irrigation amounts for high water-consuming crops are reduced or replaced with low-water-consuming crops based on priority. Pareto output is then used to filter out non-dominated solutions, providing decision-makers with a feasible solution library for multi-objective trade-offs. This ensures that the solutions meet resource constraints and ecological requirements. The resulting optimized and validated agricultural water and soil resource matching solution library contains detailed information on land plots, crops, and irrigation amounts, along with corresponding economic, ecological, and resource utilization scores, providing decision-makers with feasible solutions for multi-objective trade-offs.

[0085] The specific tasks of the multi-objective matching optimization unit are as follows:

[0086] Constraints such as land type, crop water requirement, and water availability are encoded as state vectors and input into a DQN network. Matching actions are defined as triplets of land plot-crop-irrigation amount, where each triplet represents planting a specific crop on a particular land plot and allocating a certain amount of irrigation water. The DQN outputs the Q-value of each action, i.e., the expected return of each land plot-crop-irrigation amount triplet. An ε-greedy strategy is used to select actions, generating initial matching schemes plot by plot to ensure coverage of all land and crop needs. Using the initial schemes generated by the DQN as seeds, the initial matching schemes are used as the initial population for a genetic algorithm. A fitness function is defined, considering three objectives: economic benefit (return), ecological benefit (water saving rate), and resource utilization rate. Genetic operations are then performed. For the selection operation (roulette wheel), schemes with higher fitness are probabilistically selected for the next generation based on the fitness function. For the crossover operation (single-point exchange of matching pairs), two schemes are randomly selected, and some matching pairs are exchanged to generate new ones. For the mutation operation (randomly adjusting irrigation volume), a solution is randomly selected, and the irrigation volume of a certain land plot is adjusted. The selection, crossover, and mutation operations are repeated to gradually improve the quality of the solutions and explore better solutions. The generated solutions are checked to see if they meet hard constraints, such as the total irrigation volume not exceeding the available water volume and land suitability. For solutions that exceed the limits, the irrigation volume of high water-consuming crops is reduced or replaced with low-demand crops according to priority. The priority is determined based on factors such as the economic value and ecological importance of the crops. Using the Pareto optimization method, a non-dominated solution set is selected to provide decision-makers with a feasible solution library for multi-objective trade-offs. Each solution in the non-dominated solution set achieves a balance between economic benefits, ecological benefits, and resource utilization, and cannot optimize any objective without worsening other objectives. The optimized and verified agricultural water and soil resource matching solution library is then output. Each solution in the solution library contains detailed information on land plot, crop, and irrigation volume, as well as corresponding economic benefit, ecological benefit, and resource utilization scores.

[0087] The visual interaction module includes:

[0088] The system outputs agricultural water and soil matching schemes in the form of dynamic charts and geographic heat maps, including a land parcel-level view, an interactive timeline, and a target comparison panel. In the land parcel-level view, crop distribution is marked with different colors, and bubble size represents irrigation volume. The interactive timeline allows users to observe quarterly / annual scheme trends by sliding a time bar. The target comparison panel simultaneously displays radar charts of economic benefits, water saving rate, and resource utilization rate, supporting multi-scheme overlay comparisons. It provides a drag-and-drop control panel and conditional input forms, allowing users to dynamically adjust constraints, including resource constraints, target weights, and crop preferences. The resource constraints can be modified via a slider. Water volume limits and land suitability rules are implemented. For target weights, the priority of economic / ecological / resource utilization is redistributed through star rating. For crop preferences, specific crops can be disabled or forced to be planted, triggering the system to re-optimize the solution. The system responds instantly to user operations, generates interactive reports, and provides constraint conflict warnings. If user input results in no feasible solution, conflict parameters are highlighted, and Pareto solution navigation is performed to locate the user-defined solution in the 3D target space, recommend nearby high-quality solutions, and set up a one-click export function to support exporting the solution library as Excel / PDF, including matching tables, scores, and visualization screenshots to assist in decision-making and reporting.

[0089] The specific tasks of the visual interaction module are as follows:

[0090] The system uses a geographic heatmap to display matching schemes at the plot level. Different colors mark the distribution of different crops, and bubble size represents irrigation volume, providing a visual representation of crop planting and irrigation status for each plot. A timeline slider allows users to observe quarterly or annual trends, dynamically updating charts and heatmaps to display matching schemes at different points in time. Simultaneously, radar charts showing economic benefits, water saving rates, and resource utilization rates are displayed, supporting multi-scheme overlay comparisons for intuitive comparison of their advantages and disadvantages. A drag-and-drop control panel allows users to adjust constraints, including a condition input form that allows dynamic adjustment of constraints such as resource constraints, target weights, and crop preferences. For resource constraints, a slider modifies the available water limit, adjusting land suitability rules, such as prohibiting certain crops or restricting the use of specific land types. For target weights, a star rating system redistributes the priorities of economic, ecological, and resource utilization, allowing users to adjust the weights of each target according to their needs. Checkboxes are provided for crop preferences. The system allows users to disable specific crops or force the planting of certain crops, triggering the system to re-optimize the solution to meet the user's new constraints. After the user's action, the system responds instantly, re-optimizing the solution and updating the visualization results, dynamically generating interactive reports that display the latest matching solutions and scores. If the user-input constraints result in an infeasible solution, the system highlights conflict parameters and provides detailed conflict hints to help the user understand the problem. It locates the user-defined solution in the 3D target space, recommends nearby high-quality solutions, and uses Pareto fronts to help users weigh multiple objectives. A one-click export function is provided, supporting the export of the solution library to Excel or PDF format. The exported content includes: matching tables, scores, and visualization screenshots. The matching tables record the crop planting and irrigation amounts for each land parcel in detail; the scores evaluate the economic benefits, water-saving rate, and resource utilization rate of each solution; and the visualization screenshots include land parcel-level views, timeline interaction, and target comparison panels. The exported files can be used as supplementary material for decision-making reports, helping users present and explain optimization results.

[0091] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A system for matching and predicting changes in agricultural water and soil resources based on spatiotemporal data, comprising a water and soil resources management platform, characterized in that, The water and soil resources management platform has the following communication connection modules, among which: The spatiotemporal data acquisition and processing module is used to integrate spatiotemporal data of agricultural water and soil resources from various data sources, and to process spatiotemporal sequence data using 3D-CNN, automatically extract spatiotemporal patterns, and combine autoencoders to reduce high-dimensional features to low-dimensional embedding. The DQN reinforcement learning module is used to train the agent on the agricultural water and soil resource matching problem to update the agent's decision-making strategy. The dynamic matching prediction module predicts the trend of water and soil resource changes and generates matching schemes based on the optimal strategy output by DQN. The visualization and interaction module is used to display the system's prediction results and decision-making strategies to users, and to provide an interaction interface between users and the system, allowing users to input custom parameters and conditions; The spatiotemporal data acquisition and processing module includes: By accessing multiple data sources, including meteorological stations, soil monitoring stations, and satellite remote sensing, agricultural water and soil resource data, including heterogeneous data from meteorology, sensors, and remote sensing, and by synchronizing timestamps through dynamic time warping and unifying geographic grid resolution through spatial interpolation, a spatiotemporal cube is constructed to map different data sources to the same coordinate system and time granularity. By sliding 3D-CNN convolutional kernels on a spatiotemporal cube, the spatiotemporal correlation patterns of soil moisture, precipitation, and crop growth are captured, and a spatiotemporal feature map containing high-dimensional features is output. The key features are enhanced by combining an attention mechanism, and the importance of different regions and time steps is dynamically weighted. The high-dimensional features output by 3D-CNN are input into an autoencoder, compressed into low-dimensional embeddings by the encoder, and normalized to map the values ​​to the DQN input range, and then transmitted to the DQN reinforcement learning module. The DQN reinforcement learning module includes a state-action modeling unit, an experience replay and target network unit, and a policy update unit. The state-action modeling unit is used to define the state-action-reward space of DQN, train it on the agricultural water and soil resource matching problem, match the agricultural water and soil resource management scenario, and then transform the resource matching problem into a Markov decision process. The experience replay and target network unit are used to break data correlation through experience replay, so that the target network can be trained stably and adapt to sudden changes in data distribution under climate change. The policy update unit is used to update the agent's decision-making policy based on the results of DQN training, and then apply the learned optimal policy to the actual agricultural water and soil resource matching decision-making. The dynamic matching prediction module includes a resource change prediction unit and a multi-objective matching optimization unit; The resource change prediction unit combines LSTM-DQN to process time series predictions and output management actions, while using graph neural networks to model spatial dependencies and predict future spatiotemporal resource status. The multi-objective matching optimization unit generates an initial matching scheme based on the DQN strategy, performs local search optimization by combining a genetic algorithm, introduces a constraint satisfaction mechanism to ensure the feasibility of the scheme, and then optimizes the land-crop-water resource matching scheme under resource constraints. The resource change prediction unit includes: Historical resource data is collected to construct a structured spatiotemporal dataset, which includes time-series data, spatial data, and management data. The dataset is then divided into spatiotemporal blocks, and the influence domain is divided based on Thiessen polygons. Each site represents a spatial unit. An adjacency matrix is ​​constructed, and then the temporal features of the sites are extracted using an LSTM network. A GNN is used to aggregate spatial dependencies through the adjacency matrix to generate a spatiotemporal joint embedding vector. Based on the spatiotemporal joint embedding vector, the LSTM-DQN decoder is used to predict the resource state of the next T steps time by time, and the spatial distribution is corrected synchronously with GNN to output a spatiotemporal prediction field with uncertainty. The predicted states contained in the spatiotemporal prediction field are input into the DQN main network to calculate the Q value of each management action. Combined with ε-greedy selection of the optimal action, a comprehensive prediction of the future spatiotemporal resource state is generated, and constraint verification is used to ensure compliance with water resource limits and ecological requirements.

2. The agricultural water and soil resource change matching and prediction system based on spatiotemporal data according to claim 1, characterized in that: The state-action modeling unit includes: Define the state variables in the system, quantize the state variables into state vectors that can be input into DQN, and construct the state space; Based on the needs of agricultural water and soil resource management, the feasible operations are determined, and each operation is discretized into a finite set of actions to construct an action space; Design a reward mechanism, define a reward function, and give positive or negative rewards based on the resource matching effect to balance short-term gains and long-term benefits, and guide DQN to learn the optimal strategy; The agricultural water and soil resource matching problem is transformed into a Markov decision process. The state transition probabilities are defined so that the system state transition depends only on the current state and action, satisfying the Markov property. Then, using the predefined state, action and reward space, the DQN is trained in conjunction with the Markov decision process to optimize the agent's decision strategy. The strategy is dynamically adjusted according to environmental feedback to achieve optimal resource matching.

3. The agricultural water and soil resource change matching and prediction system based on spatiotemporal data according to claim 1, characterized in that: The experience playback and target network unit includes: An experience replay mechanism is adopted to store historical interaction data containing historical state-action-reward-new state tuples into a replay buffer to form an experience replay pool. The capacity of the replay pool is managed by a FIFO strategy. In each training iteration, a small batch of data is randomly sampled from the experience replay pool to eliminate temporal correlation. Design a dual-structure architecture of main network and target network, maintain the dual structure of main network and target network, soft copy the target network parameters from main network every C steps, and then use the Q value of target network to calculate TD target, and update target network parameters by delay; A dual-network collaborative training method is adopted, in which the main network calculates the loss based on the experience replay samples and updates the parameters through backpropagation, while the target network provides a stable TD target. The parameters of the main network are synchronized every C steps to enable the policy to be progressively optimized.

4. The agricultural water and soil resource change matching and prediction system based on spatiotemporal data according to claim 1, characterized in that: The policy update unit includes: Based on the DQN training results, the main network obtains the action value in the current state based on the current state input, and outputs the Q-value vector of each action. The argmax is used to select the action with the largest Q value as the candidate policy, while retaining the ε-greedy exploration mechanism to balance utilization and exploration. The relevant actions of the candidate strategy are input into the decision base of the agent to check whether they violate hard constraints. If they do, they are replaced with suboptimal legal actions, and a corrected action probability distribution is generated to gradually optimize the strategy. The optimized strategy, which includes legal actions, is applied to actual agricultural water and soil resource matching decisions. Environmental response data is recorded simultaneously to optimize resource allocation. Actual rewards and new states are fed back to the experience replay pool, forming a continuous iterative cycle of training, deployment, and optimization.

5. The agricultural water and soil resource change matching and prediction system based on spatiotemporal data according to claim 1, characterized in that: The multi-objective matching optimization unit includes: The constraints are encoded as state vectors and input into the DQN network. The matching action is defined as a triple of land plot-crop-irrigation amount. Then, the Q value of each action is output by DQN. At the same time, the action is selected by an ε-greedy strategy to generate an initial matching scheme for each land plot. Using the initial solution generated by DQN as a seed, a genetic algorithm is combined for local search optimization, and a fitness function is constructed. The three objectives of economic benefits, ecological benefits and resource utilization are considered together. The three objectives are weighted and scored to evaluate the overall performance of each solution. Then, genetic operations including selection, crossover and mutation are performed. The genetic operations are executed iteratively to gradually improve the quality of the solution and explore better solutions. The output schemes are subjected to hard constraint verification, water resource limits and land suitability are checked, and for schemes that exceed the limits, the irrigation amount of high water-consuming crops is reduced or replaced with low-demand crops according to priority. Then, the Pareto output is used to screen out the non-dominated solution set and provide decision-makers with a library of feasible schemes for multi-objective trade-offs. The system outputs an optimized and validated database of agricultural water and soil resource matching solutions. Each solution in the database contains detailed information on land plots, crops, and irrigation amounts, as well as corresponding economic, ecological, and resource utilization scores, providing decision-makers with feasible solutions that balance multiple objectives.

6. The agricultural water and soil resource change matching and prediction system based on spatiotemporal data according to claim 5, characterized in that: The visual interaction module includes: The agricultural water and soil matching scheme output by the system is presented in the form of dynamic charts and geographic heat maps, including land block level views, time axis interaction and target comparison panel; It provides a drag-and-drop control panel and condition input forms, allowing users to dynamically adjust constraints, including resource constraints, target weights, and crop preferences; The system responds instantly to user actions, generates interactive reports, and issues constraint conflict warnings. If user input results in no feasible solution, conflict parameters are highlighted, and Pareto solution navigation is performed to locate the user-defined solution in the 3D target space, recommending nearby high-quality solutions. A one-click export function is also available, supporting the export of solution libraries as Excel / PDF, including matching tables, scores, and visualization screenshots to assist in decision-making and reporting.

Images