Spring maize drought identification diagnosis method and system based on deep learning

Abstract

The application belongs to the technical field of deep learning, and particularly relates to a spring maize drought identification and diagnosis method and system based on deep learning. The method comprises the following steps: acquiring soil moisture time series data, crop growth index data, real-time meteorological data, forecast meteorological data and water supply capacity data of each monitoring point; determining the drought time series state of each monitoring point based on a deep learning time series model; determining the water stress grade and water stress occurrence probability according to the drought time series state, the forecast meteorological data and the water supply capacity data; determining the irrigation time window and the water demand of the monitoring point to be irrigated when the conditions are met; and generating an irrigation scheduling scheme according to the irrigation time window, the water demand, the water stress grade, the water stress occurrence probability and the water supply capacity data. The application can realize the linkage processing of drought identification, risk determination and irrigation scheduling.

Description

Technical Field

[0001] This application belongs to the field of deep learning technology, specifically a method and system for identifying and diagnosing drought conditions in spring maize based on deep learning. Background Technology

[0002] In spring maize production, water supply directly affects seedling uniformity, vegetative growth stability, tasseling and pollination quality, and grain formation, thus influencing yield, irrigation costs, and regional water resource utilization efficiency. As agricultural management gradually shifts from experience-based to data-driven scheduling, utilizing intelligent analysis technologies such as deep learning to comprehensively assess multi-source farmland data has become an important development direction for crop growth monitoring and precision water management.

[0003] In existing technologies, drought assessments often rely on single soil moisture indicators, fixed threshold rules, or manual experience analysis. These methods struggle to simultaneously reflect the continuous correlation between changes in soil water loss, crop growth fluctuations, and short-term meteorological disturbances, leading to unstable identification of drought development trends. While some solutions can output drought levels, they typically remain at a static diagnostic level, lacking follow-up processing mechanisms that integrate with future precipitation changes, water supply capacity constraints, and site differences, making it difficult to formulate actionable irrigation plans. Furthermore, some solutions rely solely on single-point data, lacking a comprehensive understanding of spatial differences and scheduling priorities among multiple monitoring points, easily resulting in a disconnect between assessment results and irrigation implementation. Summary of the Invention

[0004] To address the above issues, this application provides a deep learning-based method and system for identifying and diagnosing spring maize drought, which aims to at least solve the problem of how to accurately identify spring maize drought based on multi-source time-series data and further generate an executable irrigation scheduling plan.

[0005] To achieve the above objectives, the technical solution adopted in this application is as follows: Firstly, this application provides a deep learning-based method for identifying and diagnosing drought conditions in spring maize, the method comprising: Acquire time-series data on soil moisture, crop growth indicators, real-time meteorological data, forecast meteorological data, and water supply capacity data at each monitoring point; Based on a deep learning time series model, soil moisture time series data, crop growth index data and real-time meteorological data are fused to determine the drought time series status of each monitoring point. Based on the drought time sequence, forecast meteorological data, and water supply capacity data, the water stress level and probability of water stress occurrence at each monitoring point were determined. When the water stress level reaches the preset level and the probability of water stress occurrence exceeds the preset threshold, the corresponding monitoring point is identified as the monitoring point to be irrigated, and the irrigation time window and water demand of the monitoring point to be irrigated are determined. Based on the irrigation time window, water demand, water stress level, probability of water stress occurrence, and water supply capacity data of the monitoring points to be irrigated, irrigation priorities are determined and water is allocated to generate an irrigation scheduling plan.

[0006] In one possible implementation, soil moisture time-series data, crop growth index data, and real-time meteorological data are fused using a deep learning time-series model to determine the drought time-series status of each monitoring point. This includes: time-aligning the soil moisture time-series data, crop growth index data, and real-time meteorological data to obtain multi-source time-series data; inputting the multi-source time-series data into the deep learning time-series model to obtain the time-series characteristics of each monitoring point; and determining the drought time-series status of each monitoring point based on the time-series characteristics of each monitoring point.

[0007] In one possible implementation, the drought time series status includes drought level, drought duration and drought change trend, with the drought change trend used to indicate whether the drought worsens, stabilizes or eases at the corresponding monitoring point.

[0008] In one possible implementation, determining the water stress level and probability of water stress occurrence at each monitoring point includes: determining the duration of drought based on the drought time series status; determining the rate of soil moisture decline based on soil moisture time series data; determining the change in crop growth indicators based on crop growth indicator data; determining the forecasted precipitation and forecasted number of rainless days based on forecasted meteorological data; determining the water supply capacity indicators based on water supply capacity data; and determining the water stress level and probability of water stress occurrence at each monitoring point based on the soil moisture decline rate, drought duration, change in crop growth indicators, forecasted precipitation, forecasted number of rainless days, and water supply capacity indicators.

[0009] In one possible implementation, the water stress level and probability of occurrence of water stress at each monitoring point are determined based on the rate of decrease in soil moisture, duration of drought, changes in crop growth indicators, forecasted precipitation, forecasted number of rainless days, and water supply capacity indicators. This includes: constructing a classification feature vector based on the rate of decrease in soil moisture, duration of drought, changes in crop growth indicators, forecasted precipitation, forecasted number of rainless days, and water supply capacity indicators; and inputting the classification feature vector into a random forest model to obtain the water stress level and probability of occurrence of water stress at each monitoring point.

[0010] In one possible implementation, determining the irrigation time window and water demand for the monitoring point to be irrigated includes: determining the number of forecast rainless days based on forecast meteorological data; determining the soil moisture decline rate based on the soil moisture time series data corresponding to the monitoring point to be irrigated; determining the crop growth index change trend based on the crop growth index data corresponding to the monitoring point to be irrigated; and determining the irrigation time window and water demand for the monitoring point to be irrigated based on dynamic programming, according to the forecast rainless days, the soil moisture decline rate, and the crop growth index change trend.

[0011] In one possible implementation, determining irrigation priorities and allocating water includes: grouping the monitoring points to be irrigated into multiple irrigation groups based on their geographical location, soil type, and water demand; determining the spring maize growth stage corresponding to each irrigation group based on crop growth index data of the monitoring points; determining the irrigation priority of each irrigation group based on the average probability of water stress occurrence within each group and the spring maize growth stage corresponding to each group; and allocating water based on irrigation priority, irrigation time window, water demand, and water supply capacity data to generate an irrigation scheduling plan.

[0012] In one possible implementation, the irrigation monitoring points are grouped, including: clustering the irrigation monitoring points based on their geographical location, soil type, and water demand using a clustering algorithm to obtain multiple irrigation groups; and water allocation is performed, including: determining the allocated water volume for the multiple irrigation groups based on linear programming.

[0013] In one possible implementation, the water supply capacity data includes real-time water supply and channel distribution limit; the method also includes: when the total allocated water volume corresponding to the irrigation scheduling scheme exceeds the real-time water supply, or when the planned water supply corresponding to at least one irrigation time window exceeds the channel distribution limit, the irrigation scheduling scheme is iteratively corrected according to irrigation priority, water stress level and water stress occurrence probability to obtain an executable irrigation scheduling scheme.

[0014] Secondly, this application provides a deep learning-based spring maize drought identification and diagnosis system for implementing a deep learning-based spring maize drought identification and diagnosis method. The system includes: The data acquisition module is used to acquire time-series data of soil moisture, crop growth index data, real-time meteorological data, forecast meteorological data, and water supply capacity data at each monitoring point. The status recognition module is used to perform time-series fusion of soil moisture time-series data, crop growth index data and real-time meteorological data based on a deep learning time-series model to determine the drought time-series status of each monitoring point. The stress determination module is used to determine the water stress level and the probability of water stress occurrence at each monitoring point based on the drought time sequence, forecast meteorological data, and water supply capacity data. The irrigation determination module is used to determine the corresponding monitoring point as the monitoring point to be irrigated when the water stress level reaches the preset level and the probability of water stress occurrence exceeds the preset threshold, and to determine the irrigation time window and water demand of the monitoring point to be irrigated. The scheduling generation module is used to determine irrigation priorities and allocate water based on the irrigation time window, water demand, water stress level, water stress occurrence probability and water supply capacity data of the monitoring points to be irrigated, and generate an irrigation scheduling plan.

[0015] Compared with existing technologies, the advantages and beneficial effects of this application are as follows: By acquiring time-series data on soil moisture, crop growth indicators, real-time meteorological data, forecast meteorological data, and water supply capacity data, the unified organization and collaborative utilization of multi-source data have been achieved, avoiding the shortcomings of existing technologies that rely on a single indicator, leading to biased judgments.

[0016] By using a deep learning time series model to fuse soil moisture time series data, crop growth index data, and real-time meteorological data, the continuous change process of drought conditions at monitoring points can be identified, and instantaneous observations can be extended into state results with time evolution characteristics.

[0017] By combining drought time-series data, forecast meteorological data, and water supply capacity data, the water stress level and the probability of water stress occurrence are determined, realizing the transformation from static judgment to risk assessment, thus providing a clear basis for subsequent scheduling.

[0018] By further determining the irrigation time window and water demand at monitoring points that meet the conditions, the diagnostic results are directly linked to the irrigation demand results.

[0019] By combining irrigation time windows, water demand, water stress level, water stress occurrence probability and water supply capacity data to determine irrigation priorities and allocate water, scheduling and coordination for multiple monitoring points are realized, so that the output results can directly serve irrigation execution. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the method described in this application; Figure 2 This is a block diagram of the module composition of the system in this application. Detailed Implementation

[0021] To enable those skilled in the art to better understand the technical solution, the present application will be described in detail below with reference to the embodiments. The description in this section is only exemplary and explanatory, and should not be used to limit the scope of protection of the present application in any way.

[0022] Deep learning is a data processing technique centered on multi-layer feature extraction and nonlinear mapping, capable of automatically learning high-level representational relationships with discriminative significance from raw data or multi-source heterogeneous data. Compared to processing methods that rely on manually set rules or single empirical thresholds, deep learning is more suitable for handling complex objects with temporal, coupled, and fluctuating characteristics, especially for the overall representation of continuously changing processes. In the field of agricultural information processing, there are obvious temporal correlations and interactive influences among changes in soil moisture, crop growth status, and meteorological disturbances. Using deep learning methods can transform multi-dimensional observation data into unified state representation results, providing a more stable data foundation for subsequent identification, prediction, and decision-making. Based on this, this application introduces a deep learning temporal modeling mechanism to the spring maize drought identification and diagnosis process, fusing and analyzing soil moisture time-series data, crop growth index data, and real-time meteorological data, and further conducting water stress determination and irrigation scheduling generation on this basis.

[0023] like Figure 1 As shown, a deep learning-based method for identifying and diagnosing drought conditions in spring maize includes: Acquire time-series data on soil moisture, crop growth indicators, real-time meteorological data, forecast meteorological data, and water supply capacity data at each monitoring point; In this embodiment, a unified sampling timeline is established around each monitoring point, and data output from soil moisture sensors, crop growth acquisition devices, field weather stations, weather forecast interfaces, and water supply scheduling terminals are integrated. Soil moisture time-series data reflects changes in topsoil water content; crop growth index data characterizes changes in plant height, leaf area index, canopy temperature, or leaf color; real-time meteorological data characterizes current rainfall, temperature, wind speed, and air humidity; forecast meteorological data characterizes rainfall trends and evapotranspiration conditions within a preset period; and water supply capacity data characterizes current available water volume, channel distribution capacity, and water supply time constraints. Time alignment, missing data completion, anomaly removal, and monitoring point number association are performed on all types of data to obtain a standardized time-series dataset corresponding to each monitoring point, which can then be used for subsequent drought time-series status identification, water stress determination, and irrigation scheduling calculations.

[0024] Based on a deep learning time series model, soil moisture time series data, crop growth index data and real-time meteorological data are fused to determine the drought time series status of each monitoring point. In this embodiment, a continuous time-series analysis unit is established around each monitoring point. Soil moisture time-series data, crop growth index data, and real-time meteorological data are read from a standardized time-series dataset, and these three types of data are organized into an input sequence for the same monitoring point and the same time axis. A deep learning time-series model is used to jointly encode data from different sources but with temporal correlations, extracting time-series features that can characterize the drought development process. Based on the time-series features, the drought time-series state corresponding to each monitoring point at the current analysis time is output. This drought time-series state is not a single instantaneous judgment value, but a state result with continuous evolution attributes, reflecting both the current water stress level at the monitoring point and the continuation and direction of change of the drought over a continuous time range, thus providing a unified state basis for subsequent water stress level determination and irrigation scheduling calculations.

[0025] The drought time series status of each monitoring point is determined by time-series fusion of soil moisture time series data, crop growth index data, and real-time meteorological data using a deep learning time series model. This includes: time-aligning the soil moisture time series data, crop growth index data, and real-time meteorological data to obtain multi-source time series data; inputting the multi-source time series data into the deep learning time series model to obtain the time series characteristics of each monitoring point; and determining the drought time series status of each monitoring point based on the time series characteristics.

[0026] In one embodiment, the added constraint at this stage is to establish a unified time reference for soil moisture time-series data, crop growth index data, and real-time meteorological data, and to form multi-source time-series data that can be directly input into a deep learning time-series model using this unified time reference. The reason for this constraint is that the update frequency, sampling interval, and timestamp accuracy of different data sources are usually inconsistent. If the time correspondence is not first organized, the sequences received by the deep learning time-series model will have positional mismatches, thus affecting the drought status identification results.

[0027] In practice, soil moisture records, crop growth index records, and real-time meteorological records are extracted from a preset analysis window for each monitoring point and merged according to the monitoring point number and collection time. The time base can be set at the hourly, half-day, or daily scale, with the specific value determined based on the farmland monitoring frequency, crop growth rate, and irrigation scheduling accuracy requirements. When the soil moisture sensor sampling frequency is higher than the crop growth index sampling frequency, the time granularity of the lower-frequency data is used as the base granularity for time aggregation of the higher-frequency data. When real-time meteorological data exhibits minute-level changes while the analysis window uses an hourly sequence, the temperature, air humidity, wind speed, and rainfall within the same hour are converted into real-time meteorological records for that hour according to preset aggregation rules. Specifically, temperature and air humidity can use the time-period average, rainfall can use the time-period cumulative value, and wind speed can use the time-period average or maximum value.

[0028] For data points that are missing, supplementary recording is performed. If the continuous missing time does not exceed the preset duration, interpolation between adjacent time points, filling with the most recent valid value, or correction with reference values ​​from nearby sensors in the same plot can be used. If the continuous missing time exceeds the preset duration, the time period is marked as a low confidence interval and is restricted from directly entering the model training, and is only used for supplementary judgment in the inference stage.

[0029] Data points with obvious abnormalities are removed. Anomaly judgment can be determined by combining the range of sensor physical quantities, the jump variables between adjacent time points, and the historical seasonal intervals. For example, when the sudden change in soil moisture exceeds the anomaly judgment range and there are no irrigation records or rainfall records during the same period, it can be judged as an anomaly point. The anomaly judgment range is set based on the measurement error range of the soil moisture sensor, the historical fluctuation range of the same season, and the upper limit of normal changes within the sampling period.

[0030] After cleaning, soil moisture sequences, crop growth index sequences, and real-time meteorological sequences at the same monitoring point under a unified time reference are concatenated in chronological order to form multi-source time-series data. The deep learning time-series model receives the multi-source time-series data, performs joint encoding on the inputs at each time point, and outputs the time-series features corresponding to the monitoring point. In this application, the time-series features are used to characterize the temporal correlation between changes in soil moisture, crop growth, and meteorological disturbances at the monitoring point.

[0031] Deep learning time series models can employ sequence modeling networks for multi-source time series data, consisting of an input encoding layer, a time series feature extraction layer, and a state output layer. The input encoding layer performs a unified dimensional mapping on soil moisture sequences, crop growth index sequences, and real-time meteorological sequences, forming a joint input vector arranged by time steps. The time series feature extraction layer extracts continuous dependencies between adjacent time steps and coupling changes between different data sources, and can employ one of the following: gated recurrent unit networks, long short-term memory networks, temporal convolutional networks, or temporal attention networks. The state output layer maps the extracted time series features to drought status determination results or state representation results for corresponding monitoring points. This type of model is adopted because soil moisture changes, crop growth changes, and meteorological disturbances all have continuous temporal correlations. Relying solely on single-moment data is insufficient to reflect the drought evolution process. Deep learning time series models can jointly model multi-source sequences under a unified time axis, thus providing a stable feature foundation for subsequent drought time series status determination.

[0032] When generating drought time series status based on temporal characteristics, a combination of a model's final classification layer and state mapping rules can be used. The model first outputs the confidence results corresponding to each state category, and then determines the current drought time series status of the monitoring point based on the maximum confidence result and state switching rules. The state switching rules are set based on the continuity of states between adjacent analysis windows, the minimum number of persistent analysis units, and the direction of soil moisture change. If the state results of two adjacent analysis windows change abruptly, but soil moisture and crop growth do not change accordingly, a smoothing constraint is applied to the current state result to avoid frequent jumps in drought status caused by short-term noise. Through this process, the source, meaning, and path to subsequent steps of the drought time series status remain clear.

[0033] The drought time series status includes drought level, drought duration and drought change trend. The drought change trend is used to indicate whether the drought is aggravated, stabilized or alleviated at the corresponding monitoring point.

[0034] In one embodiment, the newly added limitation in this stage is to further define the composition and usage of the drought time series status, so that the drought time series status includes not only the drought level, but also the drought duration and drought trend. The reason for introducing this limitation is that simply outputting the drought level can only reflect the state category at a certain moment, and cannot fully express the degree of drought continuation and evolution direction within the continuous analysis window, which is not conducive to subsequent judgment of water stress risk and irrigation timing.

[0035] In practice, the drought level is represented by a tiered system, indicating the combined state of current soil moisture and crop drought at the monitoring points. It can be categorized into normal, mild drought, moderate drought, and severe drought. The tier classification is determined based on the state category results output by the deep learning time-series model, and can also be validated by combining the recent deviations in soil moisture, crop growth indicators, and real-time weather conditions.

[0036] When there is a significant conflict between the model output level and the validation results, the model output should be prioritized, with the validation results used as a secondary factor for correction. The correction rule can be set to adjust the level only when the conflict persists for multiple analysis windows, in order to avoid misjudgments caused by occasional anomalies. Drought duration indicates the length of time that the current drought level at a monitoring point has been maintained continuously. It is determined by counting the number of analysis units that are continuously at the same drought level or in the same drought-affected area, starting from the current analysis time, and converting this number into duration based on the time reference. If only a slight fluctuation occurs in one analysis unit, and the state before and after is consistent, this fluctuation can be considered noise, and the duration calculation will not be interrupted.

[0037] Drought trend analysis indicates the direction of drought change at monitoring points between adjacent analysis windows, and can be categorized into worsening drought, stable drought, and mitigating drought. Worsening drought indicates that the drought severity in the current analysis window is higher than in the previous analysis window, or that the drought level remains unchanged but soil moisture decreases and crop growth weakens simultaneously. Stable drought indicates that the change in state between the current and previous analysis windows is less than a preset threshold. Mitigating drought indicates that the drought severity in the current analysis window is lower than in the previous analysis window, or that soil moisture continuously recovers along with crop growth.

[0038] The preset threshold used in determining the trend of drought is set based on the monitoring time granularity, crop growth stage, and local water fluctuation level. For example, a smaller threshold can be used during the rapid growth period to enhance the sensitivity to drought changes, while a larger threshold can be used during the slow growth period to suppress misjudgments caused by normal fluctuations. After the joint determination of drought level, drought duration, and drought trend is completed, these three factors are combined to form the drought time-series status record corresponding to the monitoring point and written into the status buffer.

[0039] The records in the state buffer are used for two purposes: firstly, to determine the water stress level at the current moment, and secondly, to update the duration and compare trends of the next analysis window. This organizational approach allows the drought time series status to serve as both an independent basis for judgment and an intermediate foundation for continuous time-series assessments in subsequent processes, making the entire identification process repeatable, traceable, and updatable.

[0040] Based on the drought time sequence, forecast meteorological data, and water supply capacity data, the water stress level and probability of water stress occurrence at each monitoring point were determined. In this embodiment, based on the drought time series status formed in the previous stage, combined with forecast meteorological data and water supply capacity data, a water stress risk assessment is conducted for each monitoring point. The drought time series status is used to characterize the current drought severity, its persistence, and changes at the monitoring point; the forecast meteorological data is used to characterize the possible rainfall replenishment conditions in the subsequent analysis period; and the water supply capacity data is used to characterize the current available irrigation support conditions. State quantities, meteorological quantities, and water supply quantities related to water stress are extracted for each monitoring point, and these various assessment quantities are organized into a unified risk assessment input. Based on this, the water stress level and water stress occurrence probability of each monitoring point are output. The water stress level is used to characterize the drought severity that the monitoring point may reach within a future preset period, and the water stress occurrence probability is used to characterize the likelihood of this drought severity occurring. Both serve as the basis for subsequent screening of monitoring points to be irrigated and irrigation scheduling calculations.

[0041] The determination of the water stress level and probability of occurrence at each monitoring point includes: determining the duration of drought based on the drought time series data; determining the rate of soil moisture decline based on soil moisture time series data; determining the changes in crop growth indicators based on crop growth indicator data; determining the forecasted precipitation and forecasted number of rainless days based on forecasted meteorological data; determining the water supply capacity indicators based on water supply capacity data; and determining the water stress level and probability of occurrence at each monitoring point based on the rate of soil moisture decline, drought duration, changes in crop growth indicators, forecasted precipitation, forecasted number of rainless days, and water supply capacity indicators.

[0042] In one embodiment, the added constraint at this stage is to clarify the input composition of the water stress level and the probability of water stress occurrence, as well as the generation method of each input quantity. This ensures that the risk assessment result not only originates from the current drought situation but also reflects future meteorological changes and water supply constraints. The reason for setting this constraint is that if the assessment is based solely on the current drought time series status, while it can reflect the current drought severity at the monitoring point, it cannot reflect the impact of future rainfall and irrigation accessibility on drought development, easily leading to inconsistencies between the current state and subsequent risks.

[0043] In practice, a water stress assessment record is established for each monitoring point. The duration of drought is directly extracted from the drought time series data to characterize the length of time the current drought condition has been sustained. The soil moisture decline rate is determined based on the soil moisture time series data for that monitoring point within a preset backtracking window. The backtracking window can be set to multiple consecutive analysis units according to the monitoring cycle. If there are abnormal fluctuations in moisture records at adjacent times, outliers are first removed according to the data cleaning rules of the previous stage, and then the decline rate is calculated based on the valid data.

[0044] The changes in crop growth indicators are determined based on the differences or magnitudes of changes in crop growth indicator data within a preset backtracking window, and are used to characterize the degree of deviation in crop growth under the current drought background. Forecast meteorological data is used to generate two key quantities: forecast precipitation, which characterizes the cumulative rainfall replenishment capacity in subsequent analysis periods, and forecast rainless days, which characterizes the duration of continuous absence of effective rainfall in the future. The forecast rainless days are determined by the continuous accumulation of forecast days below a preset rainfall threshold; the rainfall threshold can be set by combining local crop water requirements, soil infiltration conditions, and conventional irrigation experience to ensure that the rainless determination is consistent with the application scenario of this application.

[0045] Water supply capacity data is used to generate water supply capacity indicators. In this application, water supply capacity indicators characterize the water supply support level of the area where the monitoring point is located within a preset scheduling cycle. These indicators can be determined based on the current available water volume, channel transmission capacity, and available water supply periods. When monitoring points are located in the same irrigation district and share the same water supply branch, the water supply capacity indicators can be generated uniformly at the irrigation district level and then allocated to each monitoring point. When different monitoring points are located on different water supply branches, the water supply capacity indicators are generated separately at the branch level.

[0046] After extracting the aforementioned parameters, the duration of drought, the rate of soil moisture decline, changes in crop growth indicators, forecasted precipitation, forecasted number of rainless days, and water supply capacity indicators are written into the risk assessment record of the corresponding monitoring point. When the risk assessment record enters the subsequent risk classification process, it can simultaneously reflect the current drought status, future rainfall changes, and water supply constraints, thereby avoiding one-sided judgments based solely on the status at a single moment.

[0047] To ensure the usability of the assessment results, a validity check can be added when generating risk assessment records. For example, when forecast meteorological data is missing, the most recent valid forecast data is used to fill in the gaps, and the monitoring point is marked as having limited forecast confidence. When water supply capacity data is missing, water supply capacity indicators are estimated using water supply records from the previous scheduling cycle of the same irrigation district, and these are restricted from being used as a high-priority allocation basis in subsequent scheduling phases. Through this approach, the input sources, formation logic, and paths to subsequent models for water stress levels and the probability of water stress occurrence remain clear.

[0048] Based on the rate of soil moisture decline, duration of drought, changes in crop growth indicators, forecasted precipitation, forecasted number of rainless days, and water supply capacity indicators, the water stress level and probability of water stress occurrence at each monitoring point are determined. This includes: constructing a classification feature vector based on the rate of soil moisture decline, duration of drought, changes in crop growth indicators, forecasted precipitation, forecasted number of rainless days, and water supply capacity indicators; and inputting the classification feature vector into a random forest model to obtain the water stress level and probability of water stress occurrence at each monitoring point.

[0049] In one embodiment, the added constraint at this stage is that a random forest model is used to classify the risk assessment records to output the water stress level and probability of water stress occurrence for each monitoring point. This constraint is set because there is a non-linear correlation among the aforementioned assessment parameters. Under the same soil moisture decline rate, different monitoring points may exhibit different water stress risks due to differences in forecasted precipitation, changes in crop growth indicators, or water supply capacity indicators. Relying solely on fixed threshold rules is insufficient to account for the combined relationships between multiple factors.

[0050] Random forest models can employ an ensemble classification structure composed of multiple decision trees. For each monitoring point, the soil moisture decline rate, drought duration, crop growth index changes, forecast precipitation, forecast rainless days, and water supply capacity indicators from the risk assessment records are used as classification input features and input into the random forest model. Each decision tree in the random forest model completes classification judgment based on historical samples, and each decision tree outputs the corresponding water stress level category. The model then determines the water stress level of the current monitoring point based on the voting results of all decision trees, and determines the probability of water stress occurrence based on the voting percentage corresponding to the target level. Historical samples can be jointly constructed from multi-source monitoring data, irrigation records, and actual drought results from previous monitoring periods. To ensure the stability of classification results, the number of samples of different levels can be balanced during the model training phase, and consistency checks can be added for samples with abnormal fluctuations during the model inference phase. This approach can both utilize the combination relationship between multiple decision variables to complete the level classification and simultaneously provide a probabilistic representation of the risk outcome.

[0051] In practice, a classification feature vector is constructed for each monitoring point based on risk assessment records. This vector consists of soil moisture decline rate, drought duration, changes in crop growth indicators, forecasted precipitation, forecasted rainless days, and water supply capacity indicators. Each feature undergoes dimensional standardization before being fed into the random forest model. Dimensional standardization can employ interval normalization, standard deviation standardization, or hierarchical mapping based on historical distribution. The selection criterion is the degree of difference between the soil moisture range, crop growth indicator dimensions, and water supply data dimensions in different regions.

[0052] Random forest models are trained using historical samples. These historical samples consist of feature vectors of historical monitoring points within a known analysis period and the corresponding actual drought conditions. The actual drought conditions can be jointly labeled based on historical soil moisture records, crop drought survey records, irrigation execution records, and subsequent yield shifts. During training, to prevent the model from being biased towards the majority class due to insufficient samples at a particular level, the sample distribution can be balanced. Balancing methods can include oversampling, undersampling, or setting sample weights by class.

[0053] After model training, a classification feature vector is input for each monitoring point, and the classification results and voting ratios for each level are output. The classification results determine the water stress level, and the voting ratios determine the probability of water stress occurrence. Water stress levels are categorized as normal, mild, moderate, and severe. The probability of water stress occurrence indicates the likelihood of the current output level occurring; a higher value indicates a higher probability that the monitoring point will reach the corresponding water stress level in subsequent analysis periods.

[0054] To ensure the interpretability of the probability values, the vote ratio output by the random forest model can be calibrated using historical validation samples to maintain good consistency with the actual frequency of occurrence. During the model inference phase, a result stabilization rule can be set: if the change in the classification feature vector between the current and previous analysis windows is less than a change threshold, and the difference between the current output level and the previous output level exceeds one level, the current result is marked as pending review. Consistency judgment is then performed based on the results of neighboring windows to prevent frequent level jumps caused by short-term data disturbances. The change threshold is set based on the feature fluctuation range of adjacent analysis windows in historical samples.

[0055] After classification, the water stress level and probability of occurrence of each monitoring point are written into the risk results table. The contents of the risk results table are then used in the subsequent stage of determining the monitoring points to be irrigated, to screen the monitoring points that need to be included in the irrigation scheduling calculation according to their level and probability. After this processing, the judgment results generated in this stage have both a clear data source and a clear usage destination, and can form a stable connection with the subsequent irrigation time window and water demand determination process.

[0056] When the water stress level reaches the preset level and the probability of water stress occurrence exceeds the preset threshold, the corresponding monitoring point is identified as the monitoring point to be irrigated, and the irrigation time window and water demand of the monitoring point to be irrigated are determined. In this embodiment, irrigation targets and irrigation demand determination are performed based on the water stress level and probability of occurrence corresponding to each monitoring point. Monitoring points that reach a preset level and exceed a preset threshold are identified as target targets for the irrigation scheduling phase. The preset level is used to limit the lower limit of drought severity, and the preset threshold is used to limit the confidence level of risk occurrence. The combination of the two is used to avoid directly including short-term fluctuations or low-confidence results in irrigation calculations. After screening, the irrigation time window and water demand of the target targets within a preset scheduling cycle are determined by combining forecast meteorological data, soil moisture time series data, and crop growth index data. This result is written into the irrigation demand record for subsequent irrigation priority determination and water allocation.

[0057] Determining the irrigation time window and water demand for monitoring points to be irrigated includes: determining the number of forecast rainless days based on forecast meteorological data; determining the rate of soil moisture decline based on soil moisture time-series data corresponding to the monitoring points to be irrigated; determining the trend of crop growth index changes based on crop growth index data corresponding to the monitoring points to be irrigated; and determining the irrigation time window and water demand for the monitoring points to be irrigated based on dynamic programming, according to the forecast rainless days, the rate of soil moisture decline, and the trend of crop growth index changes.

[0058] In one embodiment, the added constraint at this stage is to further define the rules for determining the irrigation monitoring points and the process for generating the irrigation time window and water demand. This ensures that the monitoring points entering the irrigation scheduling stage have clear selection criteria and that the irrigation demand results have executable time and water quantity constraints. The reason for introducing this constraint is that if irrigation is directly scheduled based solely on the water stress level, short-term anomalies, local noise, or low-reliability risk results are easily included in subsequent processes, resulting in the dispersion of scheduling resources. If only the water demand is determined without limiting the feasible time range, it is difficult to connect with subsequent water supply capacity constraints.

[0059] In practice, the water stress level and probability of occurrence are read for each monitoring point, and the rules for entering the scheduling phase are set in conjunction with regional irrigation management requirements. The preset level can be set to moderate stress or above, based on the tolerance range of local spring maize to soil water shortage at different growth stages and the scheduling priority principle under the condition of scarce irrigation resources. The preset threshold is used to limit the credibility of the risk results, and can be set according to the consistency level between the classification results in historical samples and the actual drought results. For example, the lower limit of the probability interval that can stably represent the true drought risk can be used as the threshold.

[0060] When a monitoring point meets the level and threshold conditions, it is marked as a monitoring point to be irrigated, and a screening identifier is written into the irrigation demand record. For monitoring points that do not meet the conditions, only the risk assessment result is retained, and they are not included in the current round of irrigation scheduling. For each monitoring point to be irrigated, the number of forecast rainless days is first determined based on the forecast meteorological data. The number of forecast rainless days is used to characterize the number of days in the subsequent analysis period where there is a continuous lack of effective rainfall. It is determined by counting the number of consecutive days in the preset analysis period where the daily rainfall is lower than the effective rainfall judgment value.

[0061] The effective rainfall threshold is set based on local soil infiltration capacity, surface evaporation intensity, and daily water requirement of spring maize, ensuring that the no-rainfall determination results accurately reflect the degree of insufficient natural water replenishment. The soil moisture decline rate is then determined based on the time-series soil moisture data corresponding to the monitoring points to be irrigated. The soil moisture decline rate characterizes the current rate of soil water loss; it is determined by selecting effective moisture records within a preset backtracking window and removing abrupt changes to avoid sensor malfunctions or local disturbances affecting the judgment of the decline trend.

[0062] The changing trends of crop growth indicators are determined based on the crop growth indicator data corresponding to the monitoring points to be irrigated. The changing trends of crop growth indicators are used to characterize the direction of crop growth changes after being affected by water conditions. They can reflect slowed plant height growth, stagnant leaf area expansion, increased canopy temperature, or intensified leaf color changes.

[0063] After extracting the three decision variables mentioned above, the irrigation time window and water demand are determined based on dynamic programming. The dynamic programming stages are divided according to discrete time periods within a preset scheduling cycle, with each discrete time period corresponding to a candidate irrigation time point. The state variables are used to characterize the soil moisture state and crop water demand urgency at the monitoring point to be irrigated at the beginning of each discrete time period, while the decision variables are used to characterize whether irrigation should be carried out during that discrete time period and the water quantity to be used when irrigation is carried out.

[0064] In the solution process, the number of forecast rainless days is used as the basis for time constraints, the rate of decrease in soil moisture and the trend of crop growth indicators are used as the basis for the urgency of water demand, and state transition constraints are set in combination with the water demand sensitivity corresponding to the growth stage of spring maize, so that the irrigation schedule can cover the period of rain shortage and avoid premature irrigation during the period when short-term natural water replenishment may occur.

[0065] The irrigation time window is determined based on the set of discrete time periods that meet the drought control requirements and do not conflict with the rainfall forecast from the dynamic programming solution. The drought control requirements are used to limit the soil moisture status of the monitoring points to be irrigated to not be lower than the minimum suitable value corresponding to the current growth stage within the preset scheduling cycle.

[0066] The water requirement is determined based on the minimum amount of water needed to restore soil moisture to the target range within the selected irrigation time window. The target range is set based on the range of soil moisture required for normal growth of spring maize at the current growth stage, which should be neither too low nor too high to avoid ineffective irrigation.

[0067] If the solution results show that multiple time windows can satisfy the constraints, the combination of time windows that is more favorable for subsequent water supply scheduling is retained; if the solution results show that there is no time window that satisfies the conditions within the current scheduling cycle, the monitoring point to be irrigated is marked as a delayed processing object and recalculated in the next scheduling cycle. Through this process, the source of the monitoring point to be irrigated is clear, the formation path of irrigation time windows and water demand is clear, and it can directly enter the subsequent irrigation priority determination and water allocation stage.

[0068] Based on the irrigation time window, water demand, water stress level, probability of water stress occurrence, and water supply capacity data of the monitoring points to be irrigated, irrigation priorities are determined and water is allocated to generate an irrigation scheduling plan.

[0069] In this embodiment, irrigation priority determination and water allocation are performed based on the irrigation time window, water demand, water stress level, water stress occurrence probability, and water supply capacity data of the monitoring points to be irrigated. The irrigation time window is used to limit the time range for scheduled water supply; the water demand is used to characterize the minimum water replenishment requirement of each monitoring point to be irrigated within the current scheduling cycle; the water stress level and water stress occurrence probability are used to characterize the urgency of irrigation; and the water supply capacity data is used to characterize the water supply boundary within the scheduling cycle. Based on the above data, an irrigation ranking result is formed, and water allocation is completed under the premise of satisfying water supply constraints, generating an irrigation scheduling plan. The irrigation scheduling plan includes at least the monitoring point number, irrigation time period, and allocated water volume for subsequent irrigation execution or scheduling issuance.

[0070] Determining irrigation priorities and allocating water includes: grouping the monitoring points to be irrigated into multiple irrigation groups based on their geographical location, soil type, and water requirement; determining the spring maize growth stage corresponding to each irrigation group based on crop growth index data of the monitoring points; determining the irrigation priority of each irrigation group based on the average probability of water stress occurrence within each group and the corresponding spring maize growth stage; and allocating water based on irrigation priority, irrigation time window, water requirement, and water supply capacity data to generate an irrigation scheduling plan.

[0071] In one embodiment, the added constraint at this stage is to further limit the formation path of irrigation priorities and the input organization method of water allocation, so that the irrigation scheduling scheme is not generated independently based on individual monitoring points, but rather the priority ranking and resource allocation are completed on a group basis. The reason for introducing this constraint is that when there are many irrigation monitoring points, if each point is scheduled independently, it is easy to cause problems such as repeated start and stop of water supply to adjacent plots, scattered scheduling of similar soil plots, and imbalance in the allocation of limited water.

[0072] In practice, the geographical location, soil type, and water requirement are extracted from all monitoring points awaiting irrigation. Geographical location characterizes the spatial distribution of monitoring points in farmland and can be represented by plot number, latitude and longitude coordinates, or irrigation district grid location. Soil type characterizes differences in water-holding capacity and infiltration characteristics and can be determined based on farmland soil survey results, historical planting zoning results, or on-site testing results. Water requirement characterizes the amount of water replenishment needed to meet basic irrigation needs within the current scheduling cycle.

[0073] Based on three types of data, the monitoring points to be irrigated are grouped into multiple irrigation groups. After grouping, the probability of water stress occurrence for each monitoring point within the group is statistically analyzed, and the average probability of water stress occurrence within the group is calculated. In this application, the average probability of water stress occurrence within the group is used to characterize the overall drought risk level of the irrigation group, which can avoid the occasional high-risk value of a single monitoring point from causing excessive disturbance to the entire scheduling sequence.

[0074] For each irrigation group, the growth stage of spring maize is determined based on crop growth index data corresponding to the monitoring points within that group. Spring maize growth stages can be divided into seedling stage, jointing stage, tasseling stage, and grain-filling stage. The determination can be based on a combination of plant height range, leaf unfolding number, canopy cover changes, and growth duration. Spring maize growth stages reflect the crop's sensitivity to water shortage; irrigation groups in critical water-demand stages are given higher priority in the ranking.

[0075] After determining the average probability of water stress within each group and the growth stage of spring maize, irrigation priorities were established for multiple irrigation groups based on the principles of risk priority and priority for critical growth stages. Following the establishment of irrigation priorities, water allocation was carried out by combining irrigation time windows, water demand, and water supply capacity data.

[0076] In this stage, water supply capacity data is used to define the available water volume, available water periods, and transmission and distribution capacity boundaries within the scheduling cycle. If the sum of water demands of high-priority irrigation groups is less than the available water volume, the water demand of high-priority irrigation groups is prioritized. If the sum of water demands of high-priority irrigation groups is close to or exceeds the available water volume, water is allocated according to irrigation priority and overlapping irrigation time windows. After allocation, an irrigation scheduling plan is formed. The results in the irrigation scheduling plan directly enter the subsequent correction or execution stage, thus ensuring that irrigation priority determination and water allocation have clear input sources and output destinations.

[0077] Grouping the irrigation monitoring points involves: clustering the monitoring points based on their geographical location, soil type, and water demand using a clustering algorithm to obtain multiple irrigation groups; and allocating water resources by determining the allocated water volume for the multiple irrigation groups based on linear programming.

[0078] In one embodiment, the added constraint at this stage is that a clustering algorithm is used to group the monitoring points to be irrigated, and linear programming is used to determine the water allocation for multiple irrigation groups, making the grouping and allocation process have a clearer implementation path. The reason for setting this constraint is that there are combined differences between geographical location, soil type, and water demand. If irrigation groups are manually divided according to a single rule, it is easy to cause excessive differences within the same group, affecting subsequent unified scheduling; if only a sequential distribution method is used in the water allocation stage, it is also easy to cause high-priority groups to be satisfied too much, while adjacent groups cannot maintain the minimum irrigation needs.

[0079] In practice, the geographical location, soil type, and water demand of the monitoring points to be irrigated are preprocessed. The geographical location can be converted into spatial distance relationships using planar coordinates, the soil type can be converted into comparable category values ​​according to water retention, permeability, and texture, and the water demand can be kept at its original value or standardized in units according to the scheduling cycle.

[0080] After preprocessing, the three types of data are clustered using a clustering algorithm to obtain multiple irrigation groups. The clustering algorithm can be any of hierarchical clustering, partitioning clustering, or density clustering, and the selection criteria are the density of farmland plots, the dispersion of soil types, and the scale of the irrigation schedule.

[0081] The clustering results must meet two basic requirements: first, monitoring points within the same irrigation group should be relatively concentrated spatially to facilitate unified water supply; second, the difference in soil type within the same irrigation group should not exceed the preset number of soil types, and the difference in water demand should not exceed the preset water demand deviation range, to avoid excessive differences in actual irrigation demand within the same group. The preset water demand deviation range is set based on the average water demand of the plots within the scheduling cycle and the unified water supply capacity of the canals. After completing the irrigation grouping, water is allocated according to irrigation priority, irrigation time window, water demand, and water supply capacity data. Linear programming is used in this application to calculate the allocated water for multiple irrigation groups. The inputs include the total water demand corresponding to each irrigation group, the total available water supply, the upper limit of water supply under each irrigation time window, and the irrigation priority.

[0082] The objective is to prioritize meeting the minimum water demand of high-priority irrigation groups while satisfying water supply and time window constraints. Constraints include, at a minimum, that the sum of water allocated to each irrigation group within the scheduling cycle does not exceed the total available water supply, the allocated water volume within any irrigation time window does not exceed the transmission and distribution capacity of the corresponding water supply period, and the allocated water volume for a single irrigation group is not negative. If multiple feasible solutions exist in the linear programming solution, the solution with fewer water supply switching times and a higher degree of continuous water supply to adjacent plots can be prioritized for practical implementation.

[0083] After the solution is obtained, the allocated water volume of each irrigation group is mapped back to each monitoring point to be irrigated within the group, forming an irrigation scheduling plan that unfolds according to time periods. Using this approach, both the grouping and allocation processes have clear technical paths and can be seamlessly integrated with the subsequent scheduling and correction phases.

[0084] The water supply capacity data includes real-time water supply and channel distribution limit; the method also includes: when the total allocated water volume corresponding to the irrigation scheduling scheme exceeds the real-time water supply, or when the planned water supply corresponding to at least one irrigation time window exceeds the channel distribution limit, the irrigation scheduling scheme is iteratively corrected according to irrigation priority, water stress level and water stress occurrence probability to obtain an executable irrigation scheduling scheme.

[0085] In one embodiment, the added constraint at this stage is to further define the composition of the water supply capacity data and the correction trigger conditions and correction paths for the irrigation scheduling scheme, so that the output result not only satisfies the sorting logic but also the actual water supply capacity boundary. The reason for introducing this constraint is that the irrigation scheduling scheme directly generated based on irrigation priority and water demand may still be unable to be directly executed due to insufficient total water supply or excessive distribution in local periods. Without adding a correction process, although the scheduling result is valid in terms of sorting, it may not be feasible in actual irrigation district operation.

[0086] In practice, water supply capacity data includes real-time water supply volume and channel distribution limit. Real-time water supply volume represents the total amount of water available for allocation within the current scheduling cycle, and can be obtained from reservoir discharge records, pump station discharge records, or water supply plans issued by the superior scheduling platform. Channel distribution limit represents the maximum water supply capacity that a channel can carry during each water supply period, and its determination can be based on channel cross-sectional capacity, valve opening range, pump station operating capacity, and historical safe distribution records.

[0087] After generating the irrigation scheduling plan, its executability is checked. If the total allocated water volume corresponding to the irrigation scheduling plan exceeds the real-time water supply, it indicates that the total allocation exceeds the water supply boundary of the current scheduling cycle; if the planned water supply corresponding to at least one irrigation time window exceeds the channel distribution limit, it indicates that there is a risk of distribution overload in local periods. In either case, a scheduling correction process is triggered. The scheduling correction process iteratively processes water stress based on irrigation priority, water stress level, and the probability of water stress occurrence. The correction order follows the principle of prioritizing high-priority water stress, prioritizing severe stress, and prioritizing high-probability risk.

[0088] In specific processing, priority is given to compressing the water allocation for monitoring points in low-priority irrigation groups that are in non-critical growth stages and have a low probability of water stress. If the constraints still cannot be met after compression, some irrigation time windows are adjusted, and monitoring points that can be postponed are moved to subsequent water supply periods. If the total amount constraint and the time constraint still cannot be met at the same time, the minimum guaranteed water volume for each monitoring point is retained, and the remaining gap is recorded as the compensation requirement for the next scheduling cycle.

[0089] After each round of adjustments, the relationship between the total allocated water volume and the real-time water supply, as well as the relationship between the planned water supply for each irrigation time window and the upper limit of canal distribution, are re-examined until all constraints are met, at which point an executable irrigation scheduling plan is output. In addition to retaining the monitoring point numbers, water supply periods, and allocated water volumes from the original irrigation scheduling plan, the executable irrigation scheduling plan can also add correction markers to distinguish between the original allocation results and the corrected allocation results. Through this processing method, the irrigation scheduling results can simultaneously meet the requirements of irrigation urgency and water supply boundaries, making the scheduling plan ready for direct execution.

[0090] like Figure 2 As shown, a deep learning-based spring maize drought identification and diagnosis system is used to implement a deep learning-based spring maize drought identification and diagnosis method. The system includes: The data acquisition module is used to acquire soil moisture time-series data, crop growth index data, real-time meteorological data, forecast meteorological data, and water supply capacity data at each monitoring point. The data acquisition module consists of a soil moisture sensor, a crop growth information acquisition device, a field meteorological monitoring terminal, a forecast data access unit, a water supply status acquisition terminal, a communication interface circuit, and a data buffer circuit. It is used to complete the acquisition, reception, aggregation, and temporary storage of soil moisture time-series data, crop growth index data, real-time meteorological data, forecast meteorological data, and water supply capacity data at each monitoring point.

[0091] The state recognition module is used to perform time-series fusion of soil moisture time-series data, crop growth index data, and real-time meteorological data based on a deep learning time-series model to determine the drought time-series status of each monitoring point. The state recognition module consists of a processor, a memory, a deep learning inference unit, and a time-series data processing circuit. The memory stores the deep learning time-series model program, and the processor calls the deep learning inference unit to perform time-series fusion of soil moisture time-series data, crop growth index data, and real-time meteorological data to determine the drought time-series status of each monitoring point.

[0092] The stress determination module is used to determine the water stress level and probability of occurrence of water stress at each monitoring point based on the drought time series status, forecast meteorological data, and water supply capacity data. The stress determination module consists of a feature extraction circuit, a risk calculation unit, a processor, and a memory. The feature extraction circuit is used to extract the determination features from the drought time series status, forecast meteorological data, and water supply capacity data. The risk calculation unit, under the control of the processor, completes the calculation of the water stress level and probability of occurrence of water stress, and outputs the stress determination results for each monitoring point.

[0093] The irrigation determination module is used to identify monitoring points as irrigation points when the water stress level reaches a preset level and the probability of water stress occurrence exceeds a preset threshold, and to determine the irrigation time window and water demand for the monitoring points to be irrigated. The irrigation determination module consists of a condition judgment circuit, an irrigation decision unit, a processor, and a memory. The condition judgment circuit is used to determine whether the water stress level has reached a preset level and whether the probability of water stress occurrence exceeds a preset threshold. The irrigation decision unit is used to determine the irrigation time window and water demand for the monitoring points that meet the conditions, thereby determining the monitoring points to be irrigated.

[0094] The scheduling generation module is used to determine irrigation priorities and allocate water based on the irrigation time window, water demand, water stress level, water stress occurrence probability, and water supply capacity data of the monitoring points to be irrigated, thereby generating an irrigation scheduling plan. The scheduling generation module consists of a priority sorting unit, a water allocation unit, a processor, a memory, and a scheduling output interface. The priority sorting unit determines the irrigation priority based on the irrigation time window, water demand, water stress level, water stress occurrence probability, and water supply capacity data of the monitoring points to be irrigated. The water allocation unit generates the corresponding water allocation results, and the scheduling output interface outputs the irrigation scheduling plan.

[0095] It should be noted that, in this document, the terms "comprising," "including," and any other variations are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Specific examples have been used in this document to illustrate the principles and implementation methods of the technical solutions of this application. The above examples are only for the purpose of helping to understand the methods and core ideas of this application. The above descriptions are merely preferred embodiments of this application. It should be pointed out that, due to the limitations of written expression and the objective existence of infinite specific structures, those skilled in the art can make several improvements, modifications, or changes without departing from the principles of this application, and can also combine the above technical features in an appropriate manner; these improvements, modifications, changes, or combinations, or the direct application of the concept and technical solutions of this application to other situations without modification, should all be considered within the scope of protection of this application.

Claims

1. A deep learning-based method for identifying and diagnosing drought conditions in spring maize, characterized in that, The method includes: Acquire time-series data on soil moisture, crop growth indicators, real-time meteorological data, forecast meteorological data, and water supply capacity data at each monitoring point; The soil moisture time-series data, crop growth index data and real-time meteorological data are fused according to a deep learning time-series model to determine the drought time-series status of each monitoring point. Based on the drought time sequence, the forecast meteorological data, and the water supply capacity data, the water stress level and the probability of water stress occurrence at each monitoring point are determined. When the water stress level reaches a preset level and the probability of water stress occurrence exceeds a preset threshold, the corresponding monitoring point is identified as a monitoring point to be irrigated, and the irrigation time window and water demand of the monitoring point to be irrigated are determined. Based on the irrigation time window, water demand, water stress level, water stress occurrence probability, and water supply capacity data of the monitoring points to be irrigated, irrigation priority is determined and water is allocated to generate an irrigation scheduling plan.

2. The method according to claim 1, characterized in that, The step of performing time-series fusion of the soil moisture time-series data, the crop growth index data, and the real-time meteorological data based on a deep learning time-series model to determine the drought time-series status of each monitoring point includes: The soil moisture time-series data, the crop growth index data, and the real-time meteorological data are time-aligned to obtain multi-source time-series data; The multi-source time-series data is input into the deep learning time-series model to obtain the time-series features of each monitoring point; The drought temporal status of each monitoring point is determined based on the temporal characteristics of each monitoring point.

3. The method according to claim 1, characterized in that, The drought situation time series status includes drought level, drought duration and drought change trend. The drought change trend is used to indicate whether the drought is aggravated, stabilized or alleviated at the corresponding monitoring point.

4. The method according to claim 1, characterized in that, Determining the water stress level and the probability of water stress occurrence at each monitoring point includes: The duration of the drought is determined based on the drought time sequence. The rate of decrease in soil moisture was determined based on the soil moisture time-series data. The changes in crop growth indicators are determined based on the crop growth indicator data. The forecast precipitation and the number of rainless days are determined based on the forecast meteorological data. Determine water supply capacity indicators based on the aforementioned water supply capacity data; Based on the soil moisture decline rate, drought duration, crop growth index changes, forecast precipitation, forecast rainless days, and water supply capacity index, the water stress level and water stress occurrence probability of each monitoring point are determined.

5. The method according to claim 4, characterized in that, The determination of the water stress level and probability of water stress occurrence at each monitoring point based on the soil moisture decline rate, drought duration, crop growth index changes, forecasted precipitation, forecasted rainless days, and water supply capacity index includes: A classification feature vector is constructed based on the soil moisture decrease rate, drought duration, crop growth index changes, forecast precipitation, forecast number of rainless days, and water supply capacity index. The classification feature vector is input into the random forest model to obtain the water stress level and water stress occurrence probability of each monitoring point.

6. The method according to claim 1, characterized in that, Determining the irrigation time window and water demand of the monitoring point to be irrigated includes: The number of rainless days is determined based on the forecast meteorological data. The rate of decrease in soil moisture is determined based on the soil moisture time-series data corresponding to the monitoring points to be irrigated; The trend of crop growth index changes is determined based on the crop growth index data corresponding to the monitoring points to be irrigated. Based on the predicted number of rainless days, the rate of decrease in soil moisture, and the changing trends of crop growth indicators, the irrigation time window and water demand of the monitoring points to be irrigated are determined using dynamic programming.

7. The method according to claim 1, characterized in that, The process of determining irrigation priorities and allocating water includes: The monitoring points to be irrigated are grouped according to their geographical location, soil type, and water requirement, resulting in multiple irrigation groups. The spring maize growth stage corresponding to the multiple irrigation groups is determined based on the crop growth index data of the monitoring points to be irrigated. The irrigation priority of the multiple irrigation groups is determined based on the average probability of water stress occurrence within each group and the spring maize growth stage corresponding to each irrigation group. The irrigation scheduling scheme is generated by allocating water according to the irrigation priority, irrigation time window, water demand, and water supply capacity data.

8. The method according to claim 7, characterized in that, The grouping of the monitoring points to be irrigated includes: Based on the clustering algorithm, the geographical location, soil type and water demand of the monitoring points to be irrigated are clustered to obtain the multiple irrigation groups; The water allocation includes: determining the allocated water volume for the multiple irrigation groups according to a linear programming method.

9. The method according to claim 1, characterized in that, The water supply capacity data includes real-time water supply volume and channel distribution capacity. The method further includes: when the total allocated water volume corresponding to the irrigation scheduling scheme exceeds the real-time water supply, or when the planned water supply corresponding to at least one of the irrigation time windows exceeds the channel distribution limit, the irrigation scheduling scheme is iteratively corrected according to the irrigation priority, the water stress level, and the probability of water stress occurrence to obtain an executable irrigation scheduling scheme.

10. A deep learning-based spring maize drought identification and diagnosis system, used to implement the deep learning-based spring maize drought identification and diagnosis method according to any one of claims 1-9, characterized in that, The system includes: The data acquisition module is used to acquire time-series data of soil moisture, crop growth index data, real-time meteorological data, forecast meteorological data, and water supply capacity data at each monitoring point. The status recognition module is used to perform time-series fusion of the soil moisture time-series data, the crop growth index data and the real-time meteorological data according to the deep learning time-series model to determine the drought time-series status of each monitoring point. The stress determination module is used to determine the water stress level and water stress occurrence probability of each monitoring point based on the drought time sequence status, the forecast meteorological data, and the water supply capacity data. The irrigation determination module is used to determine the corresponding monitoring point as the monitoring point to be irrigated when the water stress level reaches a preset level and the probability of water stress occurrence exceeds a preset threshold, and to determine the irrigation time window and water demand of the monitoring point to be irrigated. The scheduling generation module is used to determine irrigation priority and allocate water volume based on the irrigation time window, water demand, water stress level, water stress occurrence probability and water supply capacity data of the monitoring points to be irrigated, and generate an irrigation scheduling plan.

Images