A deep learning-based drought early warning method for coated soybeans

Abstract

The present application relates to the field of agricultural informatization and intelligent irrigation early warning technology, in particular to a mulched soybean drought early warning method based on deep learning, comprising: acquiring soil moisture, soil water potential, node operation data, meteorological data containing reference evapotranspiration, crown layer image and thermal infrared image, combining plot parameters and weather forecast data for time alignment; determining root zone water deficit index; obtaining crown layer stress index; determining data reliability according to node voltage, signal strength, packet loss rate and data delay; inputting root zone water deficit index, crown layer stress index, data reliability and growth period, soil type into a drought condition evaluation model, outputting a comprehensive drought risk value, comparing and adjusting the threshold value to determine the drought grade, and triggering supplementary sampling when the reliability is low, and outputting irrigation suggestions when the heavy drought threshold is reached and the future rainfall is less than the effective rainfall threshold; the present application improves the recognition accuracy of actual drought changes.

Description

Technical Field

[0001] This invention relates to the field of agricultural informatization and intelligent irrigation early warning technology, specifically a drought early warning method for mulched soybeans based on deep learning. Background Technology

[0002] In the daily production of soybeans under plastic film mulch, soil moisture content, canopy stress, and meteorological conditions will continuously change with the growth period and plot conditions. Planting management usually needs to judge the degree of drought in a timely manner based on the field moisture status in order to avoid the impact of drought on soybean growth and yield after it develops. Under plastic film mulch, changes in film reflection, soil moisture content differences, and field communication environment will affect the accuracy and continuity of drought judgment. In traditional methods, drought warnings in the field often rely on a single soil moisture index, fixed thresholds, or human experience for judgment. The data sources are relatively limited, and the adaptability to image interference, node anomalies, weak or broken networks under mulching backgrounds, as well as differences caused by different growth stages and soil types, remains low. When data is missing, reliability decreases, or environmental conditions fluctuate, the stability of the warning results is insufficient, and irrigation decisions are delayed. Summary of the Invention

[0003] The purpose of this invention is to provide a deep learning-based drought early warning method for mulched soybeans, avoiding interference from film reflection and changes in the field communication environment on early warning data, so as to achieve continuous and accurate drought assessment and timely irrigation decisions. This invention can be achieved through the following technical solutions: Soil moisture, soil water potential, and node operation data are acquired through IoT sensor nodes; Meteorological data, including reference evapotranspiration, were obtained through weather stations; canopy images and thermal infrared images were acquired through drones. Acquire plot parameters and weather forecast data including growth period and soil type, and perform time alignment; The root zone water shortage index is determined based on soil moisture, soil water potential, and reference evapotranspiration. Based on a preset reflectance threshold, the mulch film pixels in the canopy image are identified, the vegetation index of the canopy image and the thermal features of the thermal infrared image are extracted, and the vegetation index and thermal features are corrected according to the mulch film pixels to obtain the canopy stress index. Determine data reliability based on node operation data; The root zone water shortage index, canopy stress index, data reliability, and growth period and soil type in the plot parameters are input into the pre-constructed drought assessment model, and the comprehensive drought risk value is output. The overall drought risk value is compared with the adjustment thresholds at all levels to determine the drought level. The adjustment thresholds at all levels include the severe drought threshold. If the data reliability is lower than the preset reliability benchmark threshold, the field execution unit will be controlled to re-collect the data; If the comprehensive drought risk value is greater than or equal to the severe drought threshold, and the future rainfall is judged to be less than the preset effective rainfall threshold based on weather forecast data, then irrigation recommendations will be output.

[0004] As a preferred option, this includes: establishing plot parameters, which at least include plot number, soil type, mulch film type, sensor burial depth, growth period calendar, historical irrigation records, historical drought records, and basic early warning thresholds including mild drought threshold, moderate drought threshold, and severe drought threshold.

[0005] As a preferred embodiment, soil moisture includes shallow soil moisture and deep soil moisture; the root zone water shortage index is obtained by weighting shallow soil moisture, deep soil moisture, soil water potential and reference evapotranspiration. Node operation data includes node voltage, signal strength, packet loss rate, and data latency. Data reliability is determined based on node voltage, signal strength, packet loss rate, and data latency.

[0006] As a preferred method, vegetation indices and thermal characteristics are corrected based on the mulch film pixels to obtain the canopy stress index, including: The average reflectance of the mulch film pixels and the actual canopy coverage of the canopy image are obtained. Based on the average reflectance and actual canopy coverage, the vegetation index and thermal characteristics are corrected to obtain the canopy stress index.

[0007] As a preferred option, the drought assessment model is a deep learning model based on neural networks, including an input layer, a hidden layer, and an output layer; The input layer receives a multidimensional time-series feature matrix composed of root zone water shortage index, canopy stress index, data reliability, reference evapotranspiration, growth period, and soil type; The hidden layer extracts features and models temporal dependencies from the multidimensional temporal feature matrix through a long short-term memory network to obtain hidden state features. The output layer maps the hidden state features through fully connected layers and activation functions to output a comprehensive drought risk value.

[0008] As a preferred approach, the drought assessment model is constructed in the following order: Acquire historical feature data and divide it into modeling data and validation data; Using historical drought level labels as a benchmark, the error between the predicted risk value output by the deep learning model and the benchmark is calculated; The backpropagation algorithm is used to iteratively adjust the model network weights based on the error. Training stops when the validation set loss no longer decreases for several consecutive rounds, thus obtaining the pre-constructed drought assessment model.

[0009] As a preferred option, the basic early warning threshold is modified according to the growth period and soil type to obtain the adjustment thresholds at each level. The modification items are derived from the preset growth period modification table and soil type modification table. The adjustment thresholds at each level include the threshold for mild drought, the threshold for moderate drought, and the threshold for severe drought, with the threshold for mild drought being lower than the threshold for moderate drought, and the threshold for moderate drought being lower than the threshold for severe drought. If the overall drought risk value is less than the mild drought threshold, the drought level is determined to be normal. If the comprehensive drought risk value is greater than or equal to the mild drought threshold and less than the moderate drought threshold, the drought level is determined to be mild drought. If the comprehensive drought risk value is greater than or equal to the moderate drought threshold and less than the severe drought threshold, the drought level is determined to be moderate drought. If the comprehensive drought risk value is greater than or equal to the severe drought threshold, the drought level is determined to be severe drought.

[0010] As a preferred option, the sampling interval is adjusted according to the drought severity level; When the signal strength is less than or equal to the preset signal threshold, it is determined to be a weak network condition; when the packet loss rate is greater than or equal to the preset packet loss threshold, it is determined to be a network outage condition. Under weak network conditions or network outage conditions, the on-site execution unit issues an alarm to perform local early warning, stores data locally, and performs cached retransmission after the network is restored; When the data credibility is lower than the preset credibility benchmark threshold, the features input to the drought assessment model are dynamically weighted.

[0011] The technical solution disclosed herein has the following beneficial effects: 1. This invention acquires soil moisture, soil water potential, and node operation data through IoT sensor nodes, meteorological data including reference evapotranspiration through weather stations, and canopy images and thermal infrared images through drones. It also combines plot parameters and weather forecast data for time alignment, enabling comprehensive analysis of soil water supply status, canopy stress status, meteorological water consumption demand, and subsequent rainfall conditions in the same periodic record. This improves the data integrity and comprehensiveness of drought assessment in mulched soybean fields, avoiding the one-sidedness caused by traditional single soil moisture indicators or manual experience judgment. 2. This invention determines the root zone water shortage index based on shallow soil moisture, deep soil moisture, soil water potential, and reference evapotranspiration. After performing dimensionless mapping on different physical quantities according to the direction of water shortage, it is weighted, which can simultaneously reflect the surface moisture under the film, the root zone water supply status, water availability, and stage water consumption demand. It solves the problem of inconsistency between surface water content and actual water shortage in the root zone under film covering conditions. The root zone water supply status is still the dominant judgment factor, and evapotranspiration demand is the driving correction factor, which jointly participates in the judgment, thereby improving the accuracy of identifying actual drought changes. Attached Figure Description

[0012] The above and other objects, features and advantages of the present invention will become clearer from the following description of embodiments of the invention with reference to the accompanying drawings, in which: Figure 1 This is a flowchart illustrating a deep learning-based drought early warning method for mulched soybeans, provided as an embodiment of this application. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0014] Please see Figure 1 A deep learning-based drought early warning method for mulched soybeans includes: acquiring soil moisture, soil water potential, and node operation data through IoT sensor nodes; Meteorological data, including reference evapotranspiration, are obtained from weather stations; Acquire canopy and thermal infrared images using drones; obtain plot parameters and weather forecast data including growth period and soil type, and perform time alignment; The root zone water shortage index is determined based on soil moisture, soil water potential, and reference evapotranspiration. Based on a preset reflectance threshold, the mulch film pixels in the canopy image are identified, the vegetation index of the canopy image and the thermal features of the thermal infrared image are extracted, and the vegetation index and thermal features are corrected according to the mulch film pixels to obtain the canopy stress index. Determine data reliability based on node operation data; The root zone water shortage index, canopy stress index, data reliability, and growth period and soil type parameters in the plot are input into a pre-constructed drought assessment model, which outputs a comprehensive drought risk value.

[0015] The overall drought risk value is compared with the adjustment thresholds at all levels to determine the drought level. The adjustment thresholds at all levels include the severe drought threshold. If the data reliability is lower than the preset reliability benchmark threshold, the field execution unit will be controlled to re-collect the data; If the comprehensive drought risk value is greater than or equal to the severe drought threshold, and the future rainfall is judged to be less than the preset effective rainfall threshold based on weather forecast data, then irrigation recommendations will be output. This includes: establishing plot parameters, which at least include plot number, soil type, mulch film type, sensor burial depth, growth period calendar, historical irrigation records, historical drought records, and basic early warning thresholds including mild drought threshold, moderate drought threshold, and severe drought threshold; Soil moisture includes shallow soil moisture and deep soil moisture; the root zone water shortage index is obtained by weighting shallow soil moisture, deep soil moisture, soil water potential and reference evapotranspiration. Node operation data includes node voltage, signal strength, packet loss rate, and data latency. The reliability of the data is determined based on node voltage, signal strength, packet loss rate, and data latency. The canopy stress index is obtained by correcting vegetation indices and thermal characteristics based on the mulch film pixels, including: obtaining the average reflectance of mulch film pixels and the actual canopy coverage of the canopy image; and correcting vegetation indices and thermal characteristics based on the average reflectance and actual canopy coverage to obtain the canopy stress index. This method is deployed in a field early warning system for soybean fields covered with plastic film. One side of the system is connected to IoT sensor nodes buried in the field, and the other side is connected to a weather station, a drone data collection terminal, and a management terminal for receiving weather forecast data. All types of data collection results are first collected by plot number, and then written into the data record of the same plot according to the system cycle, so as to ensure that the data from all sources are under the same judgment criteria when the drought is assessed in the future. During the plot establishment phase, a basic parameter table is first generated, which is associated with plot number, soil type, mulch film type, sensor burial depth, growth period calendar, historical irrigation records, historical drought records, and basic early warning thresholds. The basic early warning thresholds include at least the light drought threshold, the moderate drought threshold, and the severe drought threshold; this parameter table is not generated repeatedly in subsequent cycles, but is called as a fixed configuration item; the fertility calendar is used to determine the fertility period of the current cycle, historical drought records are used for labeling during the model building phase, and the mulch film type and sensor burial depth are used to explain the source of differences in different collected values ​​within the same plot; Once the warning period begins, IoT sensor nodes report soil moisture, soil water potential, and node operation data. Meteorological stations simultaneously report meteorological data including reference evapotranspiration. Drones collect canopy images and thermal infrared images, and the management terminal writes the corresponding period's weather forecast data. When aligning data from various sources by time, the system uses the plot records corresponding to the current period as the merging unit. If the reception time of a certain source data exceeds the preset time window, the records that have arrived in the current period are retained first, and the evaluation of the current period is completed after the missing sources are filled. If a certain image or meteorological item is not obtained before the end of the current period, the current period is not directly terminated, but the obtained soil and node data are retained for the purpose of performing credibility judgment and supplementary data collection control. For cycles lacking canopy images, thermal infrared images, reference evapotranspiration, or weather forecast data, which are insufficient to complete a full drought assessment, the system only performs the aforementioned record retention, reliability judgment, and supplementary data collection control. It does not directly input the cycle into the drought assessment model, nor does it perform drought level comparison in advance when the data is missing. The risk assessment is performed on the cycle only after the missing items are filled in and the time alignment is completed. This avoids the interruption of the early warning process for the entire plot due to the interruption of a single link, and also avoids the formation of logically unclosed assessment results when the input items are incomplete. The root zone water shortage index is determined by combining shallow soil moisture, deep soil moisture, soil water potential and reference evapotranspiration within the same cycle. Shallow soil moisture reflects the surface moisture under the film, deep soil moisture reflects the water supply status of the root zone, soil water potential is used to supplement water availability information, and reference evapotranspiration is used to characterize the water demand of the cycle. During the calculation, since the physical dimensions of shallow and deep soil moisture, soil water potential and reference evapotranspiration are different, the system performs dimensionless mapping on each parameter according to the direction of water shortage before participating in the weighting. The system prioritizes using historical statistical ranges for the current plot and current growth period for range normalization: for shallow and deep soil moisture, mapping is performed according to the direction where higher water content results in lower water deficit contribution. Specifically, the mapping results for shallow soil moisture are used. For example, the difference between the historical maximum shallow soil moisture and the current shallow soil moisture is divided by the difference between the historical maximum shallow soil moisture and the historical minimum shallow soil moisture. The mapping method for deep soil moisture is the same as that for shallow soil moisture. For soil water potential, the mapping is carried out in the direction that the larger the absolute value of water potential, the greater the contribution of water shortage. Specifically, the difference between the current absolute value of soil water potential and the historical minimum absolute value of water potential is divided by the difference between the historical maximum absolute value of water potential and the historical minimum absolute value of water potential. For the reference evapotranspiration, the mapping is performed in the direction that the higher the evapotranspiration demand, the higher the contribution of water shortage. That is, the difference between the current reference evapotranspiration and the historical minimum reference evapotranspiration is divided by the range between the historical maximum and minimum reference evapotranspiration. After mapping, all parameters fall into the [0,1] interval. When the historical maximum and minimum values ​​of a certain parameter are the same in the corresponding plot and the corresponding growth period, or when the sample size of the parameter is insufficient to form a stable extreme value range, the system normalizes the parameter using a preset default interval and marks the parameter as a low-discrimination item to avoid normalization failure. The root zone water deficit index is calculated using a weighted method, and the specific formula is as follows: in, , , , These are shallow soil moisture, deep soil moisture, soil water potential, and reference evapotranspiration, respectively, mapped according to the aforementioned water-deficient direction. The weighting table is pre-configured according to land parcel type, and its configuration values ​​are determined by principal component contribution rate analysis of historical measured water consumption data of the same type of land parcel. Among them, the weight of shallow soil water is... Weight of deep soil moisture The sum is greater than the soil water potential weight. ,and Not less than the reference evaporation weight The sum of all feature weights is 1, so that the root zone water supply status is still the dominant judgment factor, and evapotranspiration demand is used as a driving correction factor to participate in the judgment. The index is written into the current period's plot record and used as an input to the drought assessment model. If either shallow or deep soil moisture is missing in the current period, the system retains the other data for this period and marks the current plot record as pending re-collection so that it can be re-collected by the field execution unit. The canopy stress index is used to reduce the impact of mulch exposure on image determination. The system first performs mulch pixel identification on the canopy image, based on the reflectance threshold. When the reflectance of a pixel in the canopy image is greater than the preset reflectance threshold of 0.15, it is determined to be a mulch pixel. For pixels identified as plastic film, they are not directly included in the calculation of the uncorrected vegetation index. Instead, the average reflectance intensity and actual canopy coverage of the plastic film area are obtained first, and then the vegetation index and thermal characteristics are corrected accordingly. The vegetation index extracted from the canopy image is the uncorrected vegetation index. This uncorrected vegetation index uses an index in the canopy image that can characterize vegetation activity and moisture changes, such as the normalized difference vegetation index. The thermal features are taken from the features related to canopy temperature in the thermal infrared image, such as the difference between canopy temperature and ambient temperature. To avoid direct subtraction of features with different dimensions, the system first maps the uncorrected vegetation index and thermal characteristics to their respective dimensionless intervals, and then performs background correction for the mulch film. The mapping is performed using a maximum-minimum normalization method, and the calculation formula is as follows: in, The values ​​are current values ​​without correction for vegetation indices or thermal characteristics. This represents the maximum value of the corresponding feature in historical datasets of the same type. This is the minimum value of the corresponding feature in the historical dataset of the same type; after mapping, the feature values ​​all fall within the range of [0,1]; if the extreme values ​​of the feature are equal, resulting in a denominator of zero, then the output is a centered default value of 0.5 to ensure that the system does not cause a division by zero exception during operation; The specific correction logic is as follows: Construct a background influence factor for the mulch film, which is equal to the average reflectance of the mulch film area divided by the actual canopy coverage. When the actual canopy coverage is lower than the preset minimum coverage, for example, 5%, the minimum coverage is used to replace the actual canopy coverage in the calculation to avoid abnormal amplification of the correction amount or division-by-zero overflow when the coverage is close to zero. The corrected vegetation index is obtained by subtracting the product of the factor and the background stripping coefficient from the dimensionless uncorrected vegetation index. The correction of thermal characteristics is similar, but the background stripping coefficient corresponding to the thermal characteristics is used to remove the artificial increase caused by background thermal radiation. The background stripping coefficients corresponding to the vegetation index and thermal characteristics are obtained in advance by acquiring historical comparative image data of soybean fields with the same light and canopy coverage conditions in the uncovered bare soil state and the covered state, with the goal of eliminating the influence of the background increase of the mulch film. The coefficients are obtained by linear regression fitting calibration using the least squares method. To prevent overcorrection under extreme boundary conditions such as instantaneous intense light irradiation, which could lead to a reversal of the physical meaning of the features, the system adds a monotonicity physical constraint check to the calculated corrected features: If the decline in the corrected vegetation index compared to the previous period exceeds the preset spectral physical limit threshold, which is determined based on the maximum natural decay rate of similar crops under extreme conditions in history, then the degradation control logic is triggered. The instantaneous factor product result of the current period is discarded, and the average of the correction factors of the previous three periods is used for smooth replacement by calling the time sliding window. When the film exposure ratio is greater than the first preset ratio threshold or the canopy coverage is less than the second preset ratio threshold, the correction amount is increased accordingly to reduce the interference of high film surface reflectivity and bare areas on the canopy stress judgment. To ensure that the correction results are consistent with the input range of subsequent models, the corrected vegetation index and thermal features are truncated before being included in the calculation of the canopy stress index. When the correction value is less than 0, it is recorded as 0, and when the correction value is greater than 1, it is recorded as 1, so as to avoid the dimensionless features from going out of bounds due to excessive background correction. If the proportion of mulch film pixels identified in the canopy image exceeds the preset upper limit of the mulch film proportion, for example, set to 40%, the system retains the image record for this period, but marks the image result as high mulch film interference data so that it can be constrained in combination with data credibility in the subsequent model input stage; the corrected result forms the canopy stress index, which is stored in the current period's plot record along with the root zone water shortage index. Data reliability is determined by node operation data; node operation data includes node voltage, signal strength, packet loss rate, and data latency; the system makes a comprehensive judgment on the above four types of operation items: node voltage reflects the node's continuous operation capability, signal strength reflects the stability of the communication link, packet loss rate reflects the integrity of the uploaded data, and data latency reflects the deviation between the current period and the actual sampling. The system pre-sets a baseline range and allowable deviation range for the above four indicators, and applies corresponding penalty weights to indicators that deviate from the normal range; for example, the node voltage is set to a maximum score of 25 points, the baseline range is no points deducted if it is greater than 3.6V, 5 points are deducted for every 0.1V decrease in the 3.2V to 3.6V range, and 25 points are deducted if it is lower than 3.2V. Signal strength is scored out of 25 points, with a baseline of no deduction for a value greater than -80dBm and a deduction of 5 points for every 5dBm decrease; packet loss rate is scored out of 25 points, with a baseline of no deduction for a value less than 5% and a deduction of 5 points for every 5% increase; data delay is scored out of 25 points, with a baseline of no deduction for a value less than 2 seconds and a deduction of 5 points for every 1 second of timeout. Each individual score is truncated with the corresponding full score as the upper limit and 0 as the lower limit. That is, once the cumulative deduction of a single item reaches or exceeds the full score of that item, it will not continue to extend to a negative value. The full scores of each indicator are summed after deducting the penalty scores and divided by the total full score to obtain the data reliability quantitative score in the range of [0,1]. When displayed or recorded as a percentage, it is then converted to the range of [0,100%]. After comprehensive evaluation, the data credibility is obtained and this credibility is added to the various sensor data records in the same period. A quantitative conclusion on the availability of the current data is given before model evaluation. The root zone water shortage index, canopy stress index, data reliability, and growth period and soil type in the plot parameters are fed into a pre-constructed drought assessment model, which outputs a comprehensive drought risk value. After the model outputs, the system calls the adjustment thresholds at each level corresponding to the current plot to compare and determine the drought level. The adjustment threshold is obtained by correcting the basic early warning threshold under specific growth period and soil type conditions. The same comprehensive drought risk value corresponds to different levels under different plots and different growth periods. The level determination result is written into the plot status record in the current cycle and made available for reading by the field execution unit. Furthermore, the reference evapotranspiration, which is periodically aligned with the root zone water shortage index, is retained as an independent meteorological feature in the continuous periodic record of the plot for use in the drought assessment model for time-series modeling. The reference evapotranspiration is not substituted for the root zone water shortage index. In cases of low confidence, this implementation does not directly output a high-level warning based on the results of the cycle. Instead, it first compares the data confidence with a preset confidence benchmark threshold. If the data is lower than the threshold, the system controls the field execution unit to re-collect the data. When re-collecting data, priority is given to the nodes corresponding to missing or abnormal items to reduce duplicate sampling. If the field execution unit replenishes valid data in subsequent cycles, the system will overwrite the original abnormal records with the replenished data and re-execute the calculation of the root zone water shortage index, canopy stress index and comprehensive drought risk value for that cycle. When the comprehensive drought risk value reaches the severe drought threshold, the system also needs to combine weather forecast data to determine whether the future rainfall is less than the preset effective rainfall threshold; here the severe drought threshold is the adjusted severe drought threshold for the current plot to participate in the level comparison in the current cycle. Irrigation recommendations are only output when both conditions are met simultaneously, avoiding the output of irrigation instructions when there is already an effective expectation of rainfall in the short term; if weather forecast data is missing in the current cycle, the system retains the severe drought judgment result, but the output of irrigation recommendations is temporarily suspended until the forecast data is available before the judgment of this branch is completed. In this implementation, plot parameters, periodic acquisition results, corrected image features, confidence results, comprehensive drought risk values, and grade determination results are stored in continuous periodic records of the same plot. Subsequent model calls, threshold corrections, re-acquisitions, and irrigation suggestion outputs all directly read from this record, without the need to trace back to each original acquisition end to reassemble the data.

[0016] Specifically, the drought assessment model is a deep learning model based on neural networks, including an input layer, a hidden layer, and an output layer; The input layer receives a multidimensional time-series feature matrix composed of root zone water shortage index, canopy stress index, data reliability, reference evapotranspiration, growth period, and soil type; The hidden layer extracts features and models temporal dependencies from the multidimensional temporal feature matrix through a long short-term memory network to obtain hidden state features. The output layer maps the hidden state features through a fully connected layer and an activation function to output a comprehensive drought risk value; The drought assessment model is constructed in the following order: historical characteristic data is acquired and divided into modeling data and validation data; Using historical drought level labels as a benchmark, the error between the predicted risk value output by the deep learning model and the benchmark is calculated; The backpropagation algorithm is used to iteratively adjust the model network weights based on the error. Training stops when the validation set loss no longer decreases for several consecutive rounds, thus obtaining the pre-constructed drought assessment model.

[0017] Before the early warning system for the same mulched soybean field is put into operation, a drought assessment model is first constructed using the historical records accumulated in the field and similar fields. After the model is completed, the model parameters are deployed to the field early warning terminal for automatic inference in subsequent cycles. A deep learning model based on neural networks was adopted, and features recorded continuously over multiple periods were organized in time sequence and fed into the model for drought assessment. The input layer of the model receives a multidimensional time-series feature vector consisting of root zone water shortage index, canopy stress index, data reliability, reference evapotranspiration, growth period, and soil type. The multidimensional time-series feature vector is a vector formed by arranging records formed in multiple consecutive system cycles of the same plot in chronological order. The root zone water shortage index reflects the soil profile water supply status, the canopy stress index reflects the canopy response to water shortage, data reliability is used to limit the impact of outlier data on model output, reference evapotranspiration reflects water demand under meteorological conditions, and growth period and soil type respectively give differences in crop stage and plot water holding capacity. The above features are encoded in the input layer and then sent to the hidden layer. The reference evapotranspiration is an independent input item that is aligned with the root zone water shortage index in the same period. It has already participated in the calculation of the root zone water shortage index in the preprocessing stage, but it is still retained as a separate meteorological feature in the model input stage to characterize the change in water demand in the continuous period. Therefore, the reference evapotranspiration and the root zone water shortage index are parallel feature relationships in the model rather than a redefined relationship. To avoid the direct mixing of features with different dimensions leading to the dominance of a certain type of feature value, the input layer performs interval mapping on continuous features before encoding, and performs one-hot encoding on growth period and soil type. Specifically, four continuous features—root zone water shortage index, canopy stress index, data reliability, and reference evapotranspiration—are mapped to the [0,1] interval. The growth period is individually thermally encoded into a 6-dimensional vector according to the seedling stage, branching stage, flowering stage, pod-setting stage, grain-filling stage, and maturity stage. Soil type is individually thermally encoded into a 3-dimensional vector according to sandy soil, loam, and clay. The mapped continuous features are then concatenated with the individually thermally encoded features to form a 13-dimensional single-step feature vector. According to a preset historical time step, for example, extracting six consecutive historical cycles and the current cycle, i.e., a time step of 7, the vector is assembled into a 7-row, 13-column multi-dimensional time-series feature matrix. Furthermore, when there are supplementary sampling markers or image high film surface interference markers in a certain period, the markers are not directly concatenated into the multidimensional temporal feature vector as new input dimensions, but are only used as preprocessing constraint information to adjust the interpretation or scaling coefficients of the aforementioned six types of input features in that period, so as to keep the input dimension of 13 dimensions consistent between the training stage and the deployment stage, and not change the order of that period in the sequence. The hidden layer uses a long short-term memory network with 64 neurons to extract features from the multidimensional temporal feature matrix and model temporal dependencies, and combines the changing trends of previous and subsequent cycles to form hidden state features; For example, when the shallow soil moisture remains above the preset moisture threshold due to the influence of the mulch film while the deep soil moisture continues to decline, the Long Short-Term Memory Network can retain the information on the previous root zone water shortage accumulation, so that the output is closer to the actual drought situation; the hidden state features calculated by the model are then fed into the output layer. The output layer performs a nonlinear mapping on the feature vector of the last time step in the hidden state feature sequence through a fully connected layer and a sigmoid activation function, and outputs a comprehensive drought risk value in the continuous interval [0,1]. The risk value is limited to a unified risk expression range, which facilitates direct comparison with the adjustment thresholds corresponding to mild, moderate, and severe drought. After the model outputs, the comprehensive drought risk value is written back to the current period record of the plot, and then read by the subsequent level determination and rule execution module. When building the model, historical feature data is collected first. The historical feature data comes from existing periodic records and includes at least the root zone water shortage index, canopy stress index, data reliability, reference evapotranspiration, growth period and soil type, as well as the corresponding historical drought level labels. The sorted samples are divided into modeling data and validation data according to a predetermined method. The modeling data is used for parameter learning, and the validation data is used to monitor the model's adaptation to samples that did not participate in the current round of parameter updates during the training process. During the training phase, in order to solve the formal matching problem between discrete labels and continuous predicted values, the system maps historical drought level labels to corresponding numerical benchmarks and constructs a dynamic weighted loss function based on data credibility to calculate the error. Specifically, the data reliability score corresponding to each sample period is used as the sample penalty weight and multiplied into the mean squared error. For example, the calculation formula is: in, The total number of samples, For sample number, For the first The data credibility score for each sample For the first The numerical benchmark for a sample For the first The predicted risk value of each sample is calculated, and the adaptive error between the predicted risk value and the numerical baseline is calculated, so that low confidence samples only provide a limited gradient contribution during backpropagation. After the error is formed, the model network weights are iteratively adjusted using the backpropagation algorithm combined with the adaptive moment estimation optimizer, where the initial learning rate is set. The learning rate is set to 0.001, and an exponential decay strategy is used for learning rate decay. The specific learning rate update formula is as follows: in, For the first Learning rate and decay coefficient for each iteration Set it to 0.95 to ensure convergence stability; After each iteration, the system calculates the validation set loss on the validation data. If the validation set loss does not decrease for 10 consecutive iterations, the early stopping mechanism is triggered to stop training, and the network weights at this time are retained as the pre-built drought assessment model. To ensure consistency between the training and deployment phases, a fixed-order monotonic mapping rule is used when mapping historical drought level labels to numerical benchmarks. This ensures that normal, mild drought, moderate drought, and severe drought correspond to specific quantitative benchmarks, such as 0.1, 0.4, 0.7, and 1.0 respectively. This ensures that the numerical benchmarks increase in order of level, and that the same mapping relationship is always used for the same batch of training samples, without changing due to changes in training rounds. To ensure the continuity of the model training process, if there are missing tests or abnormal collection periods in the historical samples, the samples will not be used directly as normal training samples, but will be screened in combination with data reliability. For records whose data credibility is lower than the preset truncation threshold (e.g., 60%) and whose data credibility in each of the two adjacent periods is also lower than the truncation threshold, the records are deemed to be unable to confirm their authenticity from the adjacent periods, and the samples are removed and not included in the modeling data. For records with only partial feature loss but clear labels, their temporal position relationship is retained, and the source of the loss is marked when the features are organized to avoid misalignment of the training sequence. The missing source label is only used for sample selection and preprocessing constraints, and does not change the input layer feature category. After the above processing, the training samples are consistent with the real running conditions in the deployment stage, and the model is more stable in adapting to weak network, supplementary sampling and image missing scenarios in actual applications. After the model is trained and deployed, the field early warning terminal only needs to input the continuous time series features corresponding to the current plot into the model in each cycle to obtain the comprehensive drought risk value. The model itself does not change the original data collection process, nor does it change the execution logic of subsequent threshold judgment, supplementary data collection control and irrigation suggestions. It only provides a unified risk output interface for the same mulched soybean field early warning system to continuously call.

[0018] Specifically, the basic early warning thresholds are corrected according to the growth period and soil type to obtain the adjustment thresholds at each level. The correction items are derived from the preset growth period correction table and soil type correction table. The adjustment thresholds at each level include the threshold for mild drought, the threshold for moderate drought, and the threshold for severe drought, with the threshold for mild drought being lower than the threshold for moderate drought, and the threshold for moderate drought being lower than the threshold for severe drought. If the overall drought risk value is less than the mild drought threshold, the drought level is determined to be normal. If the comprehensive drought risk value is greater than or equal to the mild drought threshold and less than the moderate drought threshold, the drought level is determined to be mild drought. If the comprehensive drought risk value is greater than or equal to the moderate drought threshold and less than the severe drought threshold, the drought level is determined to be moderate drought. If the comprehensive drought risk value is greater than or equal to the severe drought threshold, the drought level is determined to be severe drought; The sampling interval is adjusted according to the drought level; when the signal strength is less than or equal to the preset signal threshold, it is determined to be a weak network condition; when the packet loss rate is greater than or equal to the preset packet loss threshold, it is determined to be a network outage condition. Under weak network conditions or network outage conditions, the on-site execution unit issues an alarm to perform local early warning, stores data locally, and performs cached retransmission after the network is restored; When the data credibility is lower than the preset credibility benchmark threshold, the features input to the drought assessment model are dynamically weighted.

[0019] After obtaining the comprehensive drought risk value, the early warning system for the same mulched soybean field will perform subsequent judgments and controls based on the current growth stage, soil type, and communication status, and will adjust the basic early warning threshold according to the growth stage and soil type to obtain the adjustment thresholds at each level. The system first reads the basic early warning threshold from the plot parameter table, and then queries the preset growth period correction table and soil type correction table according to the growth period and soil type corresponding to the current cycle. The growth period correction table and soil type correction table are pre-calibrated and established based on the measured water demand sensitivity coefficient of soybeans at each growth period in historical similar plots, as well as the statistical analysis of the differences in the physical water holding capacity of different soils. The specific calculation logic for the two types of corrections is as follows: Adjustment threshold = basic warning threshold + growth period correction value + soil type correction value; for example, for soybeans during the water-sensitive flowering and pod-setting stage, the growth period correction value is set to -0.05, and for the seedling stage where the drought resistance index is higher than the preset benchmark, it is set to +0.05; for sandy soil with a water holding capacity lower than the preset lower limit, the soil type correction value is set to -0.05, and for clay soil with a water holding capacity higher than the preset upper limit, it is set to +0.05. By lowering the threshold, high-level warnings are triggered earlier for sensitive periods or plots with poor water retention. After the two types of correction terms are superimposed, the light drought threshold, moderate drought threshold and severe drought threshold for the current plot in the current cycle are obtained. The three always satisfy the order relationship that the light drought threshold is less than the moderate drought threshold and the moderate drought threshold is less than the severe drought threshold. If a threshold order conflict occurs after correction, the level will not be directly output in this cycle. Instead, it will revert to the basic warning threshold for re-determination, and the correction table call result for that plot will be marked as pending verification to avoid the impact of incorrect configuration on on-site control. The light drought threshold, moderate drought threshold, and severe drought threshold involved in subsequent processes all refer to the adjusted thresholds of each level after correction in the current cycle. The basic warning threshold will only be temporarily used when the aforementioned order conflict occurs and the determination is reverted. After the threshold correction is completed, the system compares the comprehensive drought risk value with the adjustment thresholds at each level; if the comprehensive drought risk value is less than the mild drought threshold, the drought level is determined to be normal; if the comprehensive drought risk value is greater than or equal to the mild drought threshold and less than the moderate drought threshold, the drought level is determined to be mild drought. If the comprehensive drought risk value is greater than or equal to the moderate drought threshold and less than the severe drought threshold, the drought level is determined to be moderate drought; if the comprehensive drought risk value is greater than or equal to the severe drought threshold, the drought level is determined to be severe drought. The determination result is written into the current period's plot status record and read by the field execution unit. After this processing, the model output is decoupled from the subsequent control, and the model does not need to be retrained when the threshold is modified. In terms of sampling control, the system adjusts the sampling interval according to the drought level through a preset interval mapping table. Specifically, when the judgment result is normal, the field execution unit reads the status record and keeps the sampling interval at the basic interval, such as 6 hours. When the judgment result is mild drought, the sampling interval is shortened to 4 hours. When the judgment result is moderate drought, the sampling interval is shortened to 2 hours. When the judgment result is severe drought, the sampling interval is further shortened to 1 hour. This strategy is used to match the sampling density with the rate of change in drought conditions; if the status of a plot changes from low to high, the new sampling interval will take effect in subsequent cycles; if the same level is maintained for multiple consecutive cycles, the field execution unit will maintain the current sampling interval and will not issue the same control repeatedly. In terms of communication condition judgment, the system simultaneously reads the signal strength and packet loss rate from the node's operating data; when the signal strength is lower than the preset signal threshold, such as -85dBm, it is determined to be a weak network condition; when the packet loss rate is higher than the preset packet loss threshold, such as 20%, it is determined to be a network outage condition; both types of conditions are sensed by the field execution unit and the corresponding actions are executed. Under weak network conditions or network outage conditions, the field execution unit issues an alarm to perform local early warning, flips the communication state machine from online state to buffered state in the console, and stores the current period and subsequent data that failed to be uploaded with the timestamp generated by the edge node locally. Once the network is restored, for example, after receiving multiple cloud heartbeat response confirmation frames in succession, the field execution unit triggers the state machine to flip back to the online state and performs buffer retransmission according to the order of collection. When the cloud receives the retransmitted data, it performs timing correction compensation for possible clock drift at the edge: when the network is restored and the first heartbeat handshake is established, the time difference between the real-time time of the cloud and the current system time of the edge is independently calculated as the hardware clock drift reference amount. If the absolute value of the drift reference exceeds the clock drift tolerance window, for example, the tolerance is set to 5% of the sampling interval, then the global system-level correction is performed on the edge timestamp of each data packet in the buffer queue using the drift reference, and the data packets are rearranged according to the corrected absolute timestamp. The clock compensation logic ensures that the data axes fed into the drought assessment model are strictly aligned, and maintains continuous local early warning in the field even when communication is abnormal; if it is found that the cached data overlaps with the records already existing in the cloud after recovery, the comparison and deduplication are performed based on the corrected period identifier, and only the missing parts are retransmitted. In cases where the data credibility is lower than the preset credibility benchmark threshold, this implementation method does not simply rely on re-collection, but also dynamically weights the features input to the drought assessment model. Its processing logic is as follows: when the node's running status indicates a decrease in the reliability of the current period's data, it means that the underlying abnormal data reported by the IoT sensor node will cause the synthesized root zone water shortage index to deviate from the true state. The system automatically reduces the feature scaling factor of the root zone water shortage index in the model input stage. Meanwhile, for input features whose confidence scores, such as canopy stress index and reference evapotranspiration, obtained independently by drones and weather stations are greater than the set weight, the feature scaling factor is increased within the preset upper limit to reduce the amplification effect of a single anomaly source on the comprehensive drought risk value. The dynamic weighting here occurs during the model input preprocessing stage, without changing the network parameters of the trained model, nor the matrix dimension, category, and arrangement order of the multidimensional temporal feature vectors; The specific allocation strategy is as follows: Let the ratio of the current data reliability to the perfect score be... The specific scaling logic is as follows: in, This is the initial feature scaling factor corresponding to the root region water shortage index in the input feature vector, which is the system default value of 1. The scaling factor is the feature scaling factor after reduction; The system will allocate the reduced contribution margin according to a preset ratio, for example, 70% of the margin will be compensated to the more reliable canopy stress index and 30% to the reference evapotranspiration, and added to the initial feature scaling factor corresponding to the canopy stress index and the reference evapotranspiration. Dynamic weighting is performed according to a preset weight adjustment table. After mapping, the root zone water shortage index, canopy stress index and reference evapotranspiration are multiplied by the corresponding feature scaling coefficients and then assembled into a multi-dimensional time series feature vector and sent to the drought assessment model. Among them, the proportionality coefficient The [0,1] expression is used, consistent with the aforementioned data credibility quantification score; if the data credibility is stored as a percentage at the record layer, it is first converted to a [0,1] decimal before participating in the scaling calculation; To avoid the scaled feature magnitudes exceeding the coverage of the training phase, the scaling coefficients of each feature are limited to the preset upper and lower limits, and the compensation for the canopy stress index and reference evapotranspiration is only performed within the upper limit. When a certain compensation term reaches the upper limit, the remaining compensation amount will no longer overflow to other features, thereby ensuring the stability of the numerical range of the input matrix. To prevent the scaling of feature magnitudes during the inference stage from conflicting with the distribution of deep learning pre-training data, the same credibility-based dynamic weighting rule is applied to adjust the scale of the historical input matrix when processing historical feature data during the drought assessment model construction stage. This ensures the consistency of dimensional distribution and underlying structure throughout the entire lifecycle of model training and actual deployment. If a supplementary sampling marker or a high film surface interference marker exists in the current period, the feature corresponding to the marker will continue to be downgraded in the same round of weighting, but the feature will not be directly deleted, nor will a new input dimension be added, so as to keep the time series input length stable. This dynamic weighting is completed during the model input phase and exists in parallel with the re-acquisition branch; that is, on the one hand, it allows the current cycle to form a risk output with confidence constraints when necessary, and on the other hand, it retains the ability to re-evaluate after re-acquisition; if the re-acquisition result arrives and the confidence is restored, the subsequent cycle will resume the normal weighting process. In this implementation, the level determination, sampling interval adjustment, weak network and network outage handling, and dynamic weighting are all carried out around the same early warning scenario for mulched soybean fields; all results are uniformly written into the plot periodic record, and the field execution unit performs local early warning, cache storage, retransmission, and sampling interval switching according to the record content, so that communication status changes, threshold correction, and risk assessment can run continuously within the same process.

[0020] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based method for early warning of drought in mulched soybeans, characterized in that, include: Soil moisture, soil water potential, and node operation data are acquired through IoT sensor nodes; Meteorological data, including reference evapotranspiration, are obtained from weather stations; Acquire canopy images and thermal infrared images using drones; Acquire plot parameters and weather forecast data including growth period and soil type, and perform time alignment; The root zone water shortage index is determined based on the soil moisture, the soil water potential, and the reference evapotranspiration. Based on a preset reflectance threshold, the mulch film pixels in the canopy image are identified, the vegetation index of the canopy image and the thermal features of the thermal infrared image are extracted, and the vegetation index and the thermal features are corrected according to the mulch film pixels to obtain the canopy stress index. The reliability of the data is determined based on the node's operational data. The root zone water shortage index, the canopy stress index, the data reliability, and the growth period and soil type in the plot parameters are input into a pre-constructed drought assessment model to output a comprehensive drought risk value. The drought level is determined by comparing the comprehensive drought risk value with adjustment thresholds at various levels, including the severe drought threshold. If the reliability of the data is lower than the preset reliability benchmark threshold, the field execution unit is controlled to re-collect the data; If the comprehensive drought risk value is greater than or equal to the severe drought threshold, and the future rainfall is determined to be less than the preset effective rainfall threshold based on the weather forecast data, then an irrigation recommendation will be output.

2. The method for early warning of drought in mulched soybeans based on deep learning according to claim 1, characterized in that, include: Establish plot parameters, which include at least plot number, soil type, mulch film type, sensor burial depth, growth period calendar, historical irrigation records, historical drought records, and basic early warning thresholds including mild drought threshold, moderate drought threshold, and severe drought threshold.

3. The method for early warning of drought in mulched soybeans based on deep learning according to claim 1, characterized in that, The soil moisture includes shallow soil moisture and deep soil moisture; the root zone water shortage index is obtained by weighting the shallow soil moisture, deep soil moisture, soil water potential and reference evapotranspiration. The node operation data includes node voltage, signal strength, packet loss rate, and data latency. The reliability of the data is determined based on the node voltage, signal strength, packet loss rate, and data latency.

4. The method for early warning of drought in mulched soybeans based on deep learning according to claim 1, characterized in that, Based on the mulch film pixels, the vegetation index and the thermal characteristics are corrected to obtain the canopy stress index, including: The average reflectance of the mulch film pixels and the actual canopy coverage of the canopy image are obtained; the vegetation index and the thermal characteristics are corrected based on the average reflectance and the actual canopy coverage to obtain the canopy stress index.

5. The method for early warning of drought in mulched soybeans based on deep learning according to claim 1, characterized in that, The drought assessment model is a deep learning model based on a neural network, including an input layer, a hidden layer, and an output layer. The input layer receives a multidimensional time-series feature matrix composed of the root zone water shortage index, the canopy stress index, the data reliability, the reference evapotranspiration, the growth period, and the soil type; The hidden layer extracts features and models temporal dependencies from the multidimensional temporal feature matrix through a long short-term memory network to obtain hidden state features. The output layer maps the hidden state features through a fully connected layer and an activation function to output the comprehensive drought risk value.

6. A method for early warning of drought in mulched soybeans based on deep learning according to claim 5, characterized in that, The drought assessment model is constructed in the following order: Acquire historical feature data and divide it into modeling data and validation data; Using historical drought level labels as a benchmark, the error between the predicted risk value output by the deep learning model and the benchmark is calculated; The backpropagation algorithm is used to iteratively adjust the model network weights based on the error. Training stops when the validation set loss no longer decreases for several consecutive rounds, thus obtaining the pre-constructed drought assessment model.

7. A method for early warning of drought in mulched soybeans based on deep learning according to claim 2, characterized in that, The basic early warning threshold is corrected according to the growth period and the soil type to obtain the adjustment thresholds at each level, wherein the correction items are derived from the preset growth period correction table and soil type correction table. The adjustment thresholds at each level include a light drought threshold, a moderate drought threshold, and a severe drought threshold, wherein the light drought threshold is less than the moderate drought threshold, and the moderate drought threshold is less than the severe drought threshold; If the comprehensive drought risk value is less than the mild drought threshold, the drought level is determined to be normal. If the comprehensive drought risk value is greater than or equal to the mild drought threshold and less than the moderate drought threshold, then the drought level is determined to be mild drought. If the comprehensive drought risk value is greater than or equal to the moderate drought threshold and less than the severe drought threshold, then the drought level is determined to be moderate drought. If the comprehensive drought risk value is greater than or equal to the severe drought threshold, then the drought level is determined to be severe drought.

8. A method for early warning of drought in mulched soybeans based on deep learning according to claim 3, characterized in that, Adjust the sampling interval according to the drought severity level; When the signal strength is less than or equal to a preset signal threshold, it is determined to be a weak network condition; when the packet loss rate is greater than or equal to a preset packet loss threshold, it is determined to be a network outage condition. Under the conditions of weak network or network outage, the field execution unit issues an alarm to perform local early warning, stores data locally, and performs cached retransmission after the network is restored; When the data credibility is lower than a preset credibility benchmark threshold, the features input to the drought assessment model are dynamically weighted.

Images