Method for predicting and warning water and fertilizer demand of polygonatum multiflorum based on multi-task learning

Abstract

The present application belongs to the technical field of neural network and agricultural time series data processing, in particular to a kind of huangjiu phenology water and fertilizer demand prediction early warning method based on multi-task learning.The method obtains huangjiu phenology, meteorology, soil water content, soil nutrients and irrigation and fertilization records, forms time series samples according to time axis and carries out shared coding;According to the joint state representation, calculate the probability of phenological stage and the transition reliability, generate water task gating vector, nutrient task gating vector and public feature channel, output water demand, nutrient demand and water stress level through multi-task learning network;Calculate the consistency degree of each task gradient, decompose the conflict gradient and update the shared parameters, then combine the water gap, soil available water trend and phenological stage sensitivity to generate early warning results.The present application can reduce the task conflict and stage jump interference in the transition period of phenology, distinguish the high water demand state and continuous water stress in the stage, and make the prediction and early warning results correspond to the actual growth process of huangjiu.

Description

Technical Field

[0001] This invention belongs to the field of neural network and agricultural time series data processing technology, specifically a method for predicting and early warning the phenological water and fertilizer demand of Polygonatum odoratum based on multi-task learning. Background Technology

[0002] Current methods for predicting the phenological water and fertilizer requirements of Polygonatum typically use field phenological records, daily meteorological data, soil moisture content, soil nutrient status, and irrigation and fertilization records as basic data. A unified time series is established according to daily, ten-day, or phenological observation cycles. Missing data is interpolated to fill in the gaps, and data of different dimensions are normalized. Then, the environmental changes within the historical time window are used to predict subsequent water and nutrient requirements. Conventional implementation methods often involve constructing separate phenological stage classification models, water requirement regression models, nutrient requirement regression models, and water stress classification models. Alternatively, a shared temporal coding network can be used to extract common features, and multiple task branches can be connected after the shared network. During training, the losses of each task are summed with fixed weights, and the model parameters are updated in reverse using the prediction errors in historical samples. Phenological information is usually processed into manually labeled discrete stage labels, which are used as ordinary input features along with meteorological, soil, and water and fertilizer management data. After comparing the predicted water requirement output by the model with the soil moisture content, a water stress early warning result is generated based on the pre-set grading boundaries.

[0003] Similar multi-task prediction methods output phenological stages, water requirements, nutrient requirements, and stress levels simultaneously based on shared temporal coding. They adjust the impact of each output branch on shared parameters through task weights and use a fixed-length sliding window to extract the changing features of recent precipitation, temperature, soil moisture, and fertilization records. Some methods also input the phenological identification results from the previous time point into the current time point and restrict the reverse jump of phenological states according to a preset stage order. The water and fertilizer requirements branches use shared features and their respective local features to complete regression calculations. When the predicted water demand is continuously higher than the available soil water or the soil moisture content is lower than a set range, an early warning is generated. Although these methods can handle multiple related tasks within the same model, the shared parameters are usually updated directly based on the weighted results of the gradients of each task. The phenological stages are only used as static labels or auxiliary tasks. The water and nutrient tasks lack constraints on the scope of shared features that change with the phenological state. The phased lag effects of historical environmental variables on the growth of Polygonatum are also often compressed into the same time window.

[0004] When Polygonatum enters the phenological transition period, the patterns of transpiration consumption, root absorption, and nutrient distribution change synchronously. However, conventional multi-task models cannot dynamically divide the shared feature boundaries between water tasks, nutrient tasks, and phenological tasks based on phenological status. This causes the opposite training directions generated by different tasks to directly affect the same temporal representation. Phenological identification bias is further transmitted to the water and fertilizer demand branch through the shared network. Water and fertilizer prediction bias, in turn, interferes with the judgment of phenological stages. As a result, the model has difficulty in establishing a continuous and consistent correspondence between phenological progress and changes in water and fertilizer demand. During periods when data changes are concentrated, such as root and stem enlargement, stage alternation, or short-term weather fluctuations, normal stage-specific increases in water demand are easily calculated as water stress. The real continuous water shortage may also be masked by the feature changes generated by phenological transition, causing the prediction results and warning periods to deviate from the actual growth state of Polygonatum. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting and warning the phenological water and fertilizer demand of Polygonatum based on multi-task learning, which can solve the problems in the background art mentioned above.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for predicting and warning the phenological and water and fertilizer demand of Polygonatum based on multi-task learning includes: acquiring phenological records of Polygonatum, meteorological sequences, soil moisture status, soil nutrient status and historical irrigation and fertilization records, and forming time-series samples according to a unified time axis;

[0008] The time-series samples are subjected to shared temporal coding to obtain a joint state representation;

[0009] Based on the joint state characterization, output phenological stage probability and stage transition reliability, and generate water task gating vector, nutrient task gating vector and common feature channel accordingly.

[0010] Water requirements, nutrient requirements, and water stress levels were obtained through a multi-task learning network.

[0011] Calculate the consistency of training gradients for each task, decompose conflicting gradients, and update the shared temporal coding parameters.

[0012] Early warning results are generated by combining the aforementioned water demand, soil available water change trends, and phenological stage sensitivity.

[0013] Preferably, the formation of the time series sample includes: establishing a multi-scale time index based on phenological observation time points, and mapping the meteorological sequence, the soil moisture state, the soil nutrient state, and the historical irrigation and fertilization records to the corresponding index intervals;

[0014] For data with differences in sampling intervals, the original time offset is retained, and a time-series addition reflecting the duration, rate of change, and distance of the most recent effective observation is generated;

[0015] Candidate lag segments are constructed based on the response lag range of different phenological stages of Polygonatum. Each candidate lag segment and the current phenological observation are input into the shared time-series coding process to obtain a joint state characterization that includes short-term environmental disturbances and staged cumulative responses.

[0016] Preferably, obtaining the phenological stage probability and the stage transition reliability includes: establishing a stage transition relationship connected in the order of Polygonatum budding, leaf unfolding, reproductive growth, rhizome enlargement and aboveground decline;

[0017] The joint state representation is mapped to stage observation score and stage transition score, respectively;

[0018] The probability of the current stage is calculated based on the stage probability of the previous time point, the observation score of the current stage, and the allowed transfer direction.

[0019] Path constraints are imposed on candidate transitions that cross non-adjacent stages, and the reliability of the stage transitions is calculated based on the stage probability differences at consecutive time points, changes in environmental accumulation, and the completeness of phenological records, forming a phenological state sequence with temporal continuity.

[0020] Preferably, the generation of the water task gating vector, the nutrient task gating vector, and the common feature channel includes: conditionally combining the phenological stage probability with the joint state representation, and calculating the water-related feature selection weight and the nutrient-related feature selection weight respectively;

[0021] Features that simultaneously satisfy the consistency conditions of water task and nutrient task are written into the common feature channel, and features that only satisfy the consistency conditions of a single task are written into the corresponding task-specific channel.

[0022] Cross-leakage constraints are set for the dedicated channels for the water task and the nutrient task, so that neither dedicated channel can reconstruct the target sequence of the other task on its own.

[0023] The outputs of the three channels are input into the multi-task learning network according to the gating vector.

[0024] Preferably, the construction of the candidate lag segments includes: generating multiple historical segments with different start and end times for each type of meteorological variable, soil state variable, and irrigation and fertilization variable;

[0025] Calculate the joint correlation between the probability changes, water demand changes, and nutrient demand changes of each historical segment and the current phenological stage;

[0026] Retain historical fragments that have a consistent time pointer across at least two related tasks, and mark historical fragments that are significant only in a single task as task-specific fragments;

[0027] Based on the temporal coverage relationship between segments, duplicate information is merged to form a common lag path and a task-specific lag path. Then, the lag path is weighted and encoded according to the probability of the phenological stage.

[0028] Preferably, the reliability of the stage transition is further corrected by the consistency of water and fertilizer prediction, including: inputting the current stage probability into the water demand branch and the nutrient demand branch respectively to obtain the initial prediction sequence under the corresponding stage conditions;

[0029] Calculate the deviation between the initial prediction sequence and changes in soil moisture state, changes in soil nutrient state, and historical management responses;

[0030] When the transition of candidate stages causes a jump in water demand and nutrient demand that does not conform to the continuous response relationship, the credibility of the corresponding transition path is reduced, and the credibility difference is fed back to the stage probability calculation process.

[0031] The transfer paths verified at multiple consecutive time points are retained to generate the phenological state sequence constrained by the water and fertilizer tasks.

[0032] Preferably, the decomposition of the conflict gradient includes: calculating the gradient vectors of the phenological stage identification task, water demand prediction task, nutrient demand prediction task, and water stress determination task relative to the shared temporal coding parameters, respectively.

[0033] For any two gradient vectors, calculate the directionally consistent quantity and project the oppositely oriented portions onto the subspace formed by the corresponding task-specific channel representation;

[0034] The remaining in-direction components are constructed into a common effective gradient, and the shared temporal coding parameters are updated only with the common effective gradient;

[0035] The projected task-specific components are used to update the water task-specific channel, the nutrient task-specific channel, and the phenological state inference parameters, respectively.

[0036] Preferably, the cross-leakage constraint includes: reconstructing the water demand sequence using the dedicated channel for the water task, and reconstructing the nutrient demand sequence using the dedicated channel for the nutrient task;

[0037] After swapping the outputs of the two task-specific channels, cross-reconstruction sequences are generated respectively;

[0038] The leakage amount is formed based on the degree of fit of the cross-reconstruction sequence to the non-corresponding target sequence, and the leakage amount and the common explanatory amount of the common feature channel for the two target sequences are jointly written into the training constraints.

[0039] The coupled residuals, which are simultaneously affected by soil moisture and soil nutrient status, are encoded separately, and only the coupled residuals are allowed to participate in the output calculation of the two tasks through the common feature channel.

[0040] Preferably, the joint correction of the phenological state sequence and the lag path includes: reallocating the stage weights of the common lag path and the task-specific lag path based on the phenological state sequence constrained by the water and fertilizer task.

[0041] The stage observation scores and water and fertilizer demand sequences are recalculated using the re-encoded joint state representation.

[0042] Compare the phase transition direction, demand change direction, and historical response direction in the two calculations. When the three are inconsistent, backtrack along the time coverage interval of the lag path that caused the deviation and reduce the weight of the corresponding segment.

[0043] The iterative execution phase includes probability updates and lag path updates until two adjacent results satisfy the preset consistency convergence condition.

[0044] Preferably, the generation of the early warning result includes: calculating a water gap sequence at consecutive time points based on the water demand and soil available water obtained through training with the common effective gradient;

[0045] Based on the phenological stage probability, the water deficit sequence is assigned a stage-sensitive weight, and water-fertilizer coupling residuals in the common feature channel are combined to determine evidence of water shortage.

[0046] Remove anomalous components that can be explained by the dedicated channel for the nutrient task and abrupt components caused by low-confidence stage transitions;

[0047] The remaining water shortage evidence is accumulated according to its duration, growth direction, and the credibility of stage transitions, and the water stress level and warning period associated with the corresponding phenological stage are output.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0049] 1. By forming a mutually constrained multi-task computation process through shared temporal coding, phenological stage probabilistic inference, state gating, and conflict gradient decomposition, the phenological stage is no longer a static input parallel to other variables, but is used to control the proportion of water-related features, nutrient-related features, and common coupling features entering the corresponding task branches. The reliability of stage transition is jointly corrected by the water demand sequence, nutrient demand sequence, and continuous changes in soil state. When the phenological task and the water and fertilizer task have opposite update directions, the conflicting part is assigned to the task-specific representation, and the shared temporal coding parameters are updated only with the common components with consistent directions. This limits the spread of phenological identification errors in the shared network and avoids the water task or nutrient task from changing the common representation alone. During the phenological transition of Polygonatum, the temporal consistency of stage judgment, water demand change, and fertilizer demand change is maintained, so that the normal stage of high water demand and the continuous water shortage form different data evolution paths, reducing the confusion between the two in the early warning calculation.

[0050] 2. By using multi-scale time indexing to retain the sampling intervals and original time offsets of different data, and establishing common lag paths and task-specific lag paths according to the phenological stages of Polygonatum, the delayed responses caused by recent meteorological disturbances, soil condition changes, and historical irrigation and fertilization can be written into the corresponding task representations, avoiding the compression of the impact of different durations into the same input by a fixed time window; cross-leakage constraints are set between the water-specific channel and the nutrient-specific channel, and the coupling residuals simultaneously affected by water and nutrients are limited to the common feature channel, which can reduce the repeated interpretation of the same abnormal change by a single task; after iterative correction of the phenological state sequence and lag path, abnormal components that can be explained by nutrient factors and low-confidence stage jump components are removed, and the remaining water gap is accumulated according to the duration and direction of change, so that the warning results correspond to specific phenological stages and continuous risk intervals, reducing the warning jumps caused by short-term data fluctuations. Attached Figure Description

[0051] Figure 1 A flowchart illustrating the overall process for predicting and early warning the phenological, water, and fertilizer requirements of Polygonatum odoratum.

[0052] Figure 2 For shared timing-coded hysteresis response curves;

[0053] Figure 3 The stage transition path constraint curve diagram;

[0054] Figure 4 This is a curve diagram of the water and fertilizer gating channel;

[0055] Figure 5 The graph shows the conflict gradient projection decomposition curve.

[0056] Figure 6 Preserve score curves for candidate lagged segments;

[0057] Figure 7 For the convergence curve of the joint correction;

[0058] Figure 8 This is a graph showing the accumulation of evidence for water scarcity and the dynamic boundary curve. Detailed Implementation

[0059] In one embodiment, please refer to Figure 1 A method for predicting and warning the phenological water and fertilizer demand of Polygonatum odoratum based on multi-task learning is proposed, comprising: acquiring phenological records, meteorological sequences, soil moisture status, soil nutrient status, and historical irrigation and fertilization records of Polygonatum odoratum, and forming time-series samples according to a unified time axis; performing shared time-series coding on the time-series samples to obtain joint state representations; outputting phenological stage probabilities and stage transition confidence based on the joint state representations, and generating water task gating vectors, nutrient task gating vectors, and common feature channels accordingly; obtaining water demand, nutrient demand, and water stress level through a multi-task learning network; calculating the consistency of training gradients for each task, decomposing conflict gradients, and updating the shared time-series coding parameters; and generating warning results by combining water demand, soil available water change trends, and phenological stage sensitivity. The method uses data from the same strain or the same cultivation area that can correspond to a unified observation period as the source of a time-series sample, and allows the phenological inference results to participate in the feature selection and training update of water and fertilizer tasks, rather than just being output in parallel.

[0060] In this embodiment, the phenological record of Polygonatum includes at least the observation date, the current phenological performance, and continuous records that can confirm the stage transition. The meteorological sequence includes temperature, precipitation, air humidity, and radiation changes related to evapotranspiration and soil recharge. The soil moisture status is represented by a sequence of moisture changes available to Polygonatum roots. The soil nutrient status is represented by a sequence of nitrogen, phosphorus, potassium, or a converted comprehensive available nutrient sequence. The historical irrigation and fertilization records include the implementation time, duration, input amount, and interval between two adjacent management events. The data processing process retains the original timestamp, observation source identifier, and missing value marker. Missing positions are not directly replaced by adjacent values, but the most recent valid observation value, the time distance from the most recent valid observation, and the direction of change under the corresponding phenological stage are written simultaneously, so that the shared time sequence coding can distinguish between the true stable state and the apparent stable state formed due to sampling interruption.

[0061] A unified timeline is indexed primarily by the phenological observation cycle. Meteorological, soil, and administrative data are projected onto index intervals according to their actual occurrence times. For intervals containing multiple observations, the mean, final value, magnitude of variation, and duration are retained. For intervals without observations, missing data markers and time intervals are retained. The projected input vector is denoted as... ,in Indicates the first A unified observation point, the input vector includes the current interval state, change amount, management event code, missing information and the time offset of the above items. The shared time sequence code generates the current joint state representation by the current input, multiple candidate lag segments and the previous time point code state, so as to maintain the cumulative and continuous response of the phenological changes of Polygonatum to water and fertilizer.

[0062] ,

[0063] ;

[0064] in, Indicates the first Joint state representation at each time point, Indicates the first Joint state representation at each time point, Indicates the first Input vector at each time point, This represents a nonlinear mapping that preserves bounded outputs. This represents the current input mapping matrix. Indicates the first A mapping matrix for each lag position. The recursive matrix representing the previous joint state. This represents the bias vector. Indicates the number of candidate lag positions. Indicates the first The lagging position at time point Normalized weights, Indicates the original relevance score. Indicates the first The original relevance scores of each candidate lag position. The correlation between the current input and the lagged input is calculated. For example, during the rhizome enlargement stage, when there is continuous low rainfall and soil available water decreases, several lagged positions corresponding to continuous low rainfall will have larger values. After normalization, the corresponding historical fragments are... The contribution of isolated single-day rainfall increases, while isolated single-day rainfall retains a smaller weight because it cannot form a continuous response chain.

[0065] Please refer to Figure 2 , Figure 2 The horizontal axis represents the continuous sampling section, which progresses from the beginning to the stabilization point according to the observation order reconstructed by the unified time axis in the text; the vertical axis represents the normalized engineering quantity, which is used to uniformly display the three components of joint state response, hysteresis contribution and memory decay. Figure 2 The calculation process of generating a joint state representation by the current input, candidate lag segments, and the state at the previous time point in the corresponding shared temporal coding formula is illustrated in the figure, which shows the state response curve, lag contribution curve, and memory decay curve.

[0066] from Figure 2 As can be seen, the state response curve maintains a slow upward trend in the initial stage, and shows a significant enhancement in the transition to the high-level segment. This indicates that when Polygonatum sibiricum enters the rhizome enlargement-related interval from leaf expansion and reproductive growth, continuous low rainfall, decreased soil available water, and the interval of management events collectively increase the response intensity of the current combined state. The lag contribution curve continues to rise, indicating that historical fragments do not participate in the calculation all at once, but gradually accumulate influence across multiple observation segments. The memory decay curve smoothly declines from high to low, indicating that the direct dominance of previous states over the current output gradually gives way to recent environmental and water and fertilizer response evidence. The three curves together reflect that this method can distinguish between isolated single-day disturbances and continuous staged responses, providing stable temporal representation inputs for subsequent phenological stage probabilities and water and fertilizer task gating.

[0067] When inferring phenological stages based on joint state representation, each time point is not classified independently. Instead, a directed stage transition relationship is established between budding, leaf expansion, reproductive growth, rhizome enlargement, and aboveground decline. A reserved state is set for stages that cannot be confirmed temporarily in continuous observation. The model forms the observation score of each candidate stage based on the current joint state representation, and then forms the current stage probability by combining the stage probability of the previous time point and the allowed transition direction. The confidence of stage transition is jointly determined by the continuity of candidate paths, cumulative environmental changes, and the completeness of phenological records. When short-term meteorological fluctuations cause a sudden increase in the observation score of a certain stage but there is a lack of adjacent stage support before and after the time point, the candidate stage will not directly replace the original stage, but will be temporarily stored with a lower confidence and wait for subsequent observation verification.

[0068] ,

[0069] ;

[0070] in, Indicates the total number of phenological stages. The joint state representation corresponds to the first The observation scores for each stage, Indicates that the previous point in time belongs to The probability of each stage, Indicates from stage Transfer to stage The path permission value is set to 1 when allowing a move to an adjacent stage or a transition to a non-adjacent stage is permitted, and to 0 when crossing non-adjacent stages is prohibited. This indicates the amount of support that current environmental accumulation and phenological records provide for the transfer path. This prevents the logarithmic input from being a positive number that is zero. This represents the stage score after path constraints. Indicates phenological stages Stage score, This represents the probability of the current stage. For example, if the previous time point was in the leaf expansion stage and the current record only shows one root and stem enlargement-related feature, then the cross-stage path... The value is 0, and the current probability is still concentrated in the leaf unfolding or adjacent stages, avoiding stage jumps caused by a single abnormal observation.

[0071] Please refer to Figure 3 , Figure 3 The horizontal axis represents the continuous sampling section, the vertical axis represents the normalized engineering quantity, and the curves represent the transfer calculation results of three types of stages: reachable path, restricted path, and transfer support. Figure 3 The corresponding stage score and stage probability update formulas reflect the combined relationship between the current stage observation score and the stage probability, path permission, and cumulative environmental support at the previous time point. The reachable path curve in the legend is used to represent the probability channel that allows for retention or adjacent transfer, the restricted path curve is used to represent the path that is suppressed when crossing non-adjacent stages, and the transfer support curve is used to represent the degree of sustained support of meteorological, soil, and phenological records for candidate transfers.

[0072] from Figure 3 As can be seen, the reachable path curve rises rapidly after the turning point and remains at a high level during the accumulation and stabilization phase. This indicates that when continuous observation and environmental accumulation support phase progression, the phase probability smoothly propagates along the directed sequence of Polygonatum budding, leaf expansion, reproductive growth, rhizome enlargement, and aboveground decline. The restricted path curve remains at a low level with only slight fluctuations, indicating that jumps across non-adjacent phases are suppressed by path allowances, and a single anomalous record does not directly change the dominant phase. The transition support curve rises slowly, reflecting that the credibility of candidate phase transitions gradually increases with the increase in record completeness and continuous response evidence.

[0073] After the phenological stage probability input state gating process, the water task gating vector is used to select features related to evapotranspiration, precipitation recharge, soil available water, and irrigation response. The nutrient task gating vector is used to select features related to soil nutrient availability, fertilization history, phenological growth intensity, and water constraints on absorption conditions. The common feature channel retains the environmental accumulation and water-fertilizer coupling residuals that can simultaneously explain changes in water and nutrient requirements. The gating weights are calculated jointly by the stage probability and the joint state representation. Therefore, the weight of the same soil water content entering the water task is different in the leaf expansion stage and the root and stem enlargement stage. The weight of the same nutrient status entering the nutrient task is also different when soil water is limited and when water is suitable. The difference is determined by the continuous response relationship in the training samples rather than being fixed manually.

[0074] ,

[0075] ,

[0076] ,

[0077] ,

[0078] ,

[0079] ;

[0080] in, This represents a vector composed of the probabilities of each stage; the semicolon indicates vector concatenation. This indicates a gating mapping that restricts the output to between 0 and 1. , and These represent the gating mapping matrices for the water task, nutrient task, and common channel, respectively. , and These represent the corresponding bias vectors. , and Let these represent the water gating vector, nutrient gating vector, and common gating vector, respectively. This indicates multiplication at corresponding positions. , and These represent the characteristics of entering the water task, nutrient task, and common pathway, respectively, such as the rhizome enlargement stage. When the corresponding location is large, the dimensions related to continuous soil moisture decline are in A relatively large value was obtained, but short-term nutrient fluctuations still need to be analyzed. Determine whether it belongs to the coupling change under water limitation.

[0081] Please refer to Figure 4 , Figure 4 The horizontal axis represents the continuous sampling section, the vertical axis represents the normalized engineering quantity, and the three curves correspond to the gated outputs of the water channel, nutrient channel, and common channel, respectively. Figure 4 The task gating vector is generated by concatenating the phenological stage probabilities and joint state representations from the text, and then the joint state is distributed to the calculation processes of the water task, nutrient task, and common feature channel. The water channel curve reflects the intensity of features such as evapotranspiration, precipitation recharge, soil available water, and irrigation response entering the water demand branch; the nutrient channel curve reflects the intensity of soil nutrient availability, fertilization history, and absorption condition constraints entering the nutrient demand branch; and the common channel curve reflects the joint explanatory power of water-fertilizer coupling residuals and environmental accumulation for both tasks.

[0082] from Figure 4As can be seen, the water channel curve generally rises with the progress of the sampling section, and is higher than the nutrient channel curve in the later section, indicating that the continuous decrease in soil moisture content in the root and stem enlargement-related interval will give the water task a higher gating weight. The nutrient channel curve shows a gentle decline, indicating that when water limitation becomes the main constraint, the direct control of independent nutrient changes on the output is appropriately weakened. The common channel curve fluctuates around the median, indicating that some changes still need to be explained by the coupling relationship between water state and nutrient uptake.

[0083] After receiving water task representations, nutrient task representations, and common channel representations, the multi-task learning network outputs the demand sequence for one or more unified observation intervals in the future, the nutrient demand branch outputs the demand sequence corresponding to the same prediction interval, the phenology branch continuously outputs the stage probability, and the water stress branch receives water demand, changes in soil available water, phenological stage sensitivity, and common water-fertilizer coupling residuals to form water stress evidence. The training samples use actual subsequent water changes, nutrient changes, continuous phenological records, and time-verified stress states as supervision information. When supervision information is missing, only the task loss with observational basis is calculated, and a missing information mask is used to prevent unlabeled tasks from generating false gradients.

[0084] During joint training, the gradient directions of phenological stage identification, water demand prediction, nutrient demand prediction, and water stress determination relative to the shared temporal coding parameters are calculated separately. When the gradient directions of any two tasks are consistent, the corresponding common components are retained. When the directions are opposite, the conflicting parts are removed from the gradient of the shared parameters and transferred to the corresponding task-specific channel. This ensures that the shared temporal coding only learns the temporal patterns that can support multiple tasks simultaneously. Local anomalies in the water task will not cover the common representations that the nutrient task depends on, and short-term misjudgments of phenological stages will not directly drive the synchronous shift of all tasks. After training, the shared parameters, gating parameters, stage transition relationships, and task-specific parameters are saved, and predictions are performed using the same data processing rules.

[0085] ;

[0086] in, This indicates the number of tasks participating in the shared parameter update. Indicates the first The gradient vectors generated by each task with shared parameters This represents another task gradient vector used to determine the direction of conflict. Indicates the first The weights for each task are determined based on label completeness and training stability. This represents the inner product of two gradient vectors. It takes the value 1 when the inner product is less than zero and 0 in other cases. Indicates the length of the gradient vector. This represents the common effective gradient used to update shared temporal coding parameters. For example, when the water demand task requires an increase in a certain dimension while the phenology task requires a decrease in the same dimension, the inner product is negative, the corresponding conflicting projection is subtracted, and the exclusive components retained by each of the two tasks are transferred to their respective parameter updates.

[0087] Please refer to Figure 5 , Figure 5 The horizontal axis represents the continuous sampling segment, the vertical axis represents the normalized engineering quantity, and the curves represent the three types of gradient components: original conflict, projection sharing, and task exclusivity. Figure 5 This corresponds to the process of comparing the gradients of tasks such as phenological stage identification, water demand prediction, nutrient demand prediction, and water stress determination during joint training, and subtracting the calculation process of conflict projections when reverse updates occur. The original conflict curve represents the degree of directional conflict between different tasks on shared temporal coding parameters, the projection sharing curve represents the effective part that can still be used to update shared parameters after subtracting the reverse component, and the task-exclusive curve represents the local update amount that should be transferred to the water, nutrient, or phenological dedicated channel.

[0088] from Figure 5 As can be seen, the original conflict curve decreases overall as the sampling segment progresses, indicating that after the stage probability stabilizes and channel division, the direct constraints on shared parameters by each task gradually weaken. The projected sharing curve continues to rise, indicating that the retained unidirectional components increase, and the shared temporal coding begins to learn more common patterns that can simultaneously explain changes in phenology, moisture, and nutrients. The task-exclusive curve gradually decays from a relatively high level in the early stage, indicating that after the exclusive updates generated in the early stage due to local moisture anomalies or short-term misjudgments are allocated to the corresponding channels, the interference on shared parameters decreases.

[0089] During the inference phase, new data is continuously received according to the unified timeline during training. Each time a new joint state representation is formed, the phenological stage probability, water requirement, nutrient requirement, and stress evidence are updated. The difference between water requirement and available soil water constitutes the basic gap. Phenological stage sensitivity is used to distinguish the tolerance range of different stages to the same gap. The common water-fertilizer coupling residual is used to identify the limited nutrient absorption caused by insufficient water. If the anomaly can be explained by the nutrient-specific representation, it is deducted from the water shortage evidence. If the anomaly comes from a low-confidence stage jump, it remains in the pending confirmation state. Only when the direction of water shortage evidence is consistent at consecutive time points and the stage probability gradually stabilizes is the warning interval for the corresponding phenological stage formed.

[0090] In a preferred embodiment, the formation of the time series sample includes: establishing a multi-scale time index based on phenological observation time points, mapping meteorological sequences, soil moisture status, soil nutrient status, and historical irrigation and fertilization records to corresponding index intervals; retaining the original time offset for data with sampling interval differences, and generating a time series additional quantity reflecting the duration, rate of change, and distance of the most recent effective observation; constructing candidate lag segments based on the response lag range of different phenological stages of Polygonatum odoratum, and inputting each candidate lag segment and the current phenological observation into a shared time series coding process to obtain a joint state characterization including short-term environmental disturbances and staged cumulative responses.

[0091] Specifically, the multi-scale time index includes at least a main scale synchronized with phenological observations, a fine scale for describing continuous meteorological changes, and a long scale for describing water and fertilizer cumulative responses. When fine-scale data is mapped to the main scale, it is not directly compressed into a single mean, but rather forms interval start values, interval end values, the direction of maximum continuous change, and the duration of events. Long-scale data establishes an accumulation window according to the start of the phenological stage or the most recent management event. The duration is calculated by the number of intervals that appear consecutively in the same direction of change and the actual duration of each interval. The rate of change is obtained by dividing the difference between adjacent valid observations by the actual time interval. The distance of the most recent valid observation is obtained by the difference between the current time point and the most recent valid timestamp, so that data with different sampling frequencies still retain their true time structure.

[0092] Candidate lag segments are generated based on the time range in which each type of variable may respond in different phenological stages. The same variable can form multiple historical segments of different lengths and starting points. Within a segment, the original direction of change, cumulative amount, and relative position of management events are preserved. Segments are linked by time coverage. Short segments describe recent disturbances, while long segments describe the stage accumulation process. Shared time series coding first encodes each segment independently, then calculates segment weights based on the probability of the current phenological stage and combines them. When a new observation arrives, only the segment coding that overlaps with the newly added time interval is updated, thus maintaining the continuity of the historical path and avoiding recalculating the entire historical sequence for each inference.

[0093] In one embodiment, obtaining the phenological stage probability and stage transition reliability includes: establishing a stage transition relationship connected in the order of Polygonatum budding, leaf unfolding, reproductive growth, rhizome enlargement, and aboveground decline; mapping the joint state characterization to stage observation scores and stage transition scores respectively; calculating the current stage probability based on the stage probability of the previous time point, the current stage observation score, and the allowed transition direction; applying path constraints to candidate transitions that cross non-adjacent stages; and calculating the stage transition reliability based on the stage probability differences at consecutive time points, changes in environmental accumulation, and the completeness of phenological records, thereby forming a phenological state sequence with temporal continuity.

[0094] In this embodiment, the stage transition relationship is stored in a directed graph. Each stage node is connected to itself and the next allowed stage. During the observation discontinuity period, the current stage is allowed to be maintained and the retention confidence is gradually reduced. The stage observation score is obtained by the joint state representation through stage mapping. The stage transition score is obtained by the stage probability of the previous time point, the current environmental accumulation, and the duration since the previous stage transition. The phenological record completeness is calculated based on the adjacent observation interval, the completeness of the recorded fields, and the consistency of continuous observations. After the current stage probability is output, it is not immediately fixed into a single label, but the probability distribution is retained for use by water gating, nutrient gating, and subsequent stage correction.

[0095] The credibility of stage transitions is based on probability changes at multiple consecutive time points. When the probability of a candidate stage increases continuously and the probability of the previous stage decreases continuously, consistent transition evidence is formed. When the probability of a candidate stage increases only at a single time point and then falls back, it is considered a short-term observational disturbance. Changes in environmental accumulation are used to determine whether the candidate stage has the corresponding temperature accumulation, moisture change, or growth sequence conditions. The completeness of phenological records is used to adjust the credibility of the evidence. If phenological records are missing but the water and fertilizer demand sequence and environmental accumulation both support the transition between adjacent stages, a wider stage probability distribution is retained, and the distribution converges to a definite stage after subsequent records are entered.

[0096] In a preferred embodiment, the generation of the water task gating vector, the nutrient task gating vector, and the common feature channel includes: conditionally combining the phenological stage probability with the joint state representation, and calculating the selection weights of water-related features and nutrient-related features respectively; writing features that simultaneously satisfy the consistency conditions of the water task and the nutrient task into the common feature channel, and writing features that only satisfy the consistency conditions of a single task into the corresponding task-specific channel; setting cross-leakage constraints on the water task-specific channel and the nutrient task-specific channel, so that no single channel can reconstruct the target sequence of another task on its own; and inputting the outputs of the three channels into the multi-task learning network according to the gating vectors.

[0097] The consistency condition is determined based on the direction of change of the outputs of the two tasks after feature perturbation. A controlled perturbation is applied to a certain dimension in the joint state representation while keeping other dimensions unchanged. If both the water demand output and the nutrient demand output produce stable and unidirectional responses, the corresponding dimension is written into the common feature channel. If only the water demand output produces a stable response, it is written into the water task-specific channel. If only the nutrient demand output produces a stable response, it is written into the nutrient task-specific channel. If the response directions of the two tasks are opposite and cannot be explained by the phenological stage probability, the dimension is not written into the common channel for the time being, and is treated as a task-exclusive candidate in the subsequent gradient conflict decomposition.

[0098] The three channels do not fix the original variables, but dynamically divide them according to the potential dimensions of the joint state representation. The same potential dimension can enter the common channel during the leaf expansion stage and enter the water-specific channel during the root and stem enlargement stage. The gating vector is smoothly adjusted as the stage probability changes continuously, avoiding channel abrupt changes caused by slight changes in phenological stage labels. The output of the common channel simultaneously inputs the water demand branch and the nutrient demand branch, while the dedicated channel only inputs the corresponding task branch. The phenological branch uses the joint state representation that has not undergone task-specific screening, maintaining access to complete environmental information for stage inference.

[0099] Furthermore, the construction of the candidate lag segments includes: generating multiple historical segments with different start and end times for each type of meteorological variable, soil state variable, and irrigation and fertilization variable; calculating the joint correlation representation between each historical segment and the current phenological stage probability change, water demand change, and nutrient demand change; retaining historical segments with consistent time orientation in at least two associated tasks, and marking historical segments that are significant only in a single task as task-specific segments; merging duplicate information based on the temporal coverage relationship between segments to form common lag paths and task-specific lag paths, and then weighting and encoding the lag paths according to the phenological stage probability.

[0100] The start and end points of historical segments are jointly determined by the variable change events and the phenological observation points. For precipitation and irrigation records, the event occurrence point is used as the candidate starting point and extended to the point where the soil moisture content returns to stability. For temperature and radiation sequences, the point where the continuous change direction begins is used as the candidate starting point and extended to the point where the direction reverses. For soil nutrients and fertilization records, the point where the input event or available nutrients show continuous change is used as the candidate starting point. Multiple candidate segments with the same starting point but different ending points are nested. If they have different starting points but cover the same environmental event, they are overlapped. During encoding, the segment length, event location, and phenological stage span are preserved so that the model can distinguish whether the same cumulative amount is formed by short-term concentrated changes or long-term slow changes.

[0101] ;

[0102] in, Indicates the first The historical segment is relative to the first The retained score of each related task. This represents the directional correlation between the segment encoding and the corresponding change in the task objective. This indicates the consistency between the direction of a time segment and the direction of evolution of phenological stages. This indicates the amount of coverage redundancy between the fragment and the already retained fragment. This represents the fluctuation amount in the correlation direction of a segment as it repeatedly changes across adjacent training batches. , , and These represent the combination coefficients determined by the training process and kept non-negative, such as when a segment of continuous rainfall decreases while the probability of the corresponding phenological stage shifts towards root and stem enlargement and water demand increases. and Take the larger value; if another segment already covers the same time range, then... Increase and suppress duplicate retention.

[0103] Please refer to Figure 6 , Figure 6 The horizontal axis represents the continuous sampling section, the vertical axis represents the normalized engineering quantity, and the curves represent four types of quantities: retained score, directional correlation, stage consistency, and redundancy deduction. Figure 6 The corresponding candidate lag segment retention score formula demonstrates the main evidence used to determine whether a historical segment enters a common lag path or a task-specific lag path. The directional correlation curve represents the unidirectional relationship between segment encoding and changes in phenological stages, water requirements, or nutrient requirements; the stage consistency curve indicates whether the segment's temporal direction aligns with the phenological evolution direction of Polygonatum; the redundancy deduction curve represents the inhibition caused by the degree of overlap between the segment and already retained segments; and the retention score curve is the screening result after considering all the above factors.

[0104] from Figure 6 As can be seen, the retained score curve forms a distinct high-value plateau in the transition to the high-level segment, indicating that the historical segments in this segment possess both strong task relevance and stage-oriented consistency, making them suitable for subsequent path coding. The directional relevance curve gradually rises in the latter part, indicating a more stable temporal orientation between continuous low rainfall, irrigation intervals, changes in soil available water, and rising water demand. The stage-oriented consistency curve rises in the middle and later stages and then flattens out, indicating that the environmental changes corresponding to the segments can explain phenological progression rather than isolated disturbances. The redundancy deduction curve continues to decline, indicating that as duplicate coverage segments are suppressed, the retained paths better represent the true response periods.

[0105] The common lag path is formed by connecting segments that maintain a consistent time direction in at least two tasks in phenological stage, water demand, and nutrient demand. The connection conditions include that the end point of the previous segment is no later than the start point of the next segment, the environmental change directions corresponding to the two segments can be continuously explained, and there are no contradictory management events between them. The task-specific lag path is formed by connecting segments that maintain a stable association only in a single task. The path encoding adopts a recursive method from early to late, and the actual time interval is written at each connection position. If the common path and the task-specific path share some segments, the shared segment is encoded only once and enters the corresponding task through different gating weights to reduce the amplification of the joint state representation by duplicate information.

[0106] Preferably, the reliability of the stage transition is further corrected by the consistency of water and fertilizer prediction, including: inputting the current stage probability into the water demand branch and the nutrient demand branch respectively to obtain the initial prediction sequence under the corresponding stage conditions; calculating the deviation between the initial prediction sequence and the changes in soil water content, soil nutrient status, and historical management response; when the candidate stage transition causes the water demand and nutrient demand to simultaneously exhibit transitions that do not conform to the continuous response relationship, reducing the reliability of the corresponding transition path and feeding the reliability difference back to the stage probability calculation process; retaining the transition path verified by multiple consecutive time points to generate a phenological state sequence constrained by the water and fertilizer tasks.

[0107] Specifically, the water demand branch generates the water demand trajectory under the current stage conditions using the current stage probability and water task representation as input, and the nutrient demand branch generates the fertilizer demand trajectory using the same stage probability and nutrient task representation as input. The continuous response relationship is learned from the real state changes after historical management events. For example, after irrigation, the available water in the soil changes first, and then the water demand gap narrows. The nutrient status change after fertilization is constrained by the soil moisture conditions. If a candidate stage transition causes the predicted water demand and predicted fertilizer demand to jump in opposite directions at the same time without a cause, and neither the soil state nor the management records can provide a corresponding source of change, then the jump is decomposed into candidate transition paths and a confidence deduction is formed.

[0108] When the credibility difference is back-transferred, only the stage transfer score related to the candidate transfer path is adjusted, without directly modifying the stage observation score. This avoids the water and fertilizer task completely replacing phenological observation. Continuous verification at multiple time points means that the direction of change of stage probability, the direction of change of water demand, and the direction of change of nutrient demand maintain an interpretable continuous relationship in adjacent observations, and the relationship can be supported by environmental accumulation or historical management response. The verified transfer path is written into the stage state cache, and subsequent time points prioritize probability propagation along the cache path. When a new phenological record conflicts with the cache path, the corresponding interval is recalculated without retrospectively modifying the early stage that has been confirmed by multiple observations.

[0109] In one embodiment, the decomposition of the conflict gradient includes: calculating the gradient vectors of the phenological stage identification task, water demand prediction task, nutrient demand prediction task, and water stress determination task relative to the shared temporal coding parameters; calculating the directional consistency quantity for any two gradient vectors, and projecting the opposite-direction parts onto the subspace formed by the corresponding task-specific channel representation; constructing the remaining directional components into a common effective gradient, and updating the shared temporal coding parameters only with the common effective gradient; and updating the water task-specific channel, nutrient task-specific channel, and phenological state inference parameters respectively with the projected task-specific components.

[0110] Gradient calculation is based on the effective label mask of the same batch of time series samples. Samples with missing phenological labels do not generate phenological task gradients, and samples with missing nutrient observations do not generate nutrient task gradients. However, the corresponding samples can still participate in other tasks with real supervision information. The direction consistency quantity is determined by the gradient inner product and the gradient length. When the gradients of two tasks are in the same direction, they are weighted according to the label completeness and historical stability and written into the common effective gradient. When the gradients of two tasks are in opposite directions, the projection of one gradient in the direction of the other gradient is calculated, and the projection part is assigned to the task-specific parameters according to the gating channel to which the corresponding latent dimension belongs. Conflicting parts that are not received by any dedicated channel do not participate in this update.

[0111] To avoid the task update order changing the decomposition results, all task gradients within the same training batch are saved once before parameter updates. Projection calculation is completed based on the saved original gradients. The exclusive components of each task are merged after all pairwise conflict judgments are completed. The common effective gradient is only included in the components that are not judged as reversed in the comparison of all related tasks. The potential dimensions that conflict with the phenology task and the water task but are consistent with the nutrient task are further decomposed according to the gating affiliation. If the corresponding dimension is located in the common channel, the part supported by the consistent task is retained and the conflict projection is deducted. If it is located in the water-specific channel, only the water task parameters are updated. After the training batch ends, the parameter sets are updated synchronously.

[0112] In a preferred embodiment, the cross-leakage constraint includes: reconstructing a water demand sequence using the output of the water task-specific channel and reconstructing a nutrient demand sequence using the output of the nutrient task-specific channel; generating cross-reconstructed sequences after exchanging the outputs of the two task-specific channels; forming a leakage amount based on the degree of fit of the cross-reconstructed sequence to the non-corresponding target sequence, and jointly writing the leakage amount and the common explanatory amount of the common feature channel for the two target sequences into the training constraint; separately encoding the coupling residuals that are simultaneously affected by soil water content and soil nutrient status, and only allowing the coupling residuals to participate in the output calculation of the two tasks through the common feature channel.

[0113] The reconstructor for the water demand sequence only receives the outputs of the water task-specific channel and the common channel, while the reconstructor for the nutrient demand sequence only receives the outputs of the nutrient task-specific channel and the common channel. After swapping the specific channels, the common channel remains unchanged. This allows us to observe whether the non-corresponding specific representation can replace the original task-specific representation to complete the reconstruction. If the nutrient-specific representation can reconstruct the main changes in water demand when the water-specific representation is missing, it indicates that the nutrient channel carries too much water information. The training constraints will reduce the weight of the relevant latent dimensions in the nutrient gating and transfer the parts that can explain both tasks to the common channel until the reconstruction result after the swap cannot stably follow the changes of the non-corresponding target.

[0114] The coupling residuals are obtained by subtracting the interpretable portions of their respective dedicated channels from the initial outputs of water demand and nutrient demand. The two remaining sequences are time-aligned, and residual segments with consistent change directions and common phenological conditions are retained. These residual segments are encoded as common water-fertilizer coupling characteristics. Typical cases include a decrease in soil moisture content leading to no significant change in nutrient availability but a limited response to nutrient demand. The common channel retains the joint changes between water status and nutrient uptake. The water-specific channel still describes the water demand gap, and the nutrient-specific channel still describes nutrient demand independent of water constraints, thus avoiding the same anomaly being repeatedly encoded by the two dedicated channels.

[0115] Furthermore, the joint correction of the phenological state sequence and the lag path includes: reallocating the stage weights of the common lag path and the task-specific lag path based on the phenological state sequence constrained by the water and fertilizer task; recalculating the stage observation score and the water and fertilizer demand sequence using the re-encoded joint state representation; comparing the stage transition direction, demand change direction, and historical response direction in the two calculations, and when the three are inconsistent, backtracking the time coverage interval along the lag path that caused the deviation and reducing the weight of the corresponding segment; iteratively executing the stage probability update and the lag path update until the results of two adjacent calculations meet the preset consistency convergence condition.

[0116] The joint correction takes the time series intervals for initial phenological inference and initial water and fertilizer prediction as the processing object. It recalculates the soft weight of the stage to which the lagged segment belongs according to the stage probability of each time point. For segments that span two stages, the weight is assigned according to the integral ratio of the stage probability of each time point over the duration of the segment. The entire historical data is not forcibly classified into a single stage. After recoding, a new joint state representation is obtained and the stage observation score, water demand trajectory and fertilizer demand trajectory are output again. The stage transition direction is jointly determined by the stage with the highest probability at adjacent time points and the complete probability distribution. The demand change direction is jointly determined by the first-order change and duration of the prediction sequence. The historical response direction is determined by the environmental and management events in the corresponding lagged path.

[0117] ;

[0118] in, Indicates the joint correction round. This represents the set of time points that participated in the correction in this round. This indicates the number of time points in the set. Indicates the first The stage probability vector of the wheel. Indicates the first The stage probability vector for round -1. and They represent the first The water and nutrient requirements of the wheel. and They represent the first -1 round of water and nutrient requirements, and This represents the nonnegative coefficients used to unify different output scales. This represents the combined change in the results of two adjacent rounds. For example, if the stage probability has stabilized but the water demand trajectory still changes repeatedly due to a certain precipitation lag path, the third or second term remains large. The correction process continues to backtrack the corresponding segment until the stage state and water and fertilizer demand are simultaneously consistent.

[0119] Please refer to Figure 7 , Figure 7 The horizontal axis represents the continuous sampling segment, and the vertical axis represents the normalized engineering quantity. The curves represent the overall change, stage change, moisture change, and nutrient change, respectively. This figure corresponds to the calculated relationship of the differences between adjacent rounds of results during the joint calibration process, and is used to measure whether the phenological stage probability, water requirement, and nutrient requirement tend to be consistent after local recalculation. The overall change curve reflects the overall difference at all calibration time points, the stage change curve reflects the adjustment range of the stage probability distribution, and the moisture change and nutrient change curves reflect the degree of change in water and fertilizer demand output during continuous calibration, respectively.

[0120] from Figure 7 As can be seen, the comprehensive change curve rapidly declines from a high level, then continues to converge after a slight fluctuation in the middle, indicating that the initial correction phase significantly adjusted the weights of lag paths, stage affiliations, and demand sequences, subsequently stabilizing. The stage change curve declines more gently, suggesting that phenological probabilities are not directly replaced by water and fertilizer tasks, but rather gradually adjusted under the constraints of adjacent stages and record completeness. The moisture change curve is higher in the earlier stages than in the later stages, indicating that water demand is highly sensitive to the redistribution of lag paths; the nutrient change curve declines synchronously overall, indicating that nutrient output gradually aligns with the stage state after being constrained by common channels and soil moisture conditions.

[0121] When the direction of stage transition, the direction of demand change, and the direction of historical response are inconsistent, the system reverses its position from the predicted point of deviation along the gating weights to the lagging segment that contributes the most. Then, it checks whether the segment crosses the stage boundary, whether it contains missing intervals, and whether it overlaps with management events. Segments with uncertain stage affiliation have their stage weights reduced, segments caused by missing intervals have their observation reliability weights reduced, and segments that are a mixture of management events and natural changes are split into two sub-segments before and after the event. The correction process retains the source of each weight adjustment so that the same path cannot be restored to its original weight in subsequent rounds due to a single task error, until the comprehensive change enters the convergence range determined by the training stability interval.

[0122] In this embodiment, the generation of the early warning result includes: calculating a water gap sequence at consecutive time points based on the water demand and available soil water obtained through training with a common effective gradient; assigning stage-sensitive weights to the water gap sequence according to the phenological stage probability, and determining water shortage evidence by combining the water-fertilizer coupling residual in the common feature channel; removing anomalous components that can be explained by the nutrient task-specific channel and abrupt components caused by low-confidence stage transitions; accumulating the remaining water shortage evidence according to duration, growth direction, and stage transition confidence, and outputting the water stress level and early warning period associated with the corresponding phenological stage.

[0123] Soil available water is calculated on a unified time axis by considering the current soil moisture content, recent recharge events, and continuous consumption trends, and is aligned with the water demand in the same forecast interval. When the water gap is less than zero, it indicates that the current available water can cover the forecast demand and does not enter the risk accumulation phase. When the water gap is greater than zero, a stage-sensitive weighted value is obtained by combining the phenological stage probability. The common water-fertilizer coupling residual is used to retain the part that affects nutrient absorption due to water limitation. Anomalies that can be independently explained by the nutrient task-specific channel are deducted from the water shortage evidence. Instantaneous water demand jumps caused by low-confidence stage transitions are retained as unconfirmed components to subsequent time points and do not directly form a stress level.

[0124] ,

[0125] ,

[0126] ;

[0127] in, Indicates the first Moisture gap at a given time point, weighted by phenological stage sensitivity This indicates the abnormal components that can be explained by the dedicated channel for nutrient tasks. This represents the jump component caused by the low-confidence stage transition. This indicates the remaining evidence of water scarcity after excluding non-water-scarce causes. This represents the cumulative amount of continuous risk. Indicates the first -1 point in time continuous risk accumulation, This indicates that the attenuation coefficient of the preceding risk is retained and its value is between 0 and 1. Indicates the weight of the duration of the current evidence. Indicates the credibility of the stage transition. This indicates that the historical samples have been verified to belong to the first... A set of stable time points for each phenological stage. Indicates the first The division of each phenological stage Indicates the stable point of time Continuous risk accumulation This represents a dynamic hierarchical boundary obtained from the historical cumulative distribution of the corresponding stage, such as continuous water shortages. Multiple time points remain positive and As it gradually increases, Continuous accumulation and crossing the corresponding stage boundary, while isolated anomalies are due to subsequent... Return to zero and follow attenuation.

[0128] Please refer to Figure 8 , Figure 8 The horizontal axis represents the continuous sampling section, and the vertical axis represents the normalized engineering quantity. The curves represent the remaining water shortage evidence, risk accumulation, and dynamic boundary, respectively. This figure corresponds to the calculation process in the early warning result generation formula, which performs stage-sensitive weighting on the water deficit, deducts nutrient-specific anomalies and low-confidence stage jumps, and then calculates risk accumulation according to the duration and stage transition confidence. The remaining water shortage evidence curve represents the valid evidence after removing non-water shortage causes; the risk accumulation curve represents the result of continuous evidence gradually accumulating while retaining prior risks; and the dynamic boundary curve represents the graded reference line determined based on the stable sample distribution of the current and adjacent phenological stages.

[0129] from Figure 8 As can be seen, the remaining water shortage evidence remained low in the early stages, then rose steadily after accumulation, indicating that the gap between water demand and available soil water gradually widened and could not be fully explained by nutrient-specific pathways or low-confidence stage jumps. The risk accumulation curve rose faster in the latter part, reflecting the cumulative risk formed by the combined effects of duration weighting and stage confidence of continuous positive evidence. The dynamic boundary curve fluctuated around the median, used to avoid directly using fixed thresholds between different phenological stages.

[0130] The warning level is dynamically graded according to the current phenological stage with the highest probability and adjacent candidate stages. The boundaries are combined with the stage probability to avoid directly switching fixed boundaries when stages change. The start of the warning period is traced back to the earliest time when the continuous risk accumulation began to increase in the same direction and the remaining water shortage evidence was continuously positive. The end point is confirmed after the risk accumulation continuously decreases and falls below the recovery boundary of the corresponding stage. If the credibility of the phenological stage decreases within the warning interval, the existing risk evidence is retained but the level is marked as pending correction. The level is recalculated using the same water shortage evidence sequence after the stage probability stabilizes again. The observed continuous water gap is not deleted.

[0131] In a preferred embodiment, model training employs a sample organization method that divides samples according to the complete growth cycle. The same continuous phenological interval cannot simultaneously enter the training set and the validation set to prevent information leakage at adjacent time points. At the start of training, the shared temporal coding is first made able to reconstruct continuous state changes on a unified time axis, and then phenological stage probabilities, gating channels, and water and fertilizer demand tasks are added. After the stage transition path can be kept continuous, gradient conflict decomposition and cross-leakage constraints are added. Training batches are jointly extracted according to the phenological stage distribution and water status distribution, so that the stage transition interval, stable growth period, and continuous water shortage interval can all participate in parameter updates. When the model is saved, the data field order, time index rules, stage transition map, and the generation basis of dynamic grading boundaries are recorded simultaneously.

[0132] In a preferred embodiment, at the start of a new cultivation cycle, the trained shared temporal coding parameters and stage transition relationships are used, and the scale and time offset of the input data are corrected using the early phenological records of the current cycle. The correction only applies to the input mapping and phenological observation scores and does not directly modify the learned common lag paths. When continuous meteorological changes that are not historically covered occur in the current cycle, the relevant segments are saved in the candidate path area and participate in inference with a lower gating weight. After the segments simultaneously obtain support from at least two of the tasks of phenological state, water requirement, or nutrient requirement at consecutive time points, they are then incorporated into the common lag path or the corresponding task-specific lag path according to the scores of the aforementioned segments.

[0133] In this embodiment, the output data includes the phenological stage probability, stage transition confidence, water requirement, nutrient requirement, water stress level, warning start and end time periods, and the source of water shortage evidence for each unified observation point. The source of water shortage evidence is represented by lag path identifier, gating channel attribution, and corresponding time coverage interval. No explanatory labels that are detached from the input data and model calculation process are output. When subsequent phenological records or soil state records are supplemented, joint state coding, stage probability update, gating calculation, and risk accumulation are re-executed for the time interval affected by the supplemented records. The unaffected early intervals retain the original results, so that the entire working process maintains a traceable and consistent calculation chain under the condition of continuous increase in time series data.

[0134] Furthermore, before entering the shared temporal encoding, the training samples undergo a temporal consistency check. The check includes whether the phenological record time falls within the cultivation cycle, whether soil state changes have corresponding observation sources, whether irrigation and fertilization events have duplicate timestamps, and whether the meteorological sequence shows isolated mutations that cannot be supported by adjacent observations. For duplicate events, the original records are retained and duplicate tags are generated. For isolated mutations, anomaly source tags are generated instead of direct deletion. Records that exceed the cultivation cycle are transferred to the isolation area. The tags are used as an additional dimension of the input vector in the encoding, enabling the model to learn the relationship between data credibility and task output. When training supervision information conflicts, the range of samples involved in the task loss is determined by the completeness of the continuous observation chain, and records that cannot form continuous evidence are retained for subsequent supplementary verification.

[0135] Preferably, the water and nutrient requirements are output in a sequence format consistent with the unified observation interval. Each prediction location simultaneously saves the corresponding stage probability snapshot and lag path contribution. When the prediction interval is rolled forward, the prediction locations that have obtained the true soil conditions or management responses are converted into correction samples, and the stage probabilities used in the prediction are compared with the subsequently confirmed phenological states. If the deviation comes from the change in the stage transition time point, only the stage observation score and related gating weights are updated. If the deviation comes from the change in the environmental response lag, the retention score of the corresponding candidate lag segment is updated. If the deviation comes from the leakage of task-specific information, cross-reconstruction and channel division are re-executed to avoid changing all parameters at the same time using a single total error.

[0136] Specifically, the entire computation chain maintains the correspondence between data and results using time-point version identifiers. Each new, corrected, or supplementary observation generates a new data version. The joint state representation, stage probability, gating vector, water and fertilizer demand sequence, and risk accumulation record the data version and previous state version on which they depend. During local recalculation, the previous effective state is restored starting from the earliest affected unified observation point, and subsequent results are updated sequentially along the time axis. Changes to stage transition paths, common lag paths, and task-specific lag paths retain the weights and trigger sources before and after the changes. This ensures that the same input version can reproduce the same prediction and warning results, and limits the impact of new evidence on existing results to an interval with temporal causal correlation.

[0137] In a preferred embodiment, when phenological records for the current time point are missing but meteorological, soil, and management data are complete, the stage inference retains the probability of the previous time point and propagates according to the allowed transfer path. The stage is not forcibly determined using a single environmental variable. The water task gating vector and the nutrient task gating vector are calculated based on the probability distribution after propagation. The stage transition confidence gradually decreases as the missing data persists. The water and fertilizer demand output is still formed by the joint state representation and the task-specific channel. During the early warning accumulation process, the reduced stage transition confidence is written into the water shortage evidence weight. After the phenological records are supplemented, a local recalculation is performed from the starting point of the missing interval. If the supplemented records support the original propagation path, the corresponding confidence is restored. If they support another adjacent path, the lag path stage weights of that interval are reallocated and subsequent predictions are updated.

[0138] Furthermore, when soil moisture status is temporarily missing while phenological and meteorological data are continuous, the water gap does not use unconstrained single-value interpolation results. The system forms an available water status interval based on the most recent effective soil status, precipitation and irrigation supply, continuous meteorological consumption direction, and the current phenological stage. The water demand branch calculates the upper and lower boundaries of the interval and generates the gap range. Risk accumulation uses the part of the gap range that can be supported by continuous observation. If the warning levels corresponding to the upper and lower boundaries are different, the pending confirmation level is maintained and the source of the missing data is recorded. After the subsequent real soil status is entered, the status interval is replaced and recalculated along the original time axis. The nutrient task-specific abnormal component and the common water and fertilizer coupling residual are updated synchronously to avoid missing soil data being directly interpreted as continuous water shortage.

Claims

1. A method for predicting and early warning the phenological water and fertilizer demand of Polygonatum odoratum based on multi-task learning, characterized in that, include: Acquire phenological records, meteorological sequences, soil moisture content, soil nutrient status, and historical irrigation and fertilization records of Polygonatum odoratum, and form time-series samples according to a unified time axis; The time-series samples are subjected to shared temporal coding to obtain a joint state representation; Based on the joint state characterization, output phenological stage probability and stage transition reliability, and generate water task gating vector, nutrient task gating vector and common feature channel accordingly. Water requirements, nutrient requirements, and water stress levels were obtained through a multi-task learning network. Calculate the consistency of training gradients for each task, decompose conflicting gradients, and update the shared temporal coding parameters. Early warning results are generated by combining the aforementioned water demand, soil available water change trends, and phenological stage sensitivity.

2. The method for predicting and early warning the phenological water and fertilizer demand of Polygonatum odoratum based on multi-task learning according to claim 1, characterized in that, The formation of the time series sample includes: establishing a multi-scale time index based on phenological observation time points, and mapping the meteorological sequence, the soil moisture state, the soil nutrient state, and the historical irrigation and fertilization records to the corresponding index intervals; For data with differences in sampling intervals, the original time offset is retained, and a time-series addition reflecting the duration, rate of change, and distance of the most recent effective observation is generated; Candidate lag segments are constructed based on the response lag range of different phenological stages of Polygonatum. Each candidate lag segment and the current phenological observation are input into the shared time-series coding process to obtain a joint state characterization that includes short-term environmental disturbances and staged cumulative responses.

3. The method for predicting and early warning the phenological water and fertilizer demand of Polygonatum odoratum based on multi-task learning according to claim 1, characterized in that, Obtaining the probability of the phenological stages and the reliability of the stage transitions includes: establishing a stage transition relationship connected in the order of budding, leaf unfolding, reproductive growth, rhizome enlargement and aboveground decline of Polygonatum sibiricum; The joint state representation is mapped to stage observation score and stage transition score, respectively; The probability of the current stage is calculated based on the stage probability of the previous time point, the observation score of the current stage, and the allowed transfer direction. Path constraints are imposed on candidate transitions that cross non-adjacent stages, and the reliability of the stage transitions is calculated based on the stage probability differences at consecutive time points, changes in environmental accumulation, and the completeness of phenological records, forming a phenological state sequence with temporal continuity.

4. The method for predicting and early warning the phenological water and fertilizer demand of Polygonatum odoratum based on multi-task learning according to claim 1, characterized in that, The generation of the water task gating vector, the nutrient task gating vector, and the common feature channel includes: conditionally combining the phenological stage probability with the joint state representation, and calculating the selection weights of water-related features and nutrient-related features respectively; Features that simultaneously satisfy the consistency conditions of water task and nutrient task are written into the common feature channel, and features that only satisfy the consistency conditions of a single task are written into the corresponding task-specific channel. Cross-leakage constraints are set for the dedicated channels for the water task and the nutrient task, so that neither dedicated channel can reconstruct the target sequence of the other task on its own. The outputs of the three channels are input into the multi-task learning network according to the gating vector.

5. A method for predicting and early warning the phenological water and fertilizer demand of Polygonatum odoratum based on multi-task learning according to claim 2, characterized in that, The construction of the candidate lag segments includes: generating multiple historical segments with different start and end times for each type of meteorological variable, soil state variable, and irrigation and fertilization variable; Calculate the joint correlation between the probability changes, water demand changes, and nutrient demand changes of each historical segment and the current phenological stage; Retain historical fragments that have a consistent time pointer across at least two related tasks, and mark historical fragments that are significant only in a single task as task-specific fragments; Based on the temporal coverage relationship between segments, duplicate information is merged to form a common lag path and a task-specific lag path. Then, the lag path is weighted and encoded according to the probability of the phenological stage.

6. The method for predicting and early warning the phenological water and fertilizer demand of Polygonatum odoratum based on multi-task learning according to claim 3, characterized in that, The reliability of the stage transition is also corrected by the consistency of water and fertilizer prediction, including: inputting the current stage probability into the water demand branch and the nutrient demand branch respectively to obtain the initial prediction sequence under the corresponding stage conditions; Calculate the deviation between the initial prediction sequence and changes in soil moisture state, changes in soil nutrient state, and historical management responses; When the transition of candidate stages causes a jump in water demand and nutrient demand that does not conform to the continuous response relationship, the credibility of the corresponding transition path is reduced, and the credibility difference is fed back to the stage probability calculation process. The transfer paths verified at multiple consecutive time points are retained to generate the phenological state sequence constrained by the water and fertilizer tasks.

7. The method for predicting and early warning the phenological water and fertilizer demand of Polygonatum odoratum based on multi-task learning according to claim 4, characterized in that, The decomposition of the conflict gradient includes: calculating the gradient vectors of the phenological stage identification task, water demand prediction task, nutrient demand prediction task, and water stress determination task relative to the shared temporal coding parameters, respectively. For any two gradient vectors, calculate the directionally consistent quantity and project the oppositely oriented portions onto the subspace formed by the corresponding task-specific channel representation; The remaining in-direction components are constructed into a common effective gradient, and the shared temporal coding parameters are updated only with the common effective gradient; The projected task-specific components are used to update the water task-specific channel, the nutrient task-specific channel, and the phenological state inference parameters, respectively.

8. The method for predicting and early warning the phenological water and fertilizer demand of Polygonatum odoratum based on multi-task learning according to claim 4, characterized in that, The cross-leakage constraint includes: reconstructing the water demand sequence using the dedicated channel for the water task, and reconstructing the nutrient demand sequence using the dedicated channel for the nutrient task; After swapping the outputs of the two task-specific channels, cross-reconstruction sequences are generated respectively; The leakage amount is formed based on the degree of fit of the cross-reconstruction sequence to the non-corresponding target sequence, and the leakage amount and the common explanatory amount of the common feature channel for the two target sequences are jointly written into the training constraints. The coupled residuals, which are simultaneously affected by soil moisture and soil nutrient status, are encoded separately, and only the coupled residuals are allowed to participate in the output calculation of the two tasks through the common feature channel.

9. A method for predicting and early warning the phenological water and fertilizer demand of Polygonatum odoratum based on multi-task learning according to claim 6, characterized in that, The joint correction of the phenological state sequence and the lag path includes: reallocating the stage weights of the common lag path and the task-specific lag path based on the phenological state sequence constrained by the water and fertilizer task. The stage observation scores and water and fertilizer demand sequences are recalculated using the re-encoded joint state representation. Compare the phase transition direction, demand change direction, and historical response direction in the two calculations. When the three are inconsistent, backtrack along the time coverage interval of the lag path that caused the deviation and reduce the weight of the corresponding segment. The iterative execution phase includes probability updates and lag path updates until two adjacent results satisfy the preset consistency convergence condition.

10. A method for predicting and early warning the phenological water and fertilizer demand of Polygonatum odoratum based on multi-task learning according to claim 7, characterized in that, The generation of the early warning results includes: calculating a water gap sequence at consecutive time points based on the water demand and available soil water obtained through training with the common effective gradient; Based on the phenological stage probability, the water deficit sequence is assigned a stage-sensitive weight, and water-fertilizer coupling residuals in the common feature channel are combined to determine evidence of water shortage. Remove anomalous components that can be explained by the dedicated channel for the nutrient task and abrupt components caused by low-confidence stage transitions; The remaining water shortage evidence is accumulated according to its duration, growth direction, and the credibility of stage transitions, and the water stress level and warning period associated with the corresponding phenological stage are output.

Images