A method for predicting the timing of pesticide application for cotton defoliation and ripening and its application in the synergistic effect of cotton quantity and quality.

Abstract

This invention discloses a method for predicting the application time of pesticides for cotton defoliation and ripening, and its application in the synergistic development of cotton yield and quality. It relates to the field of crop cultivation and management technology, and includes acquiring comprehensive physiological and environmental stress data for cotton fields. The comprehensive stress data includes at least the incidence levels of pests and diseases, and one or more of temperature stress, water stress, and light stress. A physiological response quantitative correction model is constructed, using the comprehensive stress data as input. This method for predicting the application time of pesticides for cotton defoliation and ripening, and its application in the synergistic development of cotton yield and quality, can adaptively correct the prediction results, ensuring that the output application window simultaneously meets the requirements for defoliation rate and the synergistic goals of yield and quality. It avoids the problem of complete prediction failure when comprehensive stress occurs, leading to substandard defoliation or simultaneous deterioration of boll weight and fiber quality, maintaining the stability and feasibility of decision-making in complex and variable field environments.

Description

Technical Field

[0001] This invention relates to the field of crop cultivation and management technology, specifically to a method for predicting the timing of pesticide application for cotton defoliation and ripening and its application in the synergistic effect of cotton quantity and quality. Background Technology

[0002] Defoliation and ripening of cotton are key agronomic measures to ensure the quality of mechanized harvesting. The precise timing of pesticide application directly affects whether cotton yield and fiber quality can be optimally synergistically achieved. In existing technologies, to overcome the subjectivity of traditional experience-based decision-making, prediction methods based on accumulated temperature and growth period have been developed. However, these methods only consider single factors such as temperature or date, making them difficult to adapt to variety changes and climate fluctuations. With the development of artificial intelligence technology, machine learning prediction models integrating multi-source data such as meteorological, soil, and variety data have emerged. These models can handle nonlinear relationships and time-series characteristics, significantly improving the accuracy of defoliation rate prediction. They are also beginning to explore establishing a quantitative link between pesticide application time and yield and quality, striving to achieve a synergistic effect in both quantity and quality while ensuring effective defoliation.

[0003] However, all the aforementioned existing technical solutions implicitly assume a fundamental premise: that cotton plants are in an ideal environment and a healthy physiological state. In actual field production, not only do biotic stresses such as bollworms and aphids occur frequently, but abiotic stresses such as periods of high and low temperatures, drought, waterlogging, weak light, and nutrient imbalances also damage leaf function, interfere with the balance of endogenous hormones and the accumulation of nutrients, thus significantly altering the cotton's response to defoliation and ripening agents. When biotic stress or abiotic adverse conditions occur, the application timing given by existing predictive models often fails, resulting in defoliation rates failing to meet machine harvesting standards, or the defoliation process spiraling out of control, leading to a large number of immature bolls falling off and being forced to crack, ultimately resulting in a sharp reduction in boll weight, severe deterioration of fiber length and strength, and making the goal of synergistic yield and quality impossible to achieve. Moreover, the types, severity, and spatiotemporal distribution of pests and diseases are highly dynamic and heterogeneous, and existing predictive models completely lack this dimension in their input features, making the so-called "optimal window for quantity and quality synergy" a meaningless and erroneous signal under biotic stress.

[0004] Therefore, it is urgent to break through the current technical framework and use comprehensive stress information that reflects the true physiological state of crops as the intrinsic driving variable of the prediction model to construct a precise decision-making method for pesticide application time with dynamic adaptability to complex field environments, so as to truly achieve the quantitative and qualitative synergy of cotton defoliation and ripening in complex and ever-changing field environments. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting the application time of pesticides for cotton defoliation and ripening, and its application in the synergistic effect of cotton quantity and quality, so as to solve the problems mentioned in the background art.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for predicting the application time of pesticides for cotton defoliation and ripening, which is applied to the quantity and quality co-management of cotton, including: acquiring cotton growth and development data, meteorological data, soil moisture data and variety genetic characteristic data, and constructing a first prediction model based on multi-source data, and outputting a preliminary candidate window for pesticide application time; Includes the following steps: Acquire comprehensive physiological and environmental stress data for cotton crops; the comprehensive stress data shall include at least the occurrence level of pests and diseases, as well as one or more of temperature stress, water stress, and light stress. A physiological response quantitative correction model is constructed. This model takes the comprehensive stress data as input and outputs a correction coefficient that characterizes the degree of shift in the plant's sensitivity to defoliation ripening agent caused by the current environmental and biological comprehensive stress. The value of the correction coefficient is positively correlated with the fluctuation level of endogenous hormones and the degree of change in metabolic activity in the plant. Using the correction coefficient as a reverse constraint, the initial candidate window for application time is corrected by time domain compensation to obtain the corrected application window; The modified application window is simulated and analyzed to predict the defoliation process curve and the corresponding changes in yield and quality after application. A dual verification boundary is set, consisting of a defoliation rate threshold and a comprehensive yield and quality benefit threshold. The defoliation process curve and yield and quality changes obtained from the simulation are compared with the dual verification boundary. When the comparison results simultaneously meet the defoliation rate threshold and the comprehensive yield and quality benefit threshold, the modified application window will be output as the final application time. When the comparison results do not meet the requirements simultaneously, the deviation in leaf removal rate and the deviation in overall yield, quality and benefits are mapped inversely to the adjustment increment of the correction coefficient, and a secondary compensation correction and simulation are triggered for the application window until the comparison results fall into the double verification boundary.

[0007] Preferably, the physiological response quantitative correction model is based on the fluctuations in endogenous hormones and changes in metabolic activity of cotton plants caused by comprehensive stress as the mutagenic basis. The correction coefficient is generated by a pre-established quantitative mapping relationship between multi-level indicators reflecting the intensity of comprehensive stress and the sensitivity of defoliation response. Comprehensive stress includes biotic stress and abiotic stress. The multi-level indicators reflecting the intensity of comprehensive stress can be a comprehensive stress index generated by the fusion of parameters such as temperature, moisture, light, and pests and diseases. The magnitude of the correction coefficient is proportional to the combined characterization of the plant's delayed response time and the magnitude of the overreaction caused by the combined stress, which includes both biotic and abiotic stress.

[0008] Preferably, the quantitative mapping relationship is obtained in the following manner: Under controlled conditions, different degrees of environmental and biological stresses were applied separately or in combination, and the changes in endogenous ethylene and abscisic acid content, photosynthetic rate and transpiration rate of cotton leaves were measured under each stress condition. Environmental stresses include, but are not limited to, high temperature, low temperature, drought, waterlogging, shading, and lack of fertilizer; biological stresses include pest feeding and pathogen infection. Simultaneously, the shedding time of detached leaves under treatment with standard concentration defoliating and ripening agents was measured. A multidimensional mapping surface is established with the comprehensive stress level as input and the relative change rate of shedding time as output. The correction coefficients are generated by interpolation based on the mapped surface.

[0009] Preferably, the specific method of the time domain compensation correction is as follows: The time offset is obtained by multiplying the correction coefficient by the basic response time constant; When the correction factor indicates a delayed plant response to the pesticide, the initial candidate window for pesticide application time is shifted forward by the time offset. When the correction coefficient indicates that the plant is overreacting to the pesticide, the preliminary candidate window for pesticide application time is shifted backward along the time axis by the time offset.

[0010] Preferably, the simulation is based on a pre-built threshold model of drug effect; The drug effect threshold model uses key environmental factors and the correction coefficients as driving variables to simulate the dynamic coupling process between the defoliant release curve and the cotton plant physiological response curve under a specific application window, and outputs the daily defoliation rate sequence and the time when cotton boll dry matter accumulation stops; key environmental factors include, but are not limited to, temperature, humidity, and light intensity.

[0011] Preferably, the leaf removal rate threshold in the dual verification boundary is set to the lowest leaf removal rate level corresponding to the machine harvesting standard; The threshold for comprehensive benefits of yield and quality is determined by the relative boll weight retention rate and the comprehensive fiber quality index, wherein the comprehensive fiber quality index is a weighted composite of fiber length, breaking strength and micronaire value.

[0012] Preferably, the first prediction model is a hybrid neural network model that integrates gradient boosting decision trees and long short-term memory networks; The gradient boosting decision tree is used to process the growth and development data, soil moisture data, and variety genetic characteristic data to extract nonlinear combination features. The long short-term memory network is used to process the time series of the meteorological data and extract time-varying features. The nonlinear combined features and the time-varying features are concatenated and then output as the preliminary candidate window for drug administration through a fully connected layer.

[0013] Preferably, the method further includes converting the final output application time into a variable spraying control command and sending it to the field operation equipment to perform precise spraying of the defoliant and ripening agent.

[0014] Preferably, the cotton growth and development data includes the flowering date and boll opening rate of cotton bolls in different parts of the plant; The meteorological data includes daily average temperature, maximum temperature, minimum temperature, and effective accumulated temperature; The genetic characteristics data of the variety include boll-forming period parameters and defoliation tolerance parameters.

[0015] A device for predicting the application time of a pesticide for cotton defoliation and ripening includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of a method for predicting the application time of a pesticide for cotton defoliation and ripening.

[0016] This invention provides a method for predicting the application time of pesticides for cotton defoliation and ripening, and its application in the coordinated development of cotton quantity and quality. It has the following beneficial effects: By using comprehensive stress information (including temperature, moisture, light, pests, and diseases) as the intrinsic driving variables for pesticide application timing decisions, a calibration model is constructed by tracing back to the root causes of physiological mutations. A time-domain compensation and dual closed-loop verification mechanism are introduced, ensuring that the determination of pesticide application timing no longer relies on the implicit assumption that cotton plants are in an ideal health state. Under real-world combined stress conditions in the field, this scheme can adaptively correct prediction results, ensuring that the output application window simultaneously meets the requirements for defoliation rate and the synergistic goals of yield and quality. This avoids the problem of prediction failure during stress, leading to substandard defoliation or simultaneous deterioration of boll weight and fiber quality, maintaining the stability and feasibility of decision-making in complex and variable field environments.

[0017] By using correction coefficients as inverse constraints to perform time-domain compensation corrections on the initial prediction window, and introducing a dual closed-loop verification mechanism that integrates the defoliation rate boundary and the yield, quality, and benefit boundary, the decision on pesticide application time is transformed from an open-loop, single-prediction process to a closed-loop, iterative optimization process. When the simulation results do not simultaneously satisfy both boundary conditions, the system automatically feeds back the deviation to the correction coefficients and triggers a secondary correction until an application window that satisfies both constraints is output. This avoids the direct output of erroneous decisions and maintains the stability and reliability of decisions under complex and ever-changing integrated field stress environments. Attached Figure Description

[0018] Figure 1 This is a system module structure diagram of the method for predicting the application time of pesticides for cotton defoliation and ripening according to the present invention; Figure 2 This is the overall flowchart of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example: Please refer to Figure 1 and Figure 2 This invention provides a technical solution: a method for predicting the application time of pesticides to promote cotton defoliation and ripening. In specific implementation, the initial prediction stage driven by abiotic factors is initiated first. Various sensors deployed in the field and manual survey terminals collect real-time data from the target cotton field, including hourly air temperature, relative humidity, and sunshine duration obtained from automatic weather stations; soil volumetric water content at different depths obtained from soil moisture sensors; and plant height, number of fruiting branches, flowering date of bolls at different nodes, and boll opening progress obtained through regular field surveys.

[0021] Simultaneously, the inherent genetic parameters of the cotton varieties planted in the field were retrieved from the variety database, specifically covering boll length and basic defoliation tolerance level. These structured growth and development data, soil data, and variety parameters were input into a gradient boosting decision tree model. The model's multiple decision trees underwent iterative computation to extract the nonlinear combination relationships between various features. Meanwhile, multi-day time-series meteorological elements were input into a long short-term memory network. This network's gating mechanism captured the evolution of temperature and humidity over time, outputting time-varying feature vectors. These two feature vectors were concatenated in the model and integrated through a fully connected layer, outputting a preliminary candidate window for pesticide application based solely on abiotic factors, such as a start and end date on a timeline.

[0022] To ensure the stable reproduction and engineering implementation of this hybrid neural network model, this invention standardizes and limits the model architecture configuration, training rules, and convergence criteria: The GBDT model structure is set as a six-layer gradient boosting decision tree concatenation architecture, with each decision tree having thirty to fifty branch nodes. Layer-by-layer feature splitting and nonlinear association mining are performed on static structured data such as cotton growth and development data, soil moisture data, and varietal genetic characteristics. During model training, the ReLU activation function is used to complete node feature mapping, discarding redundant and noisy features while retaining the core information of inter-feature interaction relationships. The LSTM network is set as a three-layer temporal memory unit structure, with each memory unit configured with sixty to eighty neurons, specifically processing time series composed of continuous daily meteorological data. Automatic shearing is achieved through a gating structure. Forget about short-term meteorological fluctuation noise, long-term residual temperature, and the temporal evolution of accumulated temperature; the model training uses multi-source datasets of field measurements in cotton fields over the past five years as training samples. The samples cover field measurement data of different cotton varieties, different ecological production areas, and different degrees of pest and disease occurrence. The training set and test set are divided into a 7:3 ratio. The training process uses a conventional regression loss function for backpropagation and iteration. When the prediction error of the test set no longer decreases for ten consecutive rounds and the fluctuation range is stable within the industry allowable error range, the model is determined to have reached the convergence state and the internal parameters of the model are fixed. Subsequently, the converged model is directly called to complete feature extraction and output the preliminary candidate window for pesticide application time. The entire process of building and training the hybrid model can be reproduced according to this hierarchical configuration, number of neurons, sample division rules, and convergence judgment criteria.

[0023] After completing the above routine predictions, the system enters the physiological quantitative correction phase targeting comprehensive environmental and biotic stresses. This phase is not limited to pests and diseases but comprehensively assesses all factors currently causing plants to deviate from a healthy state, a crucial step in avoiding the failure of existing technologies. This phase does not simply add single stress data as an additional input to the original model, but rather traces back to the physiological roots of the plant's pesticide response to comprehensive stress. Specifically, data is collected in real time through a field sensor network or manually input to quantify comprehensive stress indicators in the current environment, including extreme temperature values, soil moisture content, leaf light duration, and the severity of pests and diseases.

[0024] A pre-constructed physiological response quantification correction model receives the aforementioned comprehensive stress index data. This model is built upon a multidimensional mapping surface established through controlled experiments. The mapping relationship is as follows: the comprehensive stress level is the input axis, and the relative change rate of the time it takes for cotton plants to detach from their leaves under standard defoliation and ripening agent treatment is the output axis. The comprehensive stress level encompasses multiple dimensions, including temperature deviation, soil moisture deficit, insufficient light accumulation, and pest and disease occurrence levels. During model execution, interpolation is performed on this mapping surface based on the currently input level value to generate a defoliation response sensitivity correction coefficient. This correction coefficient comprehensively quantifies the delayed or excessive response to pesticides caused by leaf tissue damage and endogenous hormone imbalance.

[0025] To achieve a quantifiable and reproducible correspondence between comprehensive stress levels, changes in endogenous hormones, and correction coefficients, a unified standard for determining comprehensive stress levels in the field and implementation specifications for surface fitting and coefficient matching were formulated. Comprehensive stress levels in the field are jointly classified according to the following dimensions: Temperature stress level: classified into mild, moderate and severe according to the degree of deviation of the daily maximum and minimum temperatures from the suitable temperature range for cotton; Water stress levels are classified into drought and waterlogging levels based on the relative soil moisture content. Light stress level: classified according to the degree of deviation between the number of consecutive cloudy and rainy days and the number of sunshine hours; Pest and disease stress level: classified according to a combination of indicators such as the number of adult pests trapped, larval density, percentage of leaf damage, and pathogen infection rate.

[0026] The stress levels of the above-mentioned dimensions are weighted and summed to form a comprehensive stress intensity index. The specific weight allocation is as follows: temperature stress level weight 0.3, water stress level weight 0.3, light stress level weight 0.2, and pest and disease stress level weight 0.2. Each stress level is quantified into a numerical value: mild corresponds to 1, moderate corresponds to 2, and severe corresponds to 3. The formula for calculating the comprehensive stress intensity index is: temperature level × 0.3 + water level × 0.3 + light level × 0.2 + pest and disease level × 0.2. If only data for some dimensions is collected (e.g., light data is missing), the weight of the missing dimension is set to zero, and the weights of the remaining dimensions are renormalized proportionally to ensure that the sum of the weights is 1. The final calculated index value ranges from 1 to 3, with a larger value indicating more severe comprehensive stress. The level classification can be completed without complex numerical calculations. When constructing the multidimensional mapping surface, the comprehensive stress intensity index is used as the main input dimension and the relative change characteristics of cotton plant leaf shedding time are used as the output dimension. Conventional spatial interpolation logic is used to continuously and smoothly fit the discrete test sample points obtained from the controlled experiment, which fully covers all working conditions from no stress to multiple extreme stress combinations. The comprehensive stress intensity index is composed of multiple single stress levels such as temperature, moisture, light, and pests and diseases. In practical applications, the Kriging space interpolation method is used for mapping. Specifically, a spherical model is selected as the variation function, and the range is set to 1.5 times the width of the stress level interval (i.e., 1.5 times the span from mild to severe). The nugget value is set to 0, and the sill value is set to the variance of the sample points. The search strategy is to select the nearest 8 to 12 controlled test sample points within a circular range centered on the target point and covering 3 level units for interpolation calculation. When there are fewer than 6 effective sample points, the search radius is automatically expanded to cover all sample points. The interpolation weight is determined by the spatial correlation between sample points and the distance to the target point. Finally, a continuous leaf removal response sensitivity value is output. In the controlled experiment, the changes in the content of endogenous ethylene and abscisic acid in the leaves of cotton plants under stress were detected simultaneously. According to the stress level gradient, a one-to-one correspondence rule was established between the hormone fluctuation range and the delay and over-excitation of the cotton plant's defoliation agent response. Based on this correlation rule, the corresponding correction coefficient was directly matched and output. The modeling and calculation process of the physiological response quantitative correction model can be completely reproduced according to the level judgment criteria, surface fitting method and the correlation rule between hormones and correction coefficients.

[0027] The correction coefficient is injected as a reverse constraint parameter into the time compensation module, which then performs a conversion: multiplying the correction coefficient by the basic response time constant of the variety in a healthy state to obtain the time offset that needs to be corrected. If the correction coefficient indicates that the plant is in a delayed response state, the aforementioned preliminary application time candidate window is shifted forward by the time offset; if it indicates an overreaction state, it is shifted backward, thereby generating a corrected application window.

[0028] After obtaining the corrected application window, the system enters a dual closed-loop verification phase. A pre-built pesticide effect threshold model is invoked. This model uses the daily average temperature, relative humidity, and the previously generated correction coefficients within the prediction period as driving variables to simulate the coupled dynamic process of pesticide penetration rate, activity duration, and cotton plant physiological metabolism on the leaf surface after defoliation and ripening agent application. The simulation outputs a daily defoliation rate sequence starting from the first day of the corrected window, and simultaneously calculates the expected changes in boll weight retention rate and fiber quality index based on the point at which boll dry matter accumulation ceases.

[0029] The system pre-stores dual verification boundaries: one is the minimum defoliation rate corresponding to the mechanized harvesting standard, and the other is a yield-quality comprehensive benefit threshold weighted by boll weight retention rate and fiber quality comprehensive index. The comparison module compares the defoliation rate obtained from simulation at a specific number of days after application with the aforementioned defoliation rate boundary, and simultaneously compares the change in yield-quality comprehensive benefit with the aforementioned benefit threshold. The corrected application window is deemed valid only when both simulation results simultaneously meet or exceed their respective boundary conditions. This window is then output as the final application instruction in the form of a specific date, driving the spraying equipment to perform precise spraying.

[0030] If any comparison result fails to meet the boundary conditions, the system triggers an iterative optimization mechanism: the deviation between the simulated leaf removal rate and the boundary leaf removal rate, as well as the deviation in yield, quality, and efficiency, are mapped inversely to the adjustment increment of the correction coefficient, thus numerically correcting the original correction coefficient. The corrected correction coefficient then undergoes a complete process from time compensation to simulation deduction and then to boundary comparison, forming an iterative closed loop, until both comparison results meet the boundary requirements, at which point the final application time is output.

[0031] By incorporating real-time monitoring data of comprehensive stress as a driving variable into the pesticide application timing decision-making process, the shortcomings of predictions that fail completely under non-ideal conditions due to the assumption of healthy cotton plants are addressed. Even when actual field conditions such as high temperature and drought, low temperature injury, waterlogging and low light, or pest and disease infestation occur, this scheme can still provide a pesticide application time window that balances defoliation effectiveness with yield and quality, avoiding problems such as substandard defoliation rates or severe declines in boll weight and fiber quality caused by ignoring the comprehensive stress dimension.

[0032] Therefore, the scheme can adaptively output the application time that effectively balances the goals of defoliation effect and yield and quality in various field scenarios, whether the cotton plants are under comprehensive stress (including biotic and abiotic stress) or in a healthy state. The entire decision-making process will no longer suffer from uncontrolled deviations due to ignoring the comprehensive stress dimension.

[0033] The difference between the actual leaf removal rate obtained from the simulation and the preset leaf removal rate threshold, and the difference between the comprehensive benefit of yield and quality and the preset benefit threshold, are divided into multi-level gradient intervals according to the magnitude of the deviation. Each level of deviation interval corresponds to a fixed magnitude of the adjustment increment of the correction coefficient. The larger the deviation, the larger the adjustment increment; the smaller the deviation, the smaller the adjustment increment. This achieves an intuitive mapping between the magnitude of the deviation and the magnitude of the coefficient adjustment. The iterative optimization process is equipped with dual termination criteria: first, the corrected application window meets both the defoliation rate threshold and the comprehensive yield and quality benefit threshold requirements through simulation; second, the change in the correction coefficient after three consecutive adjustments tends to be small and the simulation results show no significant fluctuations. Meeting either condition will terminate the iteration process. After each incremental adjustment of the correction coefficient, the entire closed-loop operation of time domain compensation correction, simulation, and dual boundary comparison is restarted. The process gradually approaches the optimal application window by using a gradual, small-gradient adjustment method to avoid decision deviations caused by single large adjustments. The entire control logic of the closed-loop iterative optimization can be fully reproduced by dividing the deviation gradient, adjusting the incremental matching rules, and setting the iteration termination conditions.

[0034] The physiological response quantitative correction model takes the fluctuations in endogenous hormones and changes in metabolic activity in cotton plants caused by comprehensive stress (including temperature, water, light, pests, and diseases) as the causative factors. It generates a correction coefficient by pre-establishing a quantitative mapping relationship between the comprehensive stress intensity index and the sensitivity to defoliation response. The value of this correction coefficient is proportional to the comprehensive characterization of the plant's response delay time and the magnitude of the overreaction to pesticides. The implementation details of this scheme are as follows: Before constructing this physiological response quantitative correction model, the basic mapping relationship needs to be calibrated. Calibration is conducted in a controlled environment, such as a growth chamber or greenhouse, to eliminate interference from external weather and soil variations. Healthy plants of the local main cotton variety are selected as experimental materials. After the cotton plants enter the peak boll-forming stage, pest feeding stress and pathogen infection stress treatments are applied respectively.

[0035] To address pest feeding stress, cotton bollworm larvae or cotton aphids were selected as model pests. By quantitatively releasing different densities of pests onto cotton plants in different groups and maintaining a specific feeding duration, multiple levels of pest feeding intensity were formed.

[0036] To address pathogen infection stress, spore suspensions of cotton wilt pathogens or verticillium wilt pathogens are used, and different concentrations of inoculum are applied for root irrigation or leaf spraying to create multiple gradients of pathogen infection levels.

[0037] Meanwhile, multiple environmental control experiments were set up to address abiotic stresses: Temperature stress treatment: Different temperature gradients were set in an artificial climate chamber, including high temperature stress of daily maximum temperature of 35℃, 38℃ and 42℃, and low temperature stress of daily minimum temperature of 5℃ and 10℃, and the treatment was carried out for several days. Water stress treatment: The relative soil moisture content was set to 40% (drought stress), 80% (normal control), and 100% (waterlogging stress) by controlling the amount of irrigation. Light stress treatment: Different shading rates (30%, 60%, 90%) were set using shade nets to simulate continuous cloudy and rainy weather. Multiple replicates were set up for each treatment group, and hormone assays and defoliation response time measurements were carried out in parallel with the pest and disease treatment group.

[0038] After stress treatment, functional leaves from the upper, middle, and lower parts of plants were collected at regular intervals from each treatment group and the blank control group to determine the content of endogenous hormones. The target for determination was the content of ethylene precursor aminocyclopropanecarboxylic acid and abscisic acid, as these hormones are core signaling substances regulating leaf abscission and boll maturation. Quantitative detection was performed using enzyme-linked immunosorbent assay (ELISA) or high-performance liquid chromatography (HPLC) to obtain the content value of the target hormones per unit fresh weight of leaves, and the rate of change of content relative to the control group was calculated. Simultaneously, under standard conditions, plants in each treatment group were treated with the same concentration of defoliation ripening agent, and the abscission response time was measured using the abscission time of detached leaves or live marked leaves as an indicator. This resulted in a multidimensional dataset with a comprehensive stress intensity index (integrating multiple stress dimensions such as temperature, moisture, light, and pests / diseases) as the independent variable and the rate of change of endogenous hormone content and the rate of change of defoliation response time as the dependent variables.

[0039] Based on this dataset, a quantitative mapping relationship was established. Specifically, a multidimensional mapping surface was constructed using the comprehensive stress intensity index as the input dimension and the relative change rate of leaf abscission time after treatment with a standard defoliant / ripening agent relative to the healthy control group as the output dimension. The comprehensive stress intensity index integrates multiple stress dimensions such as temperature, moisture, light, and pests / diseases. This mapping surface can be smoothly constructed using interpolation algorithms for scattered data, such as bilinear interpolation or Kriging interpolation, to cover any combination of levels within the continuous input space. Each point on the mapping surface corresponds to a specific defoliation response sensitivity, which directly reflects the degree of delay or acceleration in the plant's response to the agent under that stress combination.

[0040] In practical field applications, the physiological response quantification correction model receives comprehensive stress indicators acquired in real time by the field monitoring system. The specific acquisition methods are as follows: Temperature stress level is quantified by the deviation of daily maximum and minimum temperatures recorded by automatic weather stations from the suitable temperature range for cotton varieties (usually a daily average temperature of 20-30℃); the greater the deviation, the higher the stress level. Water stress level is determined by the relative water content of the 20-40cm soil layer monitored by soil moisture sensors; below 50% of field capacity is considered drought stress, and above 90% is considered waterlogging stress. Light stress level is quantified by the ratio of the number of consecutive cloudy / rainy days recorded by sunshine duration sensors to the historical average for the same period. Pest and disease stress level is categorized into mild, moderate, and severe levels based on the number of adult pests captured by pheromone traps combined with field larval density surveys, and the number of pathogen spores collected by spore traps combined with field disease incidence, according to preset threshold ranges. The stress levels of the above dimensions are weighted and fused or mapped by a lookup table to generate a comprehensive stress intensity index, which serves as the input to the model.

[0041] The magnitude of the correction coefficient directly reflects the degree of overall deviation in the plant's response characteristics to defoliation and ripening agents caused by combined stresses (including temperature, water, light, pests, and diseases). For example, when plants suffer from high temperature and drought leading to stomatal closure and decreased photosynthetic rate, or from low temperature injury leading to reduced metabolic activity, or from waterlogging and hypoxia leading to impaired root function, or from pest feeding leading to leaf tissue damage, or from pathogen infection leading to vascular bundle blockage, the synthesis levels of endogenous ethylene and abscisic acid in the plant will change abnormally. This will cause the leaf abscission layer cells to respond to the agent prematurely or delayed, or lead to obstructed agent absorption and translocation. The larger the absolute value of the correction coefficient, the greater the deviation. This value will directly serve as the proportional coefficient for subsequent time-domain compensation of the deviation, thereby achieving precise reverse constraint adjustment of the application window, ensuring that the final output application time can offset the fluctuations in agent response caused by combined stresses, and truly maintain the quantitative and qualitative goals in complex field environments.

[0042] The quantitative mapping relationship is established by applying different levels of environmental stress (including but not limited to high temperature, low temperature, drought, waterlogging, shading, and nutrient deficiency) and biological stress (including insect feeding and pathogen infection) under controlled conditions, either individually or in combination. The changes in endogenous ethylene and abscisic acid content, photosynthetic rate, and transpiration rate of cotton leaves under each stress condition are measured. Simultaneously, the abscission time of detached leaves treated with a standard concentration of defoliant is measured. A multidimensional mapping surface is established with the comprehensive stress level as input and the relative change rate of abscission time as output. Correction coefficients are generated by interpolation calculations based on this mapping surface. The implementation details of this scheme are as follows: The foundation for establishing this quantitative mapping relationship lies in the design and execution of controlled stress experiments. Experiments were conducted in artificial climate chambers or controlled greenhouses, where ambient temperature, relative humidity, photoperiod, and soil moisture supply were kept constant to eliminate cross-interference from abiotic factors. The test materials were selected from the main varieties grown in the target cotton region and cultivated in pots. Stress treatment was initiated when the cotton plants reached the peak boll-forming stage.

[0043] In the treatment of pest feeding stress, third-instar larvae of the cotton bollworm were used as the standard feeding insect. Multiple feeding intensity gradients were established in the experiment, achieved by quantitatively releasing different numbers of larvae onto a single cotton plant and controlling the duration of feeding. This created several discrete intensity levels, ranging from a blank control with zero feeding pressure to complete feeding pressure. Multiple replicate plants were set up for each intensity level. After feeding ended, the pests were uniformly removed, and the actual percentage of leaf loss was recorded as the numerical anchor for that intensity level.

[0044] In terms of pathogen infection stress treatment, conidial suspensions were prepared using strains of cotton Verticillium wilt with high pathogenicity, and quantitative infection was carried out via root irrigation inoculation. Multiple infection intensity gradients were established in the experiment, and several discrete infection levels, ranging from a blank control to high infection pressure, were formed by adjusting the concentration of the conidial suspensions. After inoculation, plant symptoms were continuously observed. Once typical symptoms such as vascular bundle browning appeared, the infection level of each treatment group was confirmed based on the plant's disease incidence and disease index.

[0045] At the point when the stress effect was fully manifested, leaf samples were taken from plants in both the treatment and control groups. Mature functional leaves in the middle of the cotton plant were uniformly sampled, and immediately after collection, the leaves were subjected to low-temperature freezing to fix the endogenous hormone state. The target hormones measured were the content of aminocyclopropanecarboxylic acid, a key substance in the ethylene biosynthesis pathway, and abscisic acid. Enzyme-linked immunosorbent assay (ELISA) was used for detection, and a standard curve was constructed using corresponding standards. The quantification unit was the number of nanograms of the target substance per unit fresh weight of leaf. The measured values ​​of each treatment group were compared with the blank control group to obtain the rate of change in endogenous hormone content.

[0046] The defoliation response time was measured concurrently with the hormone assay. Intact leaflets with intact petioles were cut from each treatment group and the control group. The petiole tips were immersed in a mixed defoliant and ripening agent solution containing a known standard concentration of thiamethoxam and ethephon to simulate field application conditions. The treated leaves were placed in a constant temperature and humidity chamber for continuous observation. The abscission layer of the petiole was gently touched with tweezers at regular intervals, and the moment the petiole detached under slight external force was recorded. The time elapsed from immersion in the agent to detachment was calculated. The relative change rate of detachment time was obtained by comparing the detachment time of each stress treatment group with that of the healthy control group.

[0047] Based on all the experimental data mentioned above, a multidimensional mapping surface was constructed. Using the comprehensive stress intensity index as the input axis and the relative change rate of abscission time as the output axis, each experimental point was calibrated in a high-dimensional coordinate system. For the blank areas between experimental points, an interpolation algorithm was used for smoothing, specifically the Kriging interpolation method based on spatial correlation. This method can provide output estimates for any combination of input levels based on the distribution characteristics and variation patterns of known data points, thus forming a smooth mapping surface covering the continuous input space. This mapping surface implicitly contains the regulatory mechanism of endogenous hormones on drug efficacy response; only the stress level needs to be input to directly output the rate of change of abscission response.

[0048] By adopting a technical approach that traces back to the physiological roots of changes caused by comprehensive stress, and by constructing a physiological response quantitative correction model with the comprehensive stress intensity index as input and the leaf drop response sensitivity as output, the adjustment of the application time is based on the quantitative mapping relationship between changes in endogenous hormones in the plant and the shift in the pesticide response, rather than simply stacking features of pest and disease data. The application time obtained in this way is more in line with the actual pesticide response state of the plant.

[0049] In actual field operation, when the physiological response quantification correction model calls this mapped surface, it receives the current comprehensive stress intensity index (generated by fusing monitoring data on temperature, moisture, light, pests, and diseases) determined by the field monitoring system. Interpolation calculations are then performed on this surface, and the output value is the defoliation response sensitivity correction coefficient. This coefficient reflects the direction and magnitude of the cotton plant's response time to the defoliant / ripening agent relative to its healthy state under the current combined stress state, providing direct quantitative basis for subsequent compensation and correction of the application window.

[0050] The time-domain compensation correction is calculated by multiplying the correction factor by the base response time constant to obtain the time offset. When the correction factor indicates a delayed plant response to the pesticide, the candidate window for initial application time is shifted forward by this time offset on the time axis. Conversely, when the correction factor indicates an overreaction in the plant response to the pesticide, the candidate window for initial application time is shifted backward by this time offset on the time axis. The implementation details of this scheme are as follows: The implementation of time-domain compensation correction relies on the preparation of two prerequisite parameters: the baseline response time constant and the correction coefficient. The baseline response time constant is a time parameter pre-calibrated through standard field trials for a specific cotton variety under healthy, stress-free conditions. The calibration method is as follows: During the peak boll-forming stage of cotton plants, healthy plant populations are selected, and a conventional concentration of defoliant and ripening agent is sprayed under standard weather conditions. Timing begins from the spraying time, and the number of days it takes for the upper, middle, and lower leaves of the cotton plants to reach the predetermined shedding ratio is continuously observed and recorded. The average number of days from multiple replicate trials is taken as the baseline response time constant for that variety. This constant reflects the typical response time required for the defoliant and ripening agent to exhibit its expected effect from spraying under healthy conditions without any comprehensive stress, i.e., when temperature, moisture, light, pests, and diseases are all within suitable ranges.

[0051] The correction coefficient is derived from the output value generated by the physiological response quantification correction model based on real-time comprehensive stress monitoring data. This correction coefficient is expressed numerically. When the value equals the baseline value, it indicates that the plant is in a healthy state, and the pesticide response characteristics are consistent with the baseline response time constant. When the value is greater than the baseline value, it indicates that the plant's pesticide response is premature and intensified due to comprehensive stress. When the value is less than the baseline value, it indicates that the plant's pesticide response is delayed and sluggish due to comprehensive stress.

[0052] The time offset is calculated by the time compensation module. This module receives two parameters: a correction coefficient and a base response time constant. It performs a multiplication operation, and the product is the required time offset. Specifically, the difference between the correction coefficient and the baseline value is used as the effective offset factor. If the correction coefficient equals the baseline value, the difference is zero, the time offset is zero, and the initial application time candidate window does not need adjustment. If the correction coefficient deviates from the baseline value, the deviation is multiplied by the base response time constant to obtain the absolute value of the time offset in days.

[0053] After obtaining the time offset, the time compensation module determines the translation direction based on the comparison between the correction coefficient and the baseline value. When the correction coefficient indicates a delayed plant response to the pesticide, it means that after the current plant is treated within the original application window, the pesticide's onset time will be later than normal due to inhibited physiological metabolism. Therefore, earlier application is needed to ensure that the defoliation process is completed before harvest. In this case, the initial candidate application window is shifted forward by the time offset. Specifically, the forward shift operation involves subtracting the time offset from both the start and end dates of the window to obtain an earlier corrected application window.

[0054] When the correction coefficient indicates that the plant's response to the pesticide is excessive, it means that after applying the pesticide within the original application window, the abscission layer cells are already in a sensitive state, causing the pesticide to take effect earlier than normal. If the pesticide is still applied within the original window, it may force the immature bolls in the upper part of the plant to crack prematurely. In this case, the initial candidate application window is shifted backward by the specified time offset. Specifically, this shift involves adding the time offset to both the start and end dates of the window to obtain a later, corrected application window.

[0055] After the above translation operation, the time span of the corrected application window remains consistent with the initial application time candidate window, with only the overall time axis shifted. The translated corrected application window will be directly used as the input parameter for subsequent simulation and derivation stages, entering the dual closed-loop verification process.

[0056] The baseline response time constant was determined using a multi-plot field calibration test of the same cotton variety. Healthy cotton fields of the same main cotton variety, with similar soil types, normal weather conditions, and no comprehensive stress were selected and divided into multiple experimental plots. During the boll-forming stage of the cotton plants, a standard concentration of defoliant and ripening agent was sprayed uniformly. The number of days required for the leaves of the upper, middle, and lower parts of the cotton plants to reach the standard defoliation ratio was continuously observed and recorded daily. The average value of the observation results from multiple experimental plots was taken as the fixed baseline response time constant for the variety. Each cotton variety completed its own parameter calibration according to a unified field test procedure and stored it in the database. The baseline correction coefficient was set as the standard baseline value under the condition of no comprehensive stress and healthy cotton plant growth, and served as the dividing reference standard for judging the delay and overreaction of the agent response. The time offset is divided into gradients according to the deviation of the correction coefficient from the benchmark value. Multiple fixed-day offset gradients are set. The smaller the deviation, the fewer the number of days to shift; the larger the deviation, the more the number of days to shift. When the drug response is determined to be delayed, the initial drug application window is uniformly matched forward to the corresponding gradient number of days. When the drug response is determined to be excessive, the initial drug application window is uniformly matched backward to the corresponding gradient number of days. The start date and end date of the window are synchronously and equally shifted while keeping the window time span unchanged. The time domain compensation correction can be fully implemented by following the test calibration process, benchmark value setting rules and gradient shifting method.

[0057] This time-domain compensation and correction mechanism enables the adjustment of the application window to be quantitatively operated based on the actual physiological response state of the plant, and the correction range is directly related to the degree of comprehensive stress, thus realizing precise reverse constraint on the application time.

[0058] The simulation was conducted based on a pre-constructed threshold model of pesticide effect. This model uses ambient temperature, relative humidity, and correction coefficients generated by a physiological response quantification model as driving variables to simulate the dynamic coupling process between the defoliant release curve and the cotton plant's physiological response curve under a specific application window. It outputs the daily defoliation rate sequence and the cessation time of cotton boll dry matter accumulation. The implementation details of this scheme are as follows: The threshold model for pesticide effect is based on a segmented mathematical description of the mechanism of action of defoliating and ripening agents and the physiological response of cotton plants. The model consists of two coupled sub-modules: a pesticide release sub-module and a physiological response sub-module. The pesticide release sub-module describes the penetration, conduction, and activity decay processes of the pesticide after it is sprayed onto the leaf surface, while the physiological response sub-module describes the abscission process of cotton plant leaf abscission cells under the stimulation of the pesticide and the regulation of endogenous hormones, as well as the dynamic changes in the accumulation of dry matter in cotton bolls.

[0059] In the specific daily simulation, the model performs iterative calculations according to the following steps: First, the application day is set as day 0, the initial effective activity value of the pesticide is set to the standard value of 1.0, and the initial defoliation rate is set to 0%. Second, based on the average temperature of the day, a preset temperature-attenuation control rule is used for determination: when the temperature is below 25℃, the daily pesticide activity attenuation is 5%; when the temperature is between 25℃ and 30℃, the attenuation is 8%; when the temperature is above 30℃, the attenuation is 12%, thus calculating the effective residual activity value of the pesticide for the day. Third, the correction coefficient is multiplied by the effective residual activity value of the pesticide for the day to obtain the corrected pesticide effect intensity for the day. Fourth, based on the relative humidity of the day, a preset humidity-ablative cell disintegration rate control rule is used for determination: when the relative humidity is between 60% and 80%, the abscission cell disintegration rate is set to the baseline value of 1.0; when the relative humidity is below 60%, for every 10% decrease... The percentage point decreases the disintegration rate by 15%; when the relative humidity is higher than 80%, the disintegration rate increases by 10% for every 10 percentage point increase, thus obtaining the baseline value of the daily abscission cell disintegration rate; the fifth step is to multiply the corrected daily agent effect intensity by the daily abscission cell disintegration rate baseline value to obtain the actual leaf loss ratio for the day; the sixth step is to multiply the total number of leaves that did not fall off the previous day by the actual leaf loss ratio for the day to obtain the newly added leaf loss amount for the day, which is added to the total leaf loss rate; the seventh step is to use the remaining agent activity value at the end of the day as the initial value for the next day, and repeat steps two to six until the leaf loss rate increment for three consecutive days is less than 1% or reaches the preset simulation duration (usually 30 days after application); the eighth step is to monitor the cumulative leaf loss rate synchronously during the daily iteration process, and when the cumulative leaf loss rate reaches the preset canopy leaf loss threshold (e.g., 60%), the day is determined as the point at which cotton boll dry matter accumulation stops, and the dry matter accumulation calculation is stopped. Through the above steps, the model outputs a daily defoliation rate sequence from the start of pesticide application and the corresponding time when the accumulation of dry matter in cotton bolls ceases.

[0060] The two submodules are coupled and iteratively calculated on the timeline with a daily step. Starting from the beginning date of the modified application window, the ambient temperature and relative humidity data for the first day are input into the two submodules respectively. The drug efficacy release submodule outputs the effective activity value of the drug for that day, and the physiological response submodule outputs the abscission cell disintegration rate for that day. These two are coupled, using the effective activity value of the drug as the driving input for the abscission cell disintegration rate, to calculate the actual proportion of cotton plant leaves that day, i.e., the leaf drop rate for that day. After the work for the day ends, the total number of remaining leaves is used as the initial condition for the next day's calculation, and the drug efficacy release submodule simultaneously updates the remaining activity value after the drug activity decays.

[0061] The following day, the meteorological data for the new day is input into the model, and the above process is repeated. The difference is that the drug release submodule adds the influence of the new data to the decay of the remaining activity from the previous application, while the physiological response submodule calculates the rate of abscission cell disintegration under the premise of a reduction in the total amount of leaves. This process is repeated day by day until the predetermined simulation duration ends, thus obtaining a complete daily sequence of leaf drop rates under the modified application window.

[0062] While simulating the defoliation process, the model simultaneously calculates the point at which boll dry matter accumulation ceases. The calculation logic is as follows: when the cumulative defoliation rate on a given day in the daily defoliation rate sequence reaches a preset canopy leaf shedding threshold, it is determined that the photosynthetic leaf area of ​​the cotton plant is insufficient to maintain normal boll dry matter accumulation, and this day is marked as the point at which boll dry matter accumulation ceases. This point at which the accumulation ceases is used to estimate subsequent yield and quality changes; specifically, based on the difference between the point at which the boll opening is completed and the standard number of days required, the ratio of boll weight retention rate to the weight of fully mature bolls is calculated. Therefore, the pesticide effect threshold model, in addition to outputting the daily defoliation rate sequence, simultaneously outputs the point at which boll dry matter accumulation ceases, providing a quantitative basis for comparing yield, quality, and benefits within a double-validation boundary.

[0063] The pesticide effect threshold model sets the gradual trend of natural decay of the defoliant and ripening agent's efficacy based on the actual field efficacy variation patterns. Combining the conventional influences of field temperature and humidity, it defines the rate of efficacy decay at different temperature and humidity ranges, achieving phased simulation of the efficacy release process without complex function calculations. Based on the comprehensive stress level and cotton plant physiological state, it classifies multiple levels of abscission cell disintegration rate, automatically matching the corresponding rate level to different stress conditions. It also incorporates the influence of field air humidity on leaf cuticle permeability, synchronously fine-tuning the body's progression pace. The daily defoliation rate is calculated using daily step-by-step cumulative statistics. The method involves accumulating the percentage of leaf shedding each day starting from the first day of the modified application window, continuously extrapolating until the leaf shedding process stabilizes, forming a complete daily leaf shedding rate sequence. The percentage of effective photosynthetic leaves remaining in the cotton canopy is used as the practical criterion for determining the cessation of boll dry matter accumulation. When the cumulative leaf shedding in the simulation causes the percentage of photosynthetic leaf area to drop to the field-recognized critical level, that day is immediately determined as the moment when dry matter accumulation ceases. Based on this, the boll weight development termination point can be calculated. The entire process of pesticide effect simulation can be reproduced according to the pesticide efficacy attenuation level, disintegration rate classification, leaf shedding rate accumulation method, and cessation time determination criteria.

[0064] The defoliation rate threshold in this dual-verification boundary is set to the lowest defoliation rate level corresponding to the machine-harvested standard. The comprehensive yield and quality benefit threshold is jointly determined by the relative boll weight retention rate and the comprehensive fiber quality index, where the comprehensive fiber quality index is a weighted composite of fiber length, breaking strength, and micronaire value. The implementation details of this scheme are as follows: The dual-verification boundary setting aims to constrain the simulation results from two dimensions: defoliation effect and quantity-quality synergy. Only when the simulation results in both dimensions meet their respective boundary conditions is the current modified application window considered valid. The boundary thresholds for the two dimensions are preset during the system initialization phase, and their settings are based on agronomic standards for mechanized harvesting and national grading standards for cotton fiber quality, respectively.

[0065] The defoliation rate threshold directly corresponds to the minimum requirement for the degree of defoliation of cotton plants in mechanized harvesting operations. In mechanized harvesting practice, the process of cotton harvester spindles grasping and conveying cotton bolls requires that the number of residual leaves on the cotton plant be controlled below a certain level; otherwise, it will lead to an increase in the impurity content of seed cotton, an increased burden on the cleaning process, and a decrease in the grade of lint cotton. The defoliation rate threshold is the minimum defoliation rate level that should be achieved within a fixed number of days after pesticide application to meet the requirements of mechanized harvesting. The specific value of this threshold can be determined through a comparative harvesting experiment: treatment plots with different defoliation rates are set up in the same cotton-growing area, and the same model of cotton harvester is used for harvesting. The impurity content of seed cotton in each plot is measured, and the lowest defoliation rate corresponding to the seed cotton impurity content reaching the upper limit of the allowable standard for mechanized harvesting is taken as the threshold.

[0066] The comprehensive yield and quality benefit threshold is determined jointly by two components: relative boll weight retention rate and fiber quality comprehensive index. The relative boll weight retention rate reflects the effect of pesticide application time on preserving the final fullness of the cotton boll. Its benchmark value is set as the standard boll weight of the cotton variety under fully normal maturity conditions. In specific calculations, the time when dry matter accumulation in the cotton boll stops, output during the simulation phase, is compared with the standard accumulated temperature days required for the variety from flowering to boll opening. The proportion of dry matter accumulation completed is calculated, and this proportion is the value of the relative boll weight retention rate. When dry matter accumulation stops prematurely, the relative boll weight retention rate is lower than the full value; the specific deviation depends on the timing of the stoppage.

[0067] The comprehensive fiber quality index is a weighted composite of three indicators: fiber length, breaking strength, and micronaire value. These three indicators are key parameters for measuring the processing value and spinning performance of cotton fibers. Fiber length is measured in millimeters and reflects the fiber's potential for yarn count. Breaking strength is measured in centineutets per tex and reflects the fiber's ability to resist breakage during spinning. The micronaire value is a comprehensive indicator reflecting fiber fineness and maturity; a value that is too high indicates that the fiber is too coarse and over-mature, while a value that is too low indicates that the fiber is too fine and under-mature. Each of the three indicators has an optimal range, with the boundaries of the ranges set according to the corresponding grade in the national grading standards for cotton fiber quality.

[0068] The composite fiber quality index is synthesized as follows: The actual values ​​of fiber length, breaking strength, and micronaire value are measured separately. The actual values ​​of each of the three indicators are then compared with the ideal values ​​in the superior grade range to obtain the individual index for each indicator. Finally, the three individual indices are weighted and summed according to preset weighting coefficients. These weighting coefficients are set based on the contribution of each indicator to the spinning value; fiber length and breaking strength are typically assigned higher weights, while micronaire value is assigned a relatively lower weight. The sum of the weights of the three is the full value. The final calculated composite fiber quality index reflects the degree to which the overall fiber quality approaches the optimal grade.

[0069] In the joint determination step of the dual-verification boundary, the system compares the defoliation rate of a fixed number of days after application with the defoliation rate threshold in the defoliation rate sequence obtained from simulation with the defoliation rate threshold. Simultaneously, it compares the comprehensive yield-quality benefit value calculated jointly based on the relative boll weight retention rate and the fiber quality comprehensive index, derived from the cessation of boll dry matter accumulation, with the comprehensive yield-quality benefit threshold. Both comparison results must simultaneously meet their respective threshold conditions for the modified application window to be deemed valid. If either comparison result fails to meet the condition, the adjustment of the correction coefficient and the iterative optimization process are triggered.

[0070] The defoliation rate threshold is directly set as the minimum acceptable level using the general industry standard for mechanized cotton harvesting in China, serving as the minimum defoliation control baseline allowed for mechanized harvesting operations. All varieties of upland cotton uniformly adhere to this fixed standard. The relative boll weight retention rate threshold uses the normal mature boll weight of the variety standard as a reference benchmark, combined with the allowable yield loss range in field production, to delineate a reasonable minimum boll weight retention control standard to ensure that yield does not experience significant reduction. The fiber quality comprehensive index selects three core indicators: fiber length, breaking strength, and micronaire value, allocating fixed weights according to their actual value contribution to the spinning industry, prioritizing these indicators. Fiber length and breaking strength are given higher reference weights, while micronaire value is configured with appropriate auxiliary weights. The weight ratios are fixed and applicable to all major cotton varieties. The comprehensive quality and benefits are evaluated using a graded comparison method. The measured levels of the three indicators are compared with the national cotton fiber quality grading standards to determine the grades. The comprehensive grades are then combined with preset weights to form a directly comparable comprehensive yield and quality benefit level. This level is then matched and verified against preset benefit thresholds. The setting and comparison of dual verification boundaries can be completed according to industry machine harvesting standards, boll weight control baselines, fixed weight allocation rules, and comprehensive evaluation methods.

[0071] By using correction coefficients as inverse constraints to perform time-domain compensation corrections on the initial prediction window, and introducing a dual closed-loop verification mechanism that integrates the defoliation rate boundary and the yield, quality, and benefit boundary, the decision on pesticide application time is transformed from an open-loop, single-prediction process to a closed-loop, iterative optimization process. When the simulation results do not simultaneously satisfy both boundary conditions, the system automatically feeds back the deviation to the correction coefficients and triggers a secondary correction until an application window that satisfies both constraints is output. This avoids the direct output of erroneous decisions and maintains the stability and reliability of decisions under complex and ever-changing integrated field stress environments.

[0072] The first prediction model is a hybrid neural network model that integrates gradient boosting decision trees and long short-term memory networks. Gradient boosting decision trees are used to process growth and development data, soil moisture data, and variety genetic characteristics data to extract nonlinear combination features. Long short-term memory networks are used to process time series meteorological data to extract time-varying features. The two types of features are concatenated and output as preliminary candidate windows for pesticide application time through a fully connected layer. The implementation details of this scheme are as follows: The hybrid neural network model is architecturally divided into three functional layers: a static feature extraction layer, a temporal feature extraction layer, and a feature fusion output layer. The static feature extraction layer uses gradient boosting decision trees as its core algorithm unit, while the temporal feature extraction layer uses long short-term memory networks as its core algorithm unit. The outputs of the two are concatenated and mapped in the feature fusion output layer.

[0073] The static feature extraction layer receives three types of structured data: cotton growth and development data, soil moisture data, and varietal genetic characteristic data. The growth and development data specifically covers the number of fruiting branches per plant, the number of bolls per plant, and the boll opening rates of upper, middle, and lower bolls recorded at multiple survey time points throughout the cotton's growth period. The soil moisture data specifically covers the daily soil volumetric water content in the 0-20 cm and 20-40 cm soil layers. The varietal genetic characteristic data specifically covers parameters such as the boll-forming period in days, basic defoliation tolerance level, and fiber quality formation cycle.

[0074] Before entering the gradient boosting decision tree, the aforementioned data undergoes data alignment processing, aggregating data from different collection frequencies into a structured feature vector with daily granularity. The gradient boosting decision tree employs a structure of multiple decision trees stacked sequentially. Each newly added decision tree uses the prediction residuals of all preceding trees as its fitting target, gradually reducing the overall prediction error through gradient descent. During the training phase, the gradient boosting decision tree learns the nonlinear combination relationships and interaction patterns between each input feature and the leaf response; these patterns are solidified in the model parameters as branch node conditions of the tree. In the prediction phase, a daily-granularity feature vector is input, and the gradient boosting decision tree propagates it down the branch paths of each decision tree, ultimately outputting a nonlinear combination feature vector reflecting the comprehensive static characteristics of that day.

[0075] The time-series feature extraction layer receives time-series meteorological data as input. Specifically, the meteorological data includes daily average temperature, daily maximum temperature, daily minimum temperature, daily precipitation, daily sunshine hours, and the effective accumulated temperature calculated from the temperature data. The time window length of the input sequence is set according to the retrospective period required for cotton defoliation and ripening decisions; typically, continuous meteorological data from several days prior to the pesticide application decision date is taken as a time-series sample.

[0076] Long Short-Term Memory (LSTM) networks consist of multiple interconnected memory units, each containing three gating structures: a forget gate, an input gate, and an output gate. The forget gate determines which information from the previous time step's state needs to be discarded, the input gate determines which information from the current input data needs to be stored in the unit's state, and the output gate determines which information from the current unit's state needs to be output as a hidden state. Through the coordinated operation of these gating mechanisms, LTM networks can retain key meteorological evolution trend information while forgetting irrelevant short-term fluctuation noise when processing long-term time series. The input meteorological time series is processed step-by-step through multiple layers of LTM networks, and the hidden state vector output at the last time step is a time-varying feature vector that captures the long-term dependencies throughout the entire input sequence.

[0077] The feature fusion output layer concatenates the nonlinear combined feature vector output from the static feature extraction layer with the time-varying feature vector output from the temporal feature extraction layer, forming a fused feature vector with a dimension equal to the sum of the dimensions of the two. This fused feature vector is input into a fully connected layer, which consists of multiple neurons. Each neuron is connected to the elements of the fused feature vector through a weight matrix, and the output is passed to the next layer after being mapped by an activation function. The output of the fully connected layer is finally transformed into a preliminary candidate window for drug administration, consisting of the start and end dates. This window only reflects the influence of abiotic factors on the defoliation process; subsequent adjustments and verifications based on the physiological response quantification correction model and a dual closed-loop verification process will be made to integrate the stress dimensions.

[0078] By fusing the nonlinear combination features extracted by gradient boosting decision trees with the meteorological time-series features extracted by long short-term memory networks, the initial application window can comprehensively reflect multi-dimensional information such as variety genetic characteristics, soil moisture, growth and development progress, and meteorological evolution patterns. On this basis, physiological response corrections based on comprehensive stress dimensions are then superimposed, forming a complete decision-making chain from abiotic factors to the fusion of biotic and abiotic factors, with each link having a clear division of labor and complementing each other.

[0079] After the system performs dual closed-loop verification and outputs the final application time, this time data exists in the system memory in the form of date and time, specifically including the time window range formed by the suggested spraying start and end dates. This time data needs to be processed by the instruction conversion module to be converted into variable spraying control instructions that can be recognized and executed by the field operation equipment in order to drive the actual field operation process.

[0080] The instruction conversion module receives two types of data at its input. The first type is the final application time window output by the system, including the start and end dates. The second type is auxiliary parameters associated with the current application decision, including the correction coefficient value used in the last simulation during the iterative optimization process, and the spraying operation parameters corresponding to that correction coefficient retrieved from the preset scheme library. Specifically, the spraying operation parameters include the unit area application rate of the defoliant and ripening agent, the pesticide solution preparation ratio, and the spraying speed.

[0081] The instruction conversion module contains a decision logic unit. This unit first compares the start date of the application time window with the current date. If the current date falls within the application time window, it determines that a spraying instruction should be generated immediately. Next, the decision logic unit adjusts the spraying parameters based on the value of the correction coefficient. When the correction coefficient indicates that the plant is in a delayed pesticide response state, the decision logic unit increases the spraying amount based on the standard application rate per unit area, with the increase proportional to the offset of the correction coefficient, to compensate for the decreased pesticide absorption efficiency caused by inhibited plant metabolism. When the correction coefficient indicates that the plant is in an overactive pesticide response state, the decision logic unit decreases the spraying amount based on the standard application rate per unit area, with the decrease also proportional to the offset of the correction coefficient, to avoid abnormal shedding caused by oversensitivity of the plant's abscission layer cells.

[0082] After the decision-making logic unit completes the parameter adjustments, it encodes and packages the adjusted application rate, pesticide concentration ratio, and spraying speed, along with the start date of the application time window, according to the communication protocol of the field operation equipment. The field operation equipment is typically a self-propelled boom sprayer or an agricultural unmanned aerial vehicle equipped with a variable-rate spraying system. This type of equipment is equipped with an onboard controller and a global navigation satellite positioning receiver module. The command conversion module establishes a data connection with the onboard controller via a wireless communication link, sending the coded variable-rate spraying control commands to the onboard controller.

[0083] The data packet of the variable spraying control command contains the following information fields: the coordinates of the four boundaries of the work area, used to delineate the electronic fence range of the spraying operation; the target spraying date and the start time of the operation on that day; a spatial distribution prescription map of the amount of pesticide applied per unit area, where the amount of pesticide applied to each grid cell is converted by a correction coefficient to ensure that the spraying intensity in different areas matches the actual stress state of the cotton plants in that area; the spray boom height or flight height parameter; and the travel speed parameter.

[0084] After receiving the variable spraying control command, the vehicle-mounted controller of the field operation equipment parses the data packet and completes a self-check. Upon reaching the preset start time, the equipment starts and autonomously travels along the route and speed specified in the prescription map. Simultaneously, the variable spraying system adjusts the nozzle flow rate in real time based on the pesticide dosage values ​​of each grid cell on the prescription map, achieving on-demand variable spraying. After the spraying operation is completed, the vehicle-mounted controller records the actual spraying trajectory, the actual pesticide dosage at each location, and the operation completion time. This operation feedback data is then fed back to the system via a wireless communication link for subsequent field management record archiving.

[0085] Through the complete conversion and execution process from application time to variable spraying control instructions, the application decision results output by the system can be automatically and accurately implemented in the field operation equipment, avoiding the delay error and application amount deviation that may be introduced by manual operation, so that the best application plan obtained through closed-loop iterative optimization under comprehensive stress can be accurately executed.

[0086] The finalized application time window is converted into a variable spraying control command and sent to the field operation equipment. The dosage per unit area in the command is dynamically adjusted according to a correction coefficient, ensuring that the spraying intensity matches the actual stress state of the plants. This guarantees both defoliation and pesticide consumption. The entire decision-making and execution process achieves fully automated integration from data acquisition, model calculation, closed-loop verification to precise spraying, reducing the uncertainty and operational delays caused by human experience intervention.

[0087] The collected growth and development data include the flowering date and boll opening rate of cotton bolls in different parts of the plant; meteorological data includes daily average temperature, maximum temperature, minimum temperature and effective accumulated temperature; and varietal genetic characteristic data includes boll opening parameters and defoliation tolerance parameters.

[0088] Data on cotton growth and development were collected through fixed-point, timed field surveys. Fixed survey points were established in the target cotton field using either the diagonal method or the five-point sampling method. Each sampling point was marked with multiple representative cotton plants as long-term observation targets. Starting from the cotton flowering period, investigators observed and recorded data on the marked cotton plants every fixed number of days. The observation involved tagging each boll at each fruiting branch node on the cotton plant, noting the flowering date of that boll. The cotton plant was divided into three parts according to node height: upper, middle, and lower. The upper part corresponds to the terminal fruiting branch, the lower part to the first to third fruiting branches, and the middle part to the fruiting branches between the two. At each survey, the number of marked bolls at each part and the developmental stage of each boll were recorded. When the boll shell cracked open, exposing the seed cotton fibers, it was recorded as boll opening. Based on the ratio of the number of bolls that had opened to the total number of flowering and boll-forming plants at each part, the boll opening rates for the upper, middle, and lower parts were calculated. The boll opening rate data from these three locations reflect the boll opening progress at different maturity levels of the cotton plant, providing spatial stratification information for the subsequent first prediction model to determine the starting point and rhythm of the overall defoliation process. The flowering date data is used to trace the development duration of each cotton boll on the timeline, providing a boll age benchmark for calculating the cessation of dry matter accumulation in the cotton boll.

[0089] Meteorological data collection is performed continuously by automatic weather stations installed in the target cotton fields. These stations are equipped with temperature, humidity, rainfall, and sunshine sensors. Each sensor samples at preset intervals, and raw data is recorded and stored hourly. The daily average temperature is the arithmetic mean of all hourly temperature samples for that day. The daily maximum temperature is the maximum value among the hourly temperature samples for that day, and the daily minimum temperature is the minimum value among the hourly temperature samples for that day. These three temperature data collectively describe the level, amplitude, and extreme conditions of the daily temperature, providing time-varying temperature characteristics information for the Long Short-Term Memory (LSTM) network. Effective accumulated temperature is calculated based on the lower limit temperature for cotton development. The difference between the daily average temperature and the lower limit temperature is calculated daily. When the difference is positive, it is accumulated and added to the effective accumulated temperature sequence; when the difference is zero or negative, the effective accumulated temperature for that day is recorded as zero. The cumulative effective accumulated temperature reflects the total amount of heat resources obtained by the cotton plant since a certain developmental starting point and is a key heat indicator for judging the progress of boll development and maturation. The daily effective accumulated temperature, together with the air temperature data, constitutes a continuous time series. After being input into the long short-term memory network, the time series characteristics such as the cumulative effect and fluctuation pattern of air temperature can be extracted, providing meteorological constraints for the estimation of the preliminary application time candidate window.

[0090] The genetic characteristics data of cotton varieties were obtained from the varietal characteristic parameters registered in the variety approval database and the measured results of multi-year, multi-location variety comparison trials. The boll-forming period parameter is the inherent number of development days required for a cotton variety from flowering to boll opening under standard temperature conditions. This parameter was measured through multi-year variety comparison trials conducted in the same ecological region. In the trials, a large number of cotton bolls flowering at the same time were selected from each tested variety and marked. The boll-opening status was observed daily, and the average number of days from flowering to boll opening for each part of the plant was recorded as the boll-forming period parameter for that variety. The defoliation tolerance parameter reflects the natural differences in the sensitivity of varieties to defoliating ripening agents. This parameter was measured through detached leaf tests using standard concentrations of the agent. Specifically, mature functional leaves of each tested variety were collected, and the petioles were immersed in a defoliating ripening agent solution of the same concentration. The average duration of leaf abscission was recorded, and the varieties were classified into multiple defoliation tolerance levels based on the duration. Boll-forming period parameters and defoliation tolerance parameters are two types of inherent genetic attributes at the variety level. These parameters are stored numerically in a variety genetic characteristic database. Each time a pesticide application time prediction is performed, the system retrieves the corresponding parameter values ​​from the database based on the actual planted variety name in the target cotton field. These parameter values ​​are input as part of a static feature vector into a gradient boosting decision tree model to extract the nonlinear combination relationship between variety characteristics and abiotic factors regarding defoliation response. This allows the initial pesticide application time candidate window to adapt to the differences in developmental rhythms and pesticide response characteristics among different varieties.

[0091] A device for predicting the application time of pesticides for cotton defoliation and ripening is constructed at the hardware level by integrating a data processing unit, a data storage unit, a communication interface unit, and a human-machine interaction unit into a single cabinet or embedded industrial control chassis. The core of the data processing unit is a central processing unit (CPU), which is a general-purpose processor with multi-core parallel computing capabilities. Its clock speed and cache capacity are sufficient to support the inference operations of the hybrid neural network model and the real-time computing needs of multiple simulations in the iterative closed loop. The CPU is connected to the data storage unit via a system bus. The data storage unit includes random access memory (RAM) and non-volatile solid-state memory (SSD). The RAM is used to temporarily store raw data received from field sensors, intermediate feature vectors during model inference, daily defoliation rate sequences generated by simulations, and updated correction coefficients generated during iterative optimization. The SSD is used to persistently store the operating system, database management system, network parameter files of the first prediction model, mapping surface data files of the physiological response quantification correction model, parameter files of the pesticide effect threshold model, a variety genetic characteristic database, and historical pesticide application decision records.

[0092] The communication interface unit includes a wired Ethernet interface and a wireless communication module. The wired Ethernet interface is used to establish a local area network connection with the automatic weather station data logger and soil moisture sensor data logger deployed in the field, and to periodically retrieve hourly meteorological data and soil volumetric water content data at a predetermined frequency. The wireless communication module has a built-in cellular mobile communication module and a local wireless communication module. The cellular mobile communication module is used to establish a data link with remote farmland environment and biological monitoring terminals (including automatic weather stations, soil moisture sensors, pest pheromone traps, spore traps, etc.) to receive real-time monitoring data of integrated stress in the field; the local wireless communication module is used to conduct short-range data exchange with the vehicle-mounted controller on the field operation equipment, send variable spraying control commands and receive operation feedback data.

[0093] The human-machine interface unit includes a touch screen and input peripherals. The touch screen presents the user with variety information for the current cotton field, real-time weather and soil data, comprehensive stress level assessment results, the final pesticide application time window output by the device, and suggested spraying parameters in a graphical interface. The input peripherals receive user input commands for variety selection, field boundary coordinates, and parameter configuration.

[0094] The computer program running in the processor is logically divided into a data preprocessing module, a first prediction module, a physiological response quantification correction module, a time compensation module, a simulation and deduction module, a dual verification module, and an instruction conversion module. The calling order and interaction relationship of each module are uniformly scheduled by the main control program.

[0095] After the device is powered on, the main control program first initializes each module, including loading the network parameters of the first prediction model, the mapped surface data, and the variety genetic characteristics database from the non-volatile solid-state memory to the random access memory. Subsequently, the data preprocessing module continuously acquires raw data from various data sources through the communication interface unit, performs missing value imputation and normalization on growth and development data, meteorological data, and soil moisture data, and retrieves the corresponding boll-forming period parameters and defoliation tolerance parameters from the variety genetic characteristics database according to the variety name. At the same time, the temperature deviation, soil moisture deficit, number of consecutive rainy days, and pest and disease occurrence levels in the comprehensive stress monitoring data are converted into corresponding temperature stress level, water stress level, light stress level, and pest and disease stress level according to the preset stress level judgment criteria, and then fused to generate a comprehensive stress intensity index, completing the structured organization of all input data.

[0096] The first prediction module calls the loaded hybrid neural network model, inputting processed growth and development data, soil moisture data, and varietal genetic characteristic parameters into the gradient boosting decision tree submodule to extract nonlinear combined feature vectors. It then inputs the meteorological time series data into the long short-term memory network submodule to extract time-varying feature vectors. The two are concatenated and output as preliminary application time candidate windows via a fully connected layer. The physiological response quantification and correction module receives comprehensive stress indicators, including temperature deviation, soil moisture deficit, number of days with insufficient light, and pest and disease severity. It performs interpolation on the loaded multidimensional mapping surface to generate correction coefficients and passes these coefficients to the time compensation module. The time compensation module calculates the time offset based on the correction coefficients and the basic response time constant, performing a translation operation on the preliminary application time candidate windows to obtain the corrected application window. The simulation and deduction module receives the corrected application window, daily weather forecast data or recent weather data, and correction coefficients. It calls the pesticide effect threshold model to perform daily step-size coupling operations, outputting the daily defoliation rate sequence and the time when cotton boll dry matter accumulation ceases. The dual verification module compares the simulation results with preset thresholds for leaf drop rate and comprehensive yield, quality, and benefits. When both comparison results meet the threshold conditions, the current corrected application window is marked as valid and output to the instruction conversion module. When any comparison result does not meet the conditions, the dual verification module feeds back the deviation to the physiological response quantification correction module, triggering adjustments to the correction coefficients and iterative calculations in subsequent modules until the threshold conditions are met. The instruction conversion module combines the final output application time window with the correction coefficients used in the last iteration to generate a variable spraying control instruction containing an application rate prescription map and operational parameters, which is then sent to the vehicle controller of the field operation equipment via the communication interface unit.

[0097] This device integrates all data acquisition, model calculation, closed-loop verification, and instruction generation functions involved in the method into a single hardware entity. Its processor and memory work together to complete the aforementioned steps, thereby achieving fully automated prediction and precise execution of the cotton defoliation and ripening time application time.

[0098] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the application time of pesticides for cotton defoliation and ripening, applied to the coordinated management of cotton quantity and quality, including: Acquire cotton growth and development data, meteorological data, soil moisture data, and variety genetic characteristics data, and construct a first prediction model based on multi-source data to output preliminary candidate windows for pesticide application time; Its characteristic is that it includes the following steps: Acquire comprehensive physiological and environmental stress data for cotton crops; the comprehensive stress data shall include at least the occurrence level of pests and diseases, as well as one or more of temperature stress, water stress, and light stress. A physiological response quantitative correction model is constructed. This model takes the comprehensive stress data as input and outputs a correction coefficient that characterizes the degree of shift in the plant's sensitivity to defoliation ripening agent caused by the current environmental and biological comprehensive stress. The value of the correction coefficient is positively correlated with the fluctuation level of endogenous hormones and the degree of change in metabolic activity in the plant. Using the correction coefficient as a reverse constraint, the initial candidate window for application time is corrected by time domain compensation to obtain the corrected application window; The modified application window is simulated and analyzed to predict the defoliation process curve and the corresponding changes in yield and quality after application. A dual verification boundary is set, consisting of a defoliation rate threshold and a comprehensive yield and quality benefit threshold. The defoliation process curve and yield and quality changes obtained from the simulation are compared with the dual verification boundary. When the comparison results simultaneously meet the defoliation rate threshold and the comprehensive yield and quality benefit threshold, the modified application window will be output as the final application time. When the comparison results do not meet the requirements simultaneously, the deviation in leaf removal rate and the deviation in overall yield, quality and benefits are mapped inversely to the adjustment increment of the correction coefficient, and a secondary compensation correction and simulation are triggered for the application window until the comparison results fall into the double verification boundary.

2. The method for predicting the application time of pesticides for cotton defoliation and ripening according to claim 1, characterized in that, The physiological response quantitative correction model is based on the fluctuations in endogenous hormones and changes in metabolic activity in cotton plants caused by comprehensive stress. The correction coefficient is generated by a pre-established quantitative mapping relationship between multi-level indicators reflecting the intensity of comprehensive stress and the sensitivity to defoliation response. The magnitude of the correction coefficient is proportional to the combined characterization of the plant's delayed response time and the magnitude of the overreaction caused by the combined stress.

3. The method for predicting the application time of pesticides for cotton defoliation and ripening according to claim 2, characterized in that, The quantitative mapping relationship is obtained in the following manner: Under controlled conditions, different degrees of environmental and biological stresses were applied separately or in combination, and the changes in endogenous ethylene and abscisic acid content, photosynthetic rate and transpiration rate of cotton leaves were measured under each stress condition. Simultaneously, the shedding time of detached leaves under treatment with standard concentration defoliating and ripening agents was measured. A multidimensional mapping surface is established with the comprehensive stress level as input and the relative change rate of shedding time as output. The correction coefficients are generated by interpolation based on the mapped surface.

4. The method for predicting the application time of pesticides for cotton defoliation and ripening according to claim 1, characterized in that, The specific method for the time-domain compensation correction is as follows: The time offset is obtained by multiplying the correction coefficient by the basic response time constant; When the correction factor indicates a delayed plant response to the pesticide, the initial candidate window for pesticide application time is shifted forward by the time offset. When the correction coefficient indicates that the plant is overreacting to the agent, the preliminary candidate window for application time is shifted backward along the time axis by the time offset.

5. The method for predicting the application time of pesticides for cotton defoliation and ripening according to claim 1, characterized in that, The simulation is based on a pre-constructed threshold model of drug action effect; The drug effect threshold model uses key environmental factors and the correction coefficient as driving variables to simulate the dynamic coupling process between the defoliant release curve and the cotton plant physiological response curve under a specific application window, and outputs the daily defoliation rate sequence and the time when the accumulation of dry matter in cotton bolls ceases.

6. The method for predicting the application time of pesticides for cotton defoliation and ripening according to claim 1, characterized in that, The leaf removal rate threshold in the dual verification boundary is set to the lowest leaf removal rate level corresponding to the machine harvesting standard. The threshold for comprehensive benefits of yield and quality is determined by the relative boll weight retention rate and the comprehensive fiber quality index, wherein the comprehensive fiber quality index is a weighted composite of fiber length, breaking strength and micronaire value.

7. The method for predicting the application time of pesticides for cotton defoliation and ripening according to claim 1, characterized in that, The first prediction model is a hybrid neural network model that integrates gradient boosting decision trees and long short-term memory networks; The gradient boosting decision tree is used to process the growth and development data, soil moisture data, and variety genetic characteristic data to extract nonlinear combination features. The long short-term memory network is used to process the time series of the meteorological data and extract time-varying features. The nonlinear combined features and the time-varying features are concatenated and then output as the preliminary candidate window for drug administration through a fully connected layer.

8. The method for predicting the application time of pesticides for cotton defoliation and ripening according to any one of claims 1 to 7, characterized in that, The method also includes converting the final output application time into a variable spraying control command and sending it to the field operation equipment to perform precise spraying of the defoliant and ripening agent.

9. The method for predicting the application time of pesticides for cotton defoliation and ripening according to claim 1, characterized in that, The cotton growth and development data include the flowering date and boll opening rate of bolls in different parts of the plant; The meteorological data includes daily average temperature, maximum temperature, minimum temperature, and effective accumulated temperature; The genetic characteristics data of the variety include boll-forming period parameters and defoliation tolerance parameters.

10. A device for predicting the application time of pesticides for cotton defoliation and ripening, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the steps of the method as described in any one of claims 1 to 9.

Images