Cotton high temperature resistance prediction system using multi-source data fusion

Abstract

The invention, which relates to the technical field of agricultural information, discloses a cotton high temperature resistance prediction system using multi-source data fusion, comprising a data acquisition module, a data processing module, a model construction module and a decision output module. The data acquisition module is used for acquiring multi-source heterogeneous data of a target area. According to the cotton high temperature resistance prediction system applying multi-source data fusion, through deep fusion of multi-source heterogeneous data and construction of a dynamic knowledge graph, accurate description of an environment-crop-soil complex interaction relationship under cotton high temperature stress is realized. A physiological development time driven mechanism model and an LSTM-XGBoost data driven model are innovatively coupled, a multi-scale prediction system formed from organ development to yield quality is established, and the dynamic cumulative influence of high temperature on the cotton shedding rate, yield composition and fiber quality can be quantified. The problems of inaccurate early warning and insufficient prediction dimension caused by single data and mechanism deficiency of a traditional method are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of agricultural information technology, specifically to a cotton high-temperature resistance prediction system that utilizes multi-source data fusion. Background Technology

[0002] In the field of agricultural information technology, predicting cotton's high-temperature tolerance is a crucial link in ensuring stable cotton production and improving planting efficiency in cotton-growing areas. Existing cotton high-temperature stress monitoring and early warning systems often rely on single data sources or simplified models. For example, they may rely solely on temperature thresholds from meteorological data for static early warning, or use traditional statistical models to roughly estimate yield losses. These methods have inherent limitations: First, due to the failure to integrate multi-source heterogeneous data, existing systems cannot construct a comprehensive dynamic feature system reflecting the interaction between the environment, cotton, and soil, leading to inaccurate identification of the cumulative effects of high-temperature stress. Second, existing models often lack physiological mechanism drivers and fail to couple the dynamics of cotton growth and development with the mechanisms of yield and quality formation, thus making it difficult to quantify the dynamic impact of high temperatures on cotton shedding rate, seed cotton yield, fiber quality, and yield loss rate. This deficiency manifests itself in several ways: In areas prone to high temperatures, existing early warning systems frequently produce false alarms or missed alarms because they neglect key factors such as soil moisture dynamics and cotton photosynthetic physiology. Furthermore, the early warning level classification is coarse and lacks the visual decision support provided by red, yellow, and green color-coded levels. Simultaneously, due to the lack of multi-dimensional feature fusion and mechanistic modeling, current technologies struggle to generate personalized prescriptions for varietal heat tolerance ranking and field regulation, leading to generalized and inefficient measures by growers in response to high-temperature disasters. This problem has become a major bottleneck restricting the digitalization and precision development of the cotton industry, urgently requiring breakthroughs in deep multi-source data fusion and dynamic prediction capabilities. Summary of the Invention

[0003] The purpose of this invention is to provide a cotton high-temperature resistance prediction system utilizing multi-source data fusion to address the problems mentioned in the background section. To solve the aforementioned technical problems, this invention provides the following technical solution:

[0004] A cotton high-temperature tolerance prediction system utilizing multi-source data fusion includes a data acquisition module, a data processing module, a model building module, and a decision output module. The data acquisition module is used to acquire multi-source heterogeneous data of a target area, which includes cotton-growing regions in Xinjiang, the Yellow River Basin, and the Yangtze River Basin. The multi-source heterogeneous data includes meteorological data, crop data, soil data, and environmental data. The meteorological data includes basic meteorological parameters and high-temperature characteristic parameters. The basic meteorological parameters include temperature, precipitation, light intensity, wind speed, and humidity. The high-temperature characteristic parameters include high-temperature intensity, duration, and frequency of extreme high temperatures. The crop data includes cotton variety genetic parameters, photosynthetic physiological dynamics, seed cotton yield composition, and fiber quality indicators. The soil data includes soil moisture dynamics, soil fertility, and soil microbial community. The environmental data includes humidity and light intensity.

[0005] The data processing module is connected to the data acquisition module and is used to perform dynamic data cleaning and heterogeneous integration of multi-source heterogeneous data. It also constructs an environment-cotton-soil interaction feature knowledge graph through multimodal feature fusion technology to realize the structured representation and dynamic updating of multi-factor coupling relationships.

[0006] The model building module is connected to the data processing module. It adopts a deep ensemble learning framework based on long short-term memory network and extreme gradient boosting tree, and couples a dynamic model of cotton boll growth and development driven by physiological development time and a model of yield and quality formation mechanism to establish a multi-scale model from physiological mechanism to phenotypic performance. This model is used to quantify the dynamic cumulative impact of high temperature on cotton shedding rate, seed cotton yield, fiber quality and yield loss rate. Fiber quality includes length, strength and micronaire value.

[0007] The decision output module is connected to the model building module and is used to output red, yellow and green three-color level early warning of high temperature stress. It also performs multi-dimensional quantitative ranking of the heat resistance of varieties based on yield, quality and stress resistance weights. At the same time, it combines knowledge graph and model output to generate personalized prescription suggestions for field stress resistance regulation, including quantitative schemes for irrigation amount and fertilizer formula. The system realizes high temperature resistance prediction and decision support for cotton through multi-source data fusion and mechanism model driving.

[0008] Preferably, the dynamic data cleaning adopts a cleaning algorithm based on data quality and consistency, including imputation of missing data, identification and correction of outliers, and time series alignment of multi-source data; the multimodal feature fusion technology integrates key features from meteorological, crop, soil and environmental data through feature extraction and feature selection, and constructs an interactive feature knowledge graph based on graph structure, where nodes represent environmental factors, cotton physiological state or soil properties, and edges represent the interaction relationships between factors.

[0009] Preferably, the physiological development time-driven cotton boll growth and development dynamic model is based on time series data of cotton physiological development stages to simulate the growth and development process of cotton bolls under different high temperature conditions; the yield and quality formation mechanism model combines cotton photosynthetic physiological dynamics and seed cotton yield composition to quantify the impact of high temperature on fiber quality indicators; in the deep ensemble learning framework, long short-term memory network is used to capture temporal dependencies, extreme gradient boosting tree is used to strengthen feature selection and nonlinear mapping, and multi-scale prediction is achieved through model fusion.

[0010] Preferably, the high-temperature stress three-color warning system in the decision output module is divided based on the high-temperature impact index output by the model, where red represents high-risk stress, yellow represents moderate-risk stress, and green represents low-risk stress; the multi-dimensional quantitative ranking of the heat resistance of varieties is based on a weighted scoring method, using yield loss rate, fiber quality change, and stress resistance indicators as ranking criteria.

[0011] Preferably, the generation of personalized prescription recommendations includes analyzing soil moisture dynamics, crop physiological state, and environmental factors based on knowledge graphs and model outputs, recommending specific values ​​for irrigation amount and fertilizer formula, and presenting them to the user through a visual interface.

[0012] Preferably, the construction of the interactive feature knowledge graph also includes a dynamic update mechanism, which adjusts the nodes and edges in the graph through real-time data input to reflect changes in the interaction relationship between the environment, cotton, and soil.

[0013] Preferably, the training process of the deep ensemble learning framework includes training the model using historical multi-source data and optimizing the model parameters through cross-validation to ensure prediction accuracy and robustness.

[0014] Preferably, the system further includes a user interface module for receiving field management data input by the user and displaying the warnings and prescription suggestions from the decision output module in a graphical manner.

[0015] Preferably, the acquisition of the multi-source heterogeneous data is achieved through sensor networks and remote monitoring equipment, including weather stations, soil probes and crop physiological sensors, and the data acquisition frequency is dynamically adjusted according to the cotton growth stage.

[0016] Preferably, the multi-scale model output in the model building module includes the cumulative impact curve of high temperature on cotton shedding rate, predicted seed cotton yield, trend of fiber quality index changes, and yield loss rate estimate. These outputs are used to support the early warning and prescription generation of the decision-making module.

[0017] This invention provides a cotton high-temperature resistance prediction system using multi-source data fusion. It has the following beneficial effects:

[0018] This cotton high-temperature tolerance prediction system, employing multi-source data fusion, achieves a precise description of the complex interaction between the environment, crop, and soil under high-temperature stress in cotton through deep fusion of heterogeneous multi-source data and dynamic knowledge graph construction. It innovatively couples a physiological development time-driven mechanistic model with an LSTM-XGBoost data-driven model, establishing a multi-scale prediction system from organ development to yield and quality formation. This system can quantify the dynamic cumulative impact of high temperature on cotton shedding rate, yield composition, and fiber quality, effectively solving the problems of inaccurate early warning and insufficient prediction dimensions caused by traditional methods due to limited data and missing mechanisms.

[0019] This cotton high-temperature tolerance prediction system, which utilizes multi-source data fusion, outputs a three-color high-temperature stress warning, multi-dimensional ranking of variety heat tolerance, and personalized field control prescriptions, forming a complete technical chain from monitoring and early warning to decision-making and execution. This solution overcomes the limitations of existing technologies that can only provide single early warnings, achieving accurate prediction and targeted control based on multi-source data fusion. It provides customizable solutions for different cotton-growing areas to cope with high-temperature disasters, improving risk management capabilities and planting management efficiency in cotton production. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the module interaction of a cotton high-temperature resistance prediction system using multi-source data fusion according to the present invention.

[0021] Figure 2 This is a flowchart of the training and optimization process for a cotton high-temperature resistance prediction system using multi-source data fusion, as described in this invention.

[0022] Figure 3 This is a logic diagram for determining the early warning level of a cotton high-temperature resistance prediction system using multi-source data fusion, as described in this invention.

[0023] Figure 4 This is a flowchart illustrating the decision generation process of a cotton high-temperature resistance prediction system that utilizes multi-source data fusion, as described in this invention. Detailed Implementation

[0024] 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.

[0025] Please see Figures 1 to 4This invention provides a technical solution: a cotton high-temperature tolerance prediction system using multi-source data fusion, comprising a data acquisition module, a data processing module, a model building module, and a decision output module; the data acquisition module is used to acquire multi-source heterogeneous data of the target area, which includes cotton-growing areas in Xinjiang, the Yellow River Basin, and the Yangtze River Basin; the multi-source heterogeneous data includes meteorological data, crop data, soil data, and environmental data, wherein the meteorological data includes basic meteorological parameters and high-temperature characteristic parameters, wherein the basic meteorological parameters include temperature, precipitation, light intensity, wind speed, and humidity; the high-temperature characteristic parameters include high-temperature intensity, duration, and frequency of extreme high temperatures, with temperature data collected in 1-hour increments, simultaneously recording the daily maximum temperature, minimum temperature, and hourly average temperature as the calculation benchmark; the crop data includes cotton variety genetic parameters, photosynthetic physiological dynamics, seed cotton yield composition, and fiber quality indicators; the soil data includes soil moisture dynamics, soil fertility, and soil microbial community; and the environmental data includes humidity and light intensity;

[0026] The data processing module is connected to the data acquisition module and is used to perform dynamic data cleaning and heterogeneous integration of multi-source heterogeneous data. It also constructs an interactive feature knowledge graph of environment-cotton-soil through multimodal feature fusion technology to realize the structured representation and dynamic updating of multi-factor coupling relationships.

[0027] The model building module is connected to the data processing module. It adopts a deep ensemble learning framework based on long short-term memory network and extreme gradient boosting tree, and couples a dynamic model of cotton boll growth and development driven by physiological development time and a model of yield and quality formation mechanism. It establishes a multi-scale model from physiological mechanism to phenotypic performance to quantify the dynamic cumulative impact of high temperature on cotton shedding rate, seed cotton yield, fiber quality and yield loss rate. Among them, fiber quality includes length, strength and micronaire value.

[0028] The decision output module connects to the model building module to output red, yellow, and green three-color early warning levels for high-temperature stress. It also performs multi-dimensional quantitative ranking of heat resistance of varieties based on yield, quality, and stress resistance weights. At the same time, it combines knowledge graphs and model outputs to generate personalized prescription suggestions for field stress resistance regulation, including quantitative schemes for irrigation amount and fertilizer formula. The system realizes high-temperature resistance prediction and decision support for cotton through multi-source data fusion and mechanism model driving.

[0029] It should be further explained that the system first collects multi-source heterogeneous data through sensor networks and remote monitoring equipment deployed in cotton-growing areas of Xinjiang, the Yellow River Basin, and the Yangtze River Basin. This includes high temperature intensity, duration, and frequency of extreme high temperatures from meteorological data; cotton variety genetic parameters, photosynthetic physiological dynamics, seed cotton yield composition, and fiber quality indicators from crop data; soil moisture dynamics, soil fertility, and soil microbial communities from soil data; and humidity and light from environmental data. After data collection, the data processing module performs dynamic data cleaning, uses an imputation algorithm based on data quality assessment to process missing values, identifies and corrects abnormal data through statistical anomaly detection methods, and uses time series alignment technology to achieve temporal consistency integration of multi-source data.

[0030] Subsequently, multimodal feature fusion technology was adopted to extract key features from various data sources through feature extraction methods, and an interactive feature knowledge graph of environment-cotton-soil was constructed based on graph structure. The nodes represent environmental factors, cotton physiological state and soil properties, respectively, and the edges represent the dynamic interaction relationship between them. The graph is dynamically updated through real-time data input to reflect the multi-factor coupling changes.

[0031] The model building module adopts a deep ensemble learning framework, using a long short-term memory network to capture the temporal dependencies of meteorological, crop, and soil data, and an extreme gradient boosting tree to enhance feature selection and nonlinear mapping. It also innovatively couples a dynamic model of cotton boll growth and development driven by physiological development time and a model of yield and quality formation mechanism. The physiological development time model simulates the cotton boll development process based on the time series of cotton growth stages, while the yield and quality model combines photosynthetic physiological dynamics to quantify the impact of high temperature on fiber quality indicators. Through multi-scale modeling, it achieves dynamic prediction from physiological mechanisms to phenotypic performance, and accurately outputs the quantitative results of the cumulative impact of high temperature on cotton shedding rate, seed cotton yield, fiber quality, and yield loss rate.

[0032] Based on model prediction results, the decision output module classifies high-temperature stress into three levels of warnings (red, yellow, and green) by setting thresholds. It also uses a weighted scoring method to quantitatively rank the heat tolerance of varieties in multiple dimensions, combining yield, quality, and stress resistance weights. Simultaneously, based on knowledge graphs and model outputs, it analyzes soil moisture, crop status, and environmental factors to generate personalized field regulation prescriptions, including specific values ​​for irrigation amount and fertilizer formula. Through the synergy of multi-source data fusion, mechanistic model embedding, and intelligent decision-making, this system has achieved a technological breakthrough in cotton high-temperature tolerance prediction, moving from single early warning to multi-dimensional regulation.

[0033] Dynamic data cleaning employs cleaning algorithms based on data quality and consistency, including imputation of missing data, identification and correction of outliers, and time series alignment of multi-source data. Multimodal feature fusion technology integrates key features from meteorological, crop, soil, and environmental data through feature extraction and selection, and constructs an interactive feature knowledge graph based on a graph structure, where nodes represent environmental factors, cotton physiological states, or soil properties, and edges represent the interaction relationships between factors.

[0034] It should be further explained that the specific implementation of its dynamic data cleaning and knowledge graph construction is as follows: The dynamic data cleaning algorithm first performs quality assessment on multi-source heterogeneous data, and uses a time series prediction-based imputation method for missing data. It uses the historical data change trend of the same monitoring point to establish a prediction model to fill in the missing values. For abnormal data identification, the sliding window statistical method is used to calculate the normal fluctuation range of each data parameter, and an adaptive threshold is set to mark data points that exceed the range. It also combines the data of nearby sensors for cross-validation and correction.

[0035] The specific implementation of dynamic data cleaning includes the following:

[0036] Missing data imputation algorithms include: for the missing data types of multi-source heterogeneous data, short-term missing data (continuous missing data ≤ 3 collection cycles) and long-term missing data (continuous missing data > 3 collection cycles), a hierarchical imputation strategy is adopted:

[0037] Short-term missing interpolation: Linear interpolation is used, and the calculation formula is as follows:

[0038] ;

[0039] Where, x t x represents the missing value at time t. t-1 x t+1 These are the valid data at adjacent times before and after t, where t is the collection timestamp, adapted to a collection frequency of 5 min / time for meteorological data and 30 min / time for soil data.

[0040] Long-term missing data imputation: LSTM time series prediction imputation was used to construct a univariate LSTM model with an input layer dimension of 12, hidden layer nodes of 64, and output layer dimension of 1. The activation function was ReLU, and the optimizer was Adam. The model used historical data from the 72 hours prior to the missing data as input to predict the missing data period. The training set consisted of no missing data from the same period in the target cotton-growing area over the past 5 years, and the validation set error was controlled within MAE < 5%. Historical data were sampled according to the collection frequency, such as 864 meteorological data samples and 144 soil data samples.

[0041] Outlier identification and correction includes using a sliding window + 3σ adaptive threshold method, with the following specific steps:

[0042] Set the sliding window size: meteorological data window = 24h, collected at 5min / time, totaling 288 data points; soil / crop data window = 72h, collected at 30min / time, totaling 144 data points; calculate the mean μ and standard deviation σ of the data within the window, with an adaptive threshold range of [μ-3σ, μ+3σ]; data exceeding the threshold are marked as outliers and corrected using the median replacement method of the valid data within the window; if the percentage of valid data within the window is <60%, cross-validation correction is performed using data from the same type of sensor in adjacent plots;

[0043] Threshold dynamic update frequency: meteorological data is updated once every 6 hours, and soil / crop data is updated once every 12 hours to ensure adaptation to dynamic changes in the cotton field environment.

[0044] In the multi-source data time-series alignment process, a unified time reference is established based on the timestamps of each data acquisition device, and non-uniformly sampled data is resampled to ensure that meteorological, soil, and crop physiological data are synchronized in time.

[0045] In the multimodal feature fusion stage, key features are first extracted from the cleaned data, including the high temperature accumulation index in meteorological data, water stress index in soil data, and photosynthetic efficiency parameters in crop data; then, the random forest algorithm is used to rank the importance of features and select the feature subset that is most sensitive to high temperature response.

[0046] The detailed steps of multimodal feature fusion include the following:

[0047] In feature extraction methods, differentiated extraction strategies are adopted for the feature attributes of different types of data, including the following:

[0048] Meteorological time-series data: An LSTM feature extraction network was used, with input being time-series data within 12 hours and hidden layer output dimension = 32, to extract time-dependent features; the meteorological time-series data includes high temperature intensity, duration, etc.; the time-series data within 12 hours was flattened into vectors according to the collection frequency; the time-dependent features include high temperature cumulative trend and extreme high temperature interval;

[0049] Soil static data: Principal component analysis (PCA) was used to reduce the dimensionality to 8 principal components, with a cumulative variance contribution rate of ≥92%, retaining core characteristics such as soil moisture stress index and organic matter content; among them, soil static data includes soil moisture and fertility;

[0050] Crop physiological data: A 1D convolutional layer of a convolutional neural network (CNN) was used to extract abrupt changes in physiological indicators, such as nodes where photosynthetic efficiency drops sharply due to high temperature. The crop physiological data included photosynthetic efficiency, fiber quality, etc. The kernel size of the convolutional layer was 3, the stride was 1, and the number of output channels was 16.

[0051] Feature selection and weight calculation include: using the random forest feature selection algorithm, with the following parameter settings:

[0052] Parameter name Value illustrate Number of decision trees 200 trees Balancing computational efficiency and feature discrimination Maximum tree depth 10th floor Avoid overfitting Feature subset sampling rate 0.8 Each split randomly selects 80% of the features Feature importance threshold 0.05 Retain features with importance ≥ 0.05.

[0053] The specific steps are as follows: First, input the feature set extracted from multiple sources, with a dimension of 64, and train a random forest model; then, calculate the importance score of each feature, and select 15 core features based on the reduction of the Gini coefficient, such as the high temperature accumulation index, soil relative humidity, photosynthetic rate, and cotton boll development stage; finally, assign feature weights: normalize according to the importance score, and the total weight = 1, such as the weight of the high temperature accumulation index = 0.22 and the weight of soil relative humidity = 0.18.

[0054] In the construction of the knowledge graph, the selected features are used as nodes. The node types include environmental factor nodes, cotton physiological state nodes, and soil attribute nodes. The relationship between nodes is determined by Pearson correlation analysis and mutual information calculation to form weighted edges. The graph is stored and updated using a graph database. When new monitoring data is input, the interaction relationship is dynamically adjusted by updating node attributes and recalculating edge weights, thereby accurately reflecting the real-time state changes of the environment-cotton-soil system.

[0055] The time-driven dynamic model of cotton boll growth and development simulates the growth and development process of cotton bolls under different high-temperature conditions based on time-series data of cotton physiological development stages. The yield and quality formation mechanism model combines the dynamics of cotton photosynthetic physiology and seed cotton yield composition to quantify the impact of high temperature on fiber quality indicators. In the deep ensemble learning framework, long short-term memory network is used to capture temporal dependencies, extreme gradient boosting tree is used to strengthen feature selection and nonlinear mapping, and multi-scale prediction is achieved through model fusion.

[0056] It should be further explained that the specific implementation of the mechanism and data-driven modeling is as follows: The dynamic model of cotton boll growth and development driven by physiological development time is based on the cumulative heat effect unit, and the cotton growth period is divided into several development stages. The nonlinear relationship between temperature and development rate is established to simulate the morphogenesis process of cotton boll under different high temperature conditions.

[0057] Physiological developmental time (PDT) is quantified using the daily heat day (GDD) method, formula:

[0058] ;

[0059] Wherein: T max，i T min，i : The highest / lowest temperature on day i; T b: Baseline temperature for cotton development, taken as 12℃, suitable for the main varieties planted in the target cotton-growing area; n: Number of days for statistics;

[0060] The PDT thresholds for each developmental stage are as follows: 120±10℃·d for the sowing-emergence stage, 280±20℃·d for the emergence-budding stage, 350±25 for the budding-flowering stage, and 420±30 for the flowering-cotyling stage.

[0061] The model takes daily temperature sequences and varietal genetic parameters as inputs and outputs the dynamics of boll dry matter accumulation and boll shedding probability. The yield and quality formation mechanism model is based on the principles of photosynthetic product allocation and transformation. By establishing a correlation function between photosynthetic physiological dynamics and fiber development, it quantifies the impact of high-temperature stress on fiber length, strength, and micronaire value formation.

[0062] In the deep ensemble learning framework, the Long Short-Term Memory Network receives time-series data after feature fusion. Its network structure includes forget gate, input gate, and output gate mechanisms to capture the long-term dependencies between meteorological factors, soil moisture, and crop physiological parameters. The Extreme Gradient Boosting Tree algorithm, on the other hand, processes non-time-series features and constructs decision tree combinations through multiple iterations to optimize the ranking of feature importance for yield composition and quality indicators.

[0063] The innovative coupling of the two models is reflected in the fact that the developmental stage index output by the physiological development time model is used as a feature input into the deep ensemble learning framework, while the temporal features extracted by the long short-term memory network and the nonlinear features extracted by the extreme gradient boosting tree are fused across modes to form a hybrid modeling architecture that combines mechanism-driven and data-driven approaches. This enables multi-scale accurate simulation of the cotton high-temperature response process from organ development to yield and quality formation.

[0064] It should be further explained that the innovative coupling of the two models includes the following:

[0065] The interaction interface between the physiological developmental time model and LSTM-XGBoost includes the following:

[0066] Feature interaction method: The physiological development time-driven model outputs three core features, which are used as supplementary features for the input layer of the LSTM, as follows:

[0067] The cotton boll development stage index ranges from 0 to 1, where 0 = flowering and boll-forming stage (initial flowering stage), 0.3 = flowering and boll-forming stage (full flowering stage), 0.7 = flowering and boll-forming stage (full boll-forming stage), and 1 = boll-opening stage. The definition standards for each stage are as follows: initial flowering stage is when 50% of the cotton plants in the field are flowering on the first fruiting branch and the first fruiting node; full flowering stage is when 50% of the cotton plants in the field are flowering on the fourth fruiting branch and the first fruiting node; full boll-forming stage is when more than 70% of the cotton plants in the field have formed cotton bolls with a diameter of ≥2cm; and boll-opening stage is from the beginning of boll opening to the basic end of the cotton harvest in the field.

[0068] Accumulated physiological development time;

[0069] The deviation from high temperature stress is the ratio of actual development time to normal development time.

[0070] The data transmission format is JSON, and the fields include timestamp, stage_index, accumulated_PDT, and stress_deviation. The transmission frequency is consistent with the crop data collection frequency, every 1 hour.

[0071] The calculation of model fusion weight coefficients includes the following:

[0072] The backpropagation optimization method using validation set error is employed, and the formula is as follows:

[0073] , ;

[0074] Where: w LSTM w XGBoost The fusion weights of LSTM and XGBoost are respectively; MSE LSTM MSE XGBoost The mean squared errors of the two models on the validation set are respectively; the validation set consists of data from the target cotton-growing area over the past three years, with a sample size of 5000.

[0075] Weight dynamic adjustment cycle: recalculated every 7 days based on the latest monitoring data to ensure prediction accuracy, and the MAE of the fused model is ≤3.5%.

[0076] The high-temperature stress three-color warning system in the decision output module is based on the high-temperature impact index output by the model. Red indicates high-risk stress, yellow indicates moderate-risk stress, and green indicates low-risk stress. The multi-dimensional quantitative ranking of heat resistance of varieties is based on a weighted scoring method, using yield loss rate, fiber quality change, and stress resistance indicators as ranking criteria.

[0077] It should be further explained that the specific implementation of its early warning and decision-making mechanism is as follows: The three-color early warning system for high-temperature stress (red, yellow, and green) is dynamically divided based on the high-temperature impact index output by the model. This index is derived by integrating the trend of cotton boll shedding rate, the prediction deviation of seed cotton yield, the fluctuation range of fiber quality indicators, and the estimated yield loss rate. The early warning level is divided using a multi-level threshold method. When the high-temperature impact index exceeds the preset upper threshold, a red warning is triggered, indicating that the risk of cotton boll shedding has increased significantly and fiber quality may be damaged. When the index is within the upper or lower threshold range, a yellow warning is triggered, indicating that there is potential stress and continuous monitoring is required. When the index is below the lower threshold, a green safe status is displayed.

[0078] The quantitative standards for the high temperature impact index are as follows:

[0079] Index calculation formula and factor weights, including: High Temperature Impact Index I HS Using a weighted summation model, the formula is:

[0080]

[0081] Among them, the shedding rate influencing factor R abscission : (Current high temperature shedding rate - Normal temperature shedding rate) / Normal temperature shedding rate, value range [0, 1]; Yield deviation factor D yield : (Normal output - Forecast output) / Normal output, value range [0, 1]; Quality fluctuation factor V quality : (fiber strength fluctuation value + micronaire value fluctuation value) / 2, with a value range of [0, 1] (fluctuation value = |measured value - standard value| / standard value); fertility period overlap factor O growth : Number of days overlapping between high temperature period and cotton boll development period / total number of days in cotton boll development period, with a value range of [0, 1].

[0082] Factor weight determination method: The analytic hierarchy process (AHP) was used, and five agricultural information technology experts were invited to score the factors. The consistency test CR < 0.1.

[0083] The red, yellow, and green warning thresholds meet the following table:

[0084] Warning Level Index range Corresponding risk description Red Alert <![CDATA[I HS >0.7]]> High risk: Cotton boll shedding rate >70%, yield loss >20% Yellow Alert <![CDATA[0.3≤I HS ≤0.7]]> Moderate risk: Shedding rate 60%-70%, yield loss 8%-20%. Green Alert <![CDATA[I HS <0.3]]> Low risk: Shedding rate <60%, yield loss <8%

[0085] The multi-dimensional quantitative ranking of heat resistance of varieties adopts a weighted scoring mechanism. The weight of the yield dimension is determined based on the deviation between the predicted seed cotton yield and the baseline yield. The weight of the quality dimension is calculated by comprehensively considering the fiber length retention rate, strength change rate, and micronaire value stability. The weight of the stress resistance dimension is adjusted according to the increase in shedding rate and physiological metabolic indicators under high temperature conditions. The final score of each variety is obtained by linear weighted summation and arranged in descending order of score to generate a heat resistance ranking list. This list is output in conjunction with the warning level to provide a quantitative basis for variety selection under different stress levels.

[0086] The generation of personalized prescription recommendations involves analyzing soil moisture dynamics, crop physiological status, and environmental factors based on knowledge graphs and model outputs, recommending specific values ​​for irrigation amounts and fertilizer formulas, and presenting them to users through a visual interface.

[0087] It should be further explained that the specific implementation of its personalized prescription generation is as follows: Based on the dynamic updated environment-cotton-soil interaction relationship in the knowledge graph and the multi-dimensional prediction results output by the model, the prescription generation engine first analyzes the current soil moisture dynamics and crop water stress status, calculates the difference between crop water requirement and soil effective water through the water balance equation, and dynamically determines the specific numerical range of irrigation amount by combining the high temperature duration and intensity data in future weather forecasts.

[0088] In personalized irrigation prescriptions, the water balance equation uses the Penman-Monteith formula to calculate the reference crop evapotranspiration ETo, and combines this with the crop coefficient Kc to obtain the actual water requirement ETC.

[0089] The calculation formula is as follows, based on crop evapotranspiration:

[0090] ;

[0091] Where: R n Net radiation, collected directly from the weather station; G: Soil heat flux, value for cotton fields = 0.1R. n T: Daily average temperature at 2m altitude; u2: Wind speed at 2m altitude; e s : Saturated vapor pressure, calculated from air temperature; e a : Actual water vapor pressure, calculated from humidity data; Δ: Slope of the water vapor pressure curve; γ: Wet and dry meter constant.

[0092] The actual water demand and irrigation volume are calculated using the following formulas: ETc = Kc × ETo, I = (ETc - PW) s )×1.2; Where: Kc: crop coefficient, budding stage=0.75, flowering and boll-forming stage=1.15, fluff-opening stage=0.85; P: daily precipitation, collected by meteorological station; W s 1.2: Effective soil water storage capacity, collected by a soil probe, taking the average value of the 0-60cm soil layer; 1.2: Safety factor, to cope with the fluctuation of evaporation caused by high temperature; I: Daily irrigation amount, in actual application, according to the irrigation cycle, such as 3 days, the total irrigation amount is calculated by accumulating.

[0093] In terms of fertilizer formula generation, the system establishes a formula calculation algorithm based on the nutrient balance model, based on soil fertility monitoring data, the nutrient requirements of cotton at different growth stages, and the impact of high temperature stress on nutrient absorption efficiency.

[0094] Fertilizer formula calculation, also known as nutrient balance model, taking nitrogen (N), phosphorus (P2O5), and potassium (K2O) as an example, is calculated using the following formula:

[0095] ;

[0096] Wherein: F xR: Amount of fertilizer applied for a certain nutrient; Y: Target seed cotton yield, based on model-predicted yield, with 90% taken as the target value, e.g., if the predicted yield is 500 kg / mu, the target value = 450 kg / mu; x Fertilizer requirements per unit yield, N=0.055, P2O5=0.022, K2O=0.048, based on the average values ​​of 3 years of field trials in the cotton-growing area of ​​the Yellow River Basin; S x A: Soil nutrient content, obtained from soil fertility testing reports, average value from the 0-30cm soil layer; B: Soil nutrient utilization coefficients: N=0.35, P2O5=0.25, K2O=0.40, suitable for cotton field loam soil type; FUE x Fertilizer utilization rate: 35% for conventional fertilization, adjusted to 28% under high temperature stress, because high temperature reduces nutrient absorption efficiency, converted to 0.28.

[0097] The algorithm comprehensively considers the soil's supply capacity for major elements such as nitrogen, phosphorus, and potassium, crop absorption characteristics, and changes in nutrient use efficiency caused by high temperatures, outputting a specialized fertilizer formula and application rate based on the current stress level. All prescription suggestions are optimized through a decision rule engine to ensure that irrigation and fertilization schemes maintain nutrient use efficiency while alleviating high-temperature stress. The final personalized prescription includes specific irrigation time, irrigation volume range, and fertilizer type and ratio scheme, and is presented to the user in a structured data form through a visual interface.

[0098] The construction of the interactive feature knowledge graph also includes a dynamic update mechanism, which adjusts the nodes and edges in the graph through real-time data input to reflect changes in the interaction relationship between the environment, cotton, and soil.

[0099] It should be further explained that the specific implementation of its knowledge graph dynamic update mechanism is as follows: The system receives multi-source data input from the monitoring network through the real-time data stream processing module. When a new batch of meteorological, soil, or crop physiological data arrives, the data quality is first verified and the format is standardized. Then, the incremental update process of the knowledge graph is started. This process includes three levels: node attribute update, edge weight adjustment, and topology optimization. In the node attribute update, the existing node attributes are weighted and corrected according to the new data. Among them, environmental factor nodes are given higher weights for recent data using a time decay model, and crop physiological nodes are updated based on the developmental stage. In the edge weight adjustment, the edge weight values ​​are recalculated by calculating the change in the correlation strength between nodes within the new data window and using correlation analysis within the sliding time window to dynamically reflect the changing trend of the environment-crop-soil interaction relationship.

[0100] A knowledge graph of the interaction characteristics between the environment, cotton, and soil includes the following:

[0101] Initial node weight assignment: The mutual information (MI) calculation method is used, as shown in the following formula:

[0102] ;

[0103] Where X and Y are two related nodes, such as "high temperature intensity" and "cotton boll shedding rate", P(x, y) is the joint probability, and P(x) and P(y) are the marginal probabilities; weight = MI(X, Y) / MI max MI max The maximum mutual information value for all node pairs, with an initial weight range of 0.1-0.9.

[0104] The dynamic update mechanism includes the following:

[0105] Update time windows: Seedling stage = 7 days, bud stage = 5 days, flowering and boll-forming stage = 3 days, fluff-opening stage = 7 days;

[0106] Update trigger conditions: When the deviation between the newly collected data and the current node attribute of the map is >10%, such as the measured soil moisture value being 15% lower than the map node value, the node weight recalculation and edge association strength update will be automatically triggered.

[0107] Storage and traceability: The Neo4j graph database is used to retain the most recent three update records and supports rollback queries.

[0108] In topology optimization, when new key feature relationships are detected or the strength of existing relationships remains below a threshold, graph structure reconstruction is automatically triggered. Emerging factor group relationships are identified through community discovery algorithms, and the connection topology in the graph is adjusted accordingly. The entire update process adopts a graph database version control mechanism to ensure data consistency and traceability of the knowledge graph during dynamic updates, thereby achieving near real-time response and representation of changes in cotton growth environment and status.

[0109] The training process of a deep ensemble learning framework includes training the model using historical multi-source data and optimizing the model parameters through cross-validation to ensure prediction accuracy and robustness.

[0110] Further explanation is needed regarding the specific implementation of its model training and optimization process as follows: The training of the deep ensemble learning framework first constructs a training set containing historical data from multiple growing seasons. This dataset covers meteorological, soil, crop physiology, and final yield and quality data from three major cotton-growing regions: Xinjiang cotton-growing area, Yellow River cotton-growing area, and Yangtze River cotton-growing area, and performs standardization processing. The training process adopts a phased strategy. First, the long short-term memory network is pre-trained, and the network parameters are optimized through the time series backpropagation algorithm. The input sequence is generated using the sliding window method, and training samples are constructed with the key growth period of cotton as the center. The training of the extreme gradient boosting tree adopts the gradient boosting iterative algorithm, and the yield prediction accuracy and quality index error are considered simultaneously through a custom loss function.

[0111] Model parameter optimization employs a k-fold cross-validation method, dividing the dataset into k mutually exclusive subsets. K-1 subsets are used alternately as the training set, and the remaining subset as the validation set. Optimal hyperparameter combinations are determined through grid search, including the number of hidden layer nodes, learning rate, and dropout rate of the Long Short-Term Memory network, as well as the maximum depth, learning rate, and number of subtrees of the extreme gradient boosting tree. To prevent overfitting, an early stopping mechanism is introduced during training, automatically terminating training when the validation set loss function no longer decreases for several consecutive training epochs.

[0112] In the final model integration stage, the prediction results of the Long Short-Term Memory Network and the Extreme Gradient Boosting Tree are fused using a weighted average method. The weight coefficients are dynamically adjusted based on the performance of each model on the validation set to ensure the stability and accuracy of the integrated model in dealing with different regions, different varieties, and different meteorological conditions.

[0113] The system also includes a user interface module, which receives field management data input by users and displays the warnings and prescription suggestions from the decision output module in a graphical manner.

[0114] It should be further explained that the specific implementation of its user interface module is as follows: The user interface module adopts a web-based visual architecture and provides two core components: a data input interface and a decision display interface. The data input interface includes structured forms and file upload functions, and receives field management data input by users, including irrigation records, fertilization time and amount, agricultural operation logs, and manually observed crop growth status. All input data is checked for format and logic through preset data validation rules to ensure consistency with the system's internal data standards.

[0115] The decision-making display interface uses a visualization component library to achieve collaborative presentation of multi-dimensional data. The high-temperature stress warning is displayed in the form of a color-coded map overlay, dynamically rendering red, yellow and green warning areas on cotton field distribution maps in different geographical regions. It also provides a timeline drag function to view the historical trend of warning level changes. The heat resistance ranking results of varieties are displayed in the form of an interactive leaderboard. Users can click on any variety to view details of its yield stability, fiber quality retention rate and stress resistance indicators under different high-temperature scenarios.

[0116] Personalized prescription recommendations employ a decision card-based design. Each card contains specific irrigation plans, fertilizer formulas, and recommended operation times, along with explanations of the regulatory basis based on knowledge graph reasoning. Users can adjust their management preferences using a slider and refresh the prescription content in real time. This interface module also provides multi-dimensional data comparison functionality, allowing users to select different time periods, different varieties, or different management measures for parallel comparison of prediction results. All visualization components support data export and report generation, forming a complete decision support closed loop.

[0117] The acquisition of multi-source heterogeneous data is achieved through sensor networks and remote monitoring equipment, including weather stations, soil probes, and crop physiological sensors. The data acquisition frequency is dynamically adjusted according to the cotton growth stage.

[0118] It should be further explained that the specific implementation of its multi-source data acquisition network is as follows: The sensor network consists of various monitoring devices deployed in typical plots of the target cotton area. Among them, the meteorological station is equipped with temperature, humidity, light intensity and wind speed sensors; the soil probe network is deployed in a grid to monitor soil temperature, volumetric water content and electrical conductivity; the crop physiological sensors include stem micro-change sensors, canopy spectrometers and leaf temperature monitors; the data acquisition frequency is dynamically adjusted according to the cotton growth stage, with a lower acquisition frequency set during the germination and seedling stages, meteorological data is collected every 30 minutes, soil data is collected every 2 hours, and crop physiological data is collected daily.

[0119] Once the plant enters the high-temperature sensitive stages such as the budding and flowering stages, the data collection frequency is automatically increased: meteorological data is collected every 5 minutes, soil data every 30 minutes, and crop physiological data every hour. All monitoring devices aggregate data through an IoT gateway, and data transmission from the field to the server is achieved using a hybrid networking method of LoRa wireless communication and 4G / 5G networks. During data transmission, key physiological parameters are encrypted and encapsulated to ensure the timeliness, integrity, and security of data collection. This dynamic data collection mechanism based on growth stages not only ensures the data density during the critical period of high-temperature stress but also achieves efficient utilization of network resources.

[0120] The multi-scale model outputs in the model building module include the cumulative impact curve of high temperature on cotton shedding rate, seed cotton yield forecast, fiber quality index change trend, and yield loss rate estimate. These outputs are used to support the decision-making module's early warning and prescription generation.

[0121] It should be further explained that the specific implementation of its multi-scale model output is as follows: The model building module outputs four types of core prediction results through a multi-scale modeling architecture. Among them, the cumulative impact curve of high temperature on cotton shedding rate is based on the collaborative calculation of physiological development time model and long short-term memory network. By integrating the intensity and duration of high temperature stress at different growth stages, a shedding probability curve that changes over time is generated. The seed cotton yield prediction value uses the extreme gradient boosting tree algorithm to process non-time-series features. Combined with the accumulation and distribution law of photosynthetic products in the mechanism model, the yield per unit area estimate based on the current environmental conditions and crop status is output.

[0122] The trend of fiber quality indicators is analyzed by coupling the yield and quality formation mechanism model with a deep ensemble learning framework. Response functions for fiber length, strength, and micronaire value with high-temperature stress parameters are established, generating dynamic change trajectories for each quality parameter during the remaining growth period. Yield loss rate estimation integrates the intensity and duration of high-temperature stress and the overlap of crop sensitive periods. A yield loss assessment model is constructed to quantify the potential yield loss ratio under different warning levels. These multi-scale outputs are fused and analyzed in the decision-making module. The cumulative impact curve is used to determine the key time nodes for warning level transitions. Yield forecasts and quality change trends serve as the core basis for variety ranking, while yield loss rate estimation directly participates in the economic evaluation of prescription recommendations, forming a complete prediction chain from physiological processes to economic impacts. This provides multi-dimensional data support for accurate decision-making under high-temperature stress.

[0123] This system achieves a precise description of the complex interaction between the environment, crop, and soil in cotton under high-temperature stress through deep fusion of multi-source heterogeneous data and construction of a dynamic knowledge graph. It innovatively couples a physiological development time-driven mechanistic model with an LSTM-XGBoost data-driven model to establish a multi-scale prediction system from organ development to yield and quality formation. This system can quantify the dynamic cumulative impact of high temperature on cotton shedding rate, yield composition, and fiber quality, effectively solving the problems of inaccurate early warning and insufficient prediction dimensions caused by the lack of data and mechanisms in traditional methods.

[0124] The system outputs a three-color early warning system for high-temperature stress, a multi-dimensional ranking of varietal heat tolerance, and personalized field control prescriptions, forming a complete technical chain from monitoring and early warning to decision-making and execution. This solution overcomes the limitations of existing technologies that can only provide a single early warning, achieving accurate prediction and targeted control based on multi-source data fusion. It provides customizable solutions for different cotton-growing areas to cope with high-temperature disasters, improving the risk management capabilities and planting management efficiency of cotton production.

[0125] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0126] 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 cotton high-temperature resistance prediction system using multi-source data fusion, characterized in that, It includes a data acquisition module, a data processing module, a model building module, and a decision output module. The data acquisition module is used to acquire multi-source heterogeneous data of the target area, which includes cotton-growing areas in Xinjiang, the Yellow River Basin, and the Yangtze River Basin. The multi-source heterogeneous data includes meteorological data, crop data, soil data, and environmental data. The meteorological data includes basic meteorological parameters and high-temperature characteristic parameters. The basic meteorological parameters include temperature, precipitation, light intensity, wind speed, and humidity. The high-temperature characteristic parameters include high-temperature intensity, duration, and frequency of extreme high temperatures. The crop data includes cotton variety genetic parameters, photosynthetic physiological dynamics, seed cotton yield composition, and fiber quality indicators. The soil data includes soil moisture dynamics, soil fertility, and soil microbial community. The environmental data includes humidity and light intensity. The data processing module is connected to the data acquisition module and is used to perform dynamic data cleaning and heterogeneous integration of multi-source heterogeneous data. It also constructs an environment-cotton-soil interaction feature knowledge graph through multimodal feature fusion technology for the structured representation and updating of multi-factor coupling relationships. The model building module is connected to the data processing module. It adopts a deep ensemble learning framework based on long short-term memory network and extreme gradient boosting tree, and couples a dynamic model of cotton boll growth and development driven by physiological development time and a model of yield and quality formation mechanism to establish a multi-scale model from physiological mechanism to phenotypic performance. This model is used to quantify the dynamic cumulative impact of high temperature on cotton shedding rate, seed cotton yield, fiber quality and yield loss rate. Fiber quality includes length, strength and micronaire value. The decision output module is connected to the model building module and is used to output a three-color warning system for high temperature stress (red, yellow, and green). It also performs multi-dimensional quantitative ranking of heat resistance of varieties based on yield, quality, and stress resistance weights. At the same time, it combines knowledge graphs and model outputs to generate personalized prescription suggestions for field stress resistance regulation, including quantitative schemes for irrigation amount and fertilizer formula.

2. The cotton high-temperature resistance prediction system using multi-source data fusion according to claim 1, characterized in that: The dynamic data cleaning adopts a cleaning algorithm based on data quality and consistency, including imputation of missing data, identification and correction of outliers, and time series alignment of multi-source data; the multimodal feature fusion technology integrates key features from meteorological, crop, soil and environmental data through feature extraction and feature selection, and constructs an interactive feature knowledge graph based on graph structure, where nodes represent environmental factors, cotton physiological state or soil properties, and edges represent the interaction relationships between factors.

3. The cotton high-temperature resistance prediction system using multi-source data fusion according to claim 1, characterized in that: The physiological development time-driven dynamic model of cotton boll growth and development simulates the growth and development process of cotton bolls under different high temperature conditions based on time series data of cotton physiological development stages. The yield and quality formation mechanism model combines the dynamics of cotton photosynthetic physiology and seed cotton yield composition to quantify the impact of high temperature on yield and fiber quality indicators. In the deep ensemble learning framework, long short-term memory network is used to capture temporal dependencies, extreme gradient boosting tree is used to strengthen feature selection and nonlinear mapping, and multi-scale prediction is achieved through model fusion.

4. The cotton high-temperature resistance prediction system using multi-source data fusion according to claim 1, characterized in that: The high-temperature stress three-color warning system in the decision output module is based on the high-temperature impact index output by the model, where red indicates high-risk stress, yellow indicates moderate-risk stress, and green indicates low-risk stress. The multi-dimensional quantitative ranking of the heat resistance of varieties is based on a weighted scoring method, using yield loss rate, fiber quality changes, and stress resistance indicators as ranking criteria.

5. The cotton high-temperature resistance prediction system using multi-source data fusion according to claim 1, characterized in that: The generation of personalized prescription recommendations involves analyzing soil moisture dynamics, crop physiological state, and environmental factors based on knowledge graphs and model outputs, recommending specific values ​​for irrigation amount and fertilizer formula, and presenting them to users through a visual interface.

6. A cotton high-temperature resistance prediction system using multi-source data fusion according to claim 2, characterized in that: The construction of the interactive feature knowledge graph also includes a dynamic update mechanism, which adjusts the nodes and edges in the graph through real-time data input to reflect changes in the interaction relationship between the environment, cotton, and soil.

7. A cotton high-temperature resistance prediction system using multi-source data fusion according to claim 3, characterized in that: The training process of the deep ensemble learning framework includes training the model using historical multi-source data and optimizing the model parameters through cross-validation to ensure prediction accuracy and robustness.

8. The cotton high-temperature resistance prediction system using multi-source data fusion according to claim 1, characterized in that: The system also includes a user interface module, which receives field management data input by the user and displays the warnings and prescription suggestions from the decision output module in a graphical manner.

9. A cotton high-temperature resistance prediction system using multi-source data fusion according to claim 1, characterized in that: The acquisition of the multi-source heterogeneous data is achieved through sensor networks and remote monitoring equipment, including weather stations, soil probes, and crop physiological sensors. The data acquisition frequency is adjusted according to the cotton growth stage.

10. A cotton high-temperature resistance prediction system using multi-source data fusion according to claim 1, characterized in that: The multi-scale model outputs in the model building module include the cumulative impact curve of high temperature on cotton shedding rate, seed cotton yield forecast, fiber quality index change trend, and yield loss rate estimate. These outputs are used to support the decision-making module's early warning and prescription generation.

Images