Visual analysis method and system based on agricultural big data model

Abstract

The invention relates to the technical field of agricultural information, and discloses a visual analysis method and system based on an agricultural big data model. According to the system, crop growth time sequence data are processed by using a multi-scale feature extraction algorithm combining a recurrent neural network and wavelet transform, and key conversion points of a crop growth stage are automatically identified by using a deep learning architecture combining a variational auto-encoder and a hidden Markov model. A standard growth track model and a multi-dimensional deviation measurement index are constructed to evaluate the abnormal growth condition of crops, a dynamic threshold model based on reinforcement learning is applied to realize early warning of abnormal growth, and a multi-dimensional visual interface based on a clock metaphor is designed to visually display crop growth information. The crop growth stage recognition precision and early warning timeliness are improved, and the agricultural production management efficiency is remarkably enhanced.

Description

Technical Field

[0001] The present invention relates to the field of agricultural information technology, and more specifically, to a visualization analysis method and system based on an agricultural big data model. Background Art

[0002] With the development of precision agriculture, the demand for refined monitoring and management of crop growth processes is growing. The existing crop growth monitoring systems have the following main technical limitations: First, they lack the ability to automatically identify the transition points of crop growth stages, making it difficult to accurately capture changes in key physiological stages such as vegetative growth to reproductive growth; second, they have limited multi-scale feature extraction and analysis capabilities for time-series growth data, and are unable to effectively capture short-term fluctuations and long-term trend characteristics at the same time; third, the quantitative measurement standards for growth anomaly deviations are imperfect, making it difficult to detect potential growth problems at an early stage; fourth, the early warning trigger mechanism is too simple, and most of them use fixed threshold judgments, which are difficult to adapt to the dynamic changes of different crop varieties and growth environments; fifth, there is a lack of visualization methods to intuitively present complex growth stages and early warning information, resulting in low efficiency in information transmission. These problems have greatly restricted the development of intelligent and precise agricultural production management.

[0003] In the prior art, there are crop growth status monitoring methods based on time series deep learning. Although deep learning methods are used to process crop growth data, their model structure mainly focuses on feature extraction on a single time scale and cannot effectively capture the dynamic characteristics of growth on multiple time scales. In addition, there is a lack of a special mechanism for automatic division of growth stages. There are crop disease and insect pest early warning methods, which are mainly based on fixed thresholds and expert rules for early warning judgments. They have limited flexibility and adaptability and cannot dynamically adjust the early warning threshold according to different environmental conditions. There are methods for visualizing the crop growth process, but their visual expressions are mainly based on traditional line charts and bar charts. The information level is single, and it is difficult to intuitively display multi-dimensional growth status and early warning information.

[0004] Therefore, there is an urgent need for a system that can automatically identify the growth stage of crops, provide early warning of growth abnormalities and provide an intuitive visual interface to improve the accuracy and efficiency of agricultural production management. Summary of the invention

[0005] The present invention provides a visualization analysis method and system based on an agricultural big data model, which solves the technical problems in related technologies such as difficulty in automatically identifying key transition points in crop growth stages, lack of multi-scale feature extraction capabilities and early warning of growth anomalies, and unintuitive information visualization expression.

[0006] The first aspect of the present invention provides a visualization analysis method based on an agricultural big data model, comprising the following steps:

[0007] A multi-scale feature extraction algorithm combining recursive neural network and wavelet transform is used to process crop growth time series data and generate representation vectors of multi-time scale features.

[0008] A deep learning architecture based on variational autoencoders and hidden Markov models performs unsupervised segmentation of the extracted multi-scale time series features to identify key transition points in the growth stage;

[0009] Construct a standard growth trajectory model and calculate the degree of deviation of current crop growth from the standard trajectory through multi-dimensional deviation metrics;

[0010] Apply a dynamic threshold model based on reinforcement learning to make early warning trigger judgments, achieve early warning of growth abnormalities, and reduce the false positive warning rate;

[0011] Build a multi-dimensional visualization interface based on the clock metaphor to intuitively display growth stages and warning information and improve the efficiency of information transmission.

[0012] Furthermore, the multi-scale feature extraction algorithm processes crop growth time series data including:

[0013] For the original time series dataset:

[0014] ;

[0015] Apply multi-level wavelet transform to decompose into different frequency components ,in, , indicating that at a point in time Collected dimensional growth parameter, where Indicates Level of detail factor, Indicates The level approximation coefficient captures the variation patterns on different time scales;

[0016] Construct a multi-branch bidirectional long short-term memory network to process wavelet coefficients of different scales;

[0017] The attention fusion module is used to calculate the importance weights of representations at different scales. , and generate a comprehensive multi-scale temporal feature representation vector .

[0018] Furthermore, the deep learning architecture based on variational autoencoder and hidden Markov model includes:

[0019] Multi-scale time series features Input the variational autoencoder and map it to the latent space through the encoder network to obtain the latent variables The mean and variance ;

[0020] The latent variables are transformed through the decoder network Reconstruct into original feature representation;

[0021] Constructing Hidden Markov Models for Latent Variables The time series is divided into states;

[0022] The Baum-Welch algorithm is used to optimize the HMM parameters, and the Viterbi algorithm is used to decode the most likely state sequence.

[0023] Furthermore, the construction of a standard growth trajectory model based on statistical learning and the calculation of the degree of deviation of the current crop growth from the standard trajectory through a multi-dimensional deviation metric index include:

[0024] Based on the historical growth data of high-quality crops, a functional data analysis method is used to fit the standard growth trajectory:

[0025] ;

[0026] in, For the The standard trajectory model of the growth stage, is the B-spline basis function, is the corresponding coefficient, is the residual term, is the number of basis functions;

[0027] For each growth indicator dimension, the confidence interval boundaries are calculated to form a dynamic envelope;

[0028] Calculate the multidimensional deviation metric between the currently monitored crop growth trajectory and the standard trajectory ;

[0029] Calculate the time cumulative deviation index Assess the severity of ongoing deviations.

[0030] Furthermore, the application of the dynamic threshold model based on reinforcement learning to perform early warning trigger judgment includes:

[0031] Construct a multi-level threshold model:

[0032] ;

[0033] in, Indicates that in the growth stage , Deviation type The next Differentiated warning threshold systems are set for different deviation types and growth stages;

[0034] Defining the state space is the combination of the current growth stage, deviation measure and historical warning effect, action space The direction and magnitude of the threshold adjustment;

[0035] Using double Q-learning algorithm to train threshold adjustment strategy ;

[0036] Based on the calculated deviation metric, cumulative deviation index and deviation type, dynamic thresholds are applied to make early warning trigger judgments.

[0037] Furthermore, the construction of a multi-dimensional visualization interface based on the clock metaphor includes:

[0038] Construct a clock-style growth stage visualization module to map the entire crop growth period onto a circular clock interface, with each growth stage represented by a different color and angle interval;

[0039] Design a multi-layer nested information expression structure, which includes the growth stage area, standard growth trajectory area, current growth indicator area, and early warning information area from the outside to the inside;

[0040] Develop an interactive warning information visualization component to map the warning information onto the clock interface in a visually encoded manner.

[0041] Furthermore, the multidimensional deviation metric is calculated by the following formula:

[0042] ;

[0043] in, is a deviation measure in the form of Mahalanobis distance, For the The weight of the dimension indicator, is the currently monitored crop growth trajectory, Indicates Growth stage The standard trajectory of the dimension indicator, is the corresponding standard deviation function.

[0044] Furthermore, the double Q learning algorithm trains the threshold adjustment strategy through the following formula:

[0045] ;

[0046] in, For the status Take action The Q value, is the learning rate, is the discount factor, For time The state of being For the status Select the action The maximum Q value of the output is the threshold adjustment strategy after optimization , For time The immediate reward at The calculation results are:

[0047] ;

[0048] in, Indicates in status Take action The reward value, are the weight coefficients of each indicator, corresponding to the accuracy, recall, timeliness and false alarm rate of the warning respectively.

[0049] Furthermore, the interactive warning information visualization component uses the following formula to calculate the warning flashing frequency:

[0050] ;

[0051] in, is the flashing frequency, is the basic frequency, and To adjust the parameters, and They are the urgency and level of the warning respectively.

[0052] The second aspect of the present invention provides a visualization analysis system based on an agricultural big data model, which is used to execute the above-mentioned visualization analysis method based on an agricultural big data model, including a data acquisition module, a multi-scale feature extraction module, a growth stage identification module, a deviation metric calculation module, a dynamic threshold warning module and a visualization interface module. Data flow and processing links are formed between the modules. The data acquisition module obtains crop growth time series data, the multi-scale feature extraction module and the growth stage identification module respectively realize feature extraction and stage division, the deviation metric calculation module and the dynamic threshold warning module collaborate to complete anomaly detection and warning triggering, and the visualization interface module integrates analysis results and provides an interactive operation interface.

[0053] The visualization analysis and system based on the agricultural big data model provided by the present invention has the following beneficial effects:

[0054] The algorithm that combines recursive neural networks with wavelet transforms can capture both short-term fluctuations and long-term trend characteristics of crop growth. The deep learning architecture based on variational autoencoders and hidden Markov models can improve the accuracy of growth stage recognition. Through multidimensional deviation metrics and a dynamic threshold model based on reinforcement learning, the system can issue warnings in the early stages of abnormal growth and identify potential problems in advance. The multidimensional visualization interface based on the clock metaphor enables agronomists to intuitively understand crop growth stages and warning information. It effectively reduces agricultural resource input and improves overall economic benefits. It has good generalization capabilities for unseen crop varieties and growth conditions, meeting the needs of a variety of agricultural production scenarios. It effectively solves the key technical problems faced by crop growth monitoring systems in existing technologies, provides a powerful decision-making support tool for precision agricultural management, and has significant technical innovation and practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flow chart of the main steps of the visualization analysis system based on the agricultural big data model of the present invention;

[0056] Figure 2 is a flow chart of the multi-scale temporal feature extraction steps of the present invention;

[0057] Figure 3 is a flow chart of the automatic identification steps of the growth stage of the present invention;

[0058] Figure 4 is a flow chart of the growth trajectory deviation metric calculation steps of the present invention;

[0059] Figure 5 is a flow chart of the steps of the dynamic threshold warning algorithm of the present invention;

[0060] Figure 6 It is a flow chart of the steps of constructing a multi-dimensional clock-type visual interface of the present invention. DETAILED DESCRIPTION

[0061] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is only to enable those skilled in the art to better understand and implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the present specification. Various examples may omit, replace, or add various processes or components as needed. In addition, the features described in some examples may also be combined in other examples.

[0062] The visualization analysis method based on the agricultural big data model provided in this embodiment is as follows: Figure 1 As shown, it mainly includes the following steps:

[0063] Step 1: Multi-scale temporal feature extraction

[0064] like Figure 2 As shown in the figure, a multi-scale feature extraction algorithm combining recursive neural network and wavelet transform is used to process crop growth time series data to generate a representation vector containing multiple time scale features. Specifically, it includes:

[0065] Step 1.1: Input the crop growth time series dataset:

[0066] ;

[0067] in, , indicating that at a point in time Collected Dimensional growth parameters, including quantitative indicators such as plant height, stem diameter, leaf area, and chlorophyll content; in the rice monitoring application of this embodiment, , including key physiological indicators such as plant height (cm), number of tillers (pieces), stem diameter (mm), number of leaves (pieces), leaf area index, chlorophyll content (SPAD value), canopy temperature (°C) and biomass (g). The sampling frequency is once a day, and the continuous sampling period is sky;

[0068] Step 1.2: Apply multi-level wavelet transform to decompose the original time series data, decompose the signal into different frequency components, and obtain multi-scale representation ,in Indicates Level of detail factor, Indicates The approximate coefficient of the level is used to capture the change pattern at different time scales; in rice growth monitoring, Daubechies wavelet (db4) is selected as the basis function to decompose the level , corresponding to short-term (1-3 days), near-term (4-7 days), medium-term (1-2 weeks) and long-term (>2 weeks) temporal variation patterns, which can simultaneously capture short-term fluctuations caused by day-night temperature differences and long-term trends in growth stage transitions;

[0069] Step 1.3: Construct a multi-branch bidirectional long short-term memory network to process wavelet coefficients of different scales. The network contains There are parallel branches, each of which processes a scale of wavelet coefficients and extracts temporal dynamic features from them. , extract the hidden state representation through bidirectional LSTM:

[0070] ;

[0071] in, Indicates The wavelet coefficients at the time point The hidden state representation of For the The wavelet decomposition coefficients of level , For time point Each BiLSTM branch is configured as a 2-layer structure with 32 hidden units in each layer. A Dropout rate of 0.2 is used to prevent overfitting, which can effectively handle time series dependencies of different lengths.

[0072] Step 1.4: Use the attention fusion module to integrate multi-scale features and calculate the importance weights of different scale representations , and generate a comprehensive multi-scale temporal feature representation :

[0073] ;

[0074] ;

[0075] in, Indicates The attention weight of the wavelet coefficients, is the learnable weight vector, express The transpose of For the The hidden state of the wavelet coefficients at the end of the sequence, is the number of levels of wavelet decomposition, for In the rice growth monitoring example, the feature vector dimension is The feature representation vector encodes the growth dynamic information at different time scales. In practical applications, experiments have shown that the fused feature representation can improve the F1 score of crop growth anomaly detection by 18.7% compared with the single time scale feature, especially for complex growth anomaly patterns such as tillering inhibition and panicle differentiation delay.

[0076] Step 2: Automatic identification of growth stages

[0077] like Figure 3 As shown in the figure, based on the deep learning architecture combining variational autoencoder and hidden Markov model, the extracted multi-scale time series features are segmented unsupervised to identify the key transition points of crop growth stages. Specifically, it includes:

[0078] Step 2.1: Input the multi-scale time series features into the variational autoencoder, map them to the latent space through the encoder network, and obtain the latent variables The mean and variance :

[0079] ;

[0080] ;

[0081] in, Represents the multi-scale temporal features of the input, is the encoder network, which contains multiple layers of fully connected neural networks. is a latent variable The mean of is a latent variable The variance of for dimensional latent variables, Indicates the mean , the variance is Normal distribution; in this embodiment, the encoder network adopts a three-layer fully connected structure, the number of hidden layer neurons is 128, 64 and 32 respectively, the activation function is LeakyReLU, and the latent variable dimension is ,Comparative experiments show that this dimension achieves the best balance between maintaining information integrity and reducing complexity, and is most sensitive to the characteristic expression of rice at different growth stages;

[0082] Step 2.2: Reconstruct the latent variables into the original feature representation through the decoder network and calculate the reconstruction error as part of the model training objective:

[0083] ;

[0084] ;

[0085] in, The decoder network is responsible for transforming the latent variables Decoded into reconstructed feature representation ; is the original multi-scale temporal feature representation, is the reconstructed feature representation, is the reconstruction loss, representing the original features and reconstruction features The decoder uses a three-layer fully connected structure symmetrical to the encoder, with 32, 64, and 128 neurons respectively. At the same time, the KL divergence loss is introduced:

[0086] ;

[0087] in, represents the Kullback-Leibler divergence, The mean is , the variance is The normal distribution of is a standard normal distribution. The total loss function is ,in It is a balancing factor used to adjust the weight between reconstruction loss and KL divergence loss. Through testing, it is determined that this value can achieve the best effect in feature extraction of rice growth stages.

[0088] Step 2.3: Construct a hidden Markov model to divide the time series of latent variables into states and define the state transition probability matrix , the emission probability distribution , and the initial state distribution Assuming that there are Potential state , each state corresponds to a growth stage:

[0089] ;

[0090] ;

[0091] ;

[0092] in, Indicates the initial state for The probability of For potential state, Indicates time point status, represents the probability, From potential state Transfer to latent state The probability of For time point The potential variables, Indicates potential state Observed below The probability of multivariate Gaussian distribution is used for modeling; in the rice growth monitoring application, the state number is determined by agronomic expert knowledge , which correspond to the six key physiological stages of seedling stage, tillering stage, panicle differentiation stage, heading stage, flowering stage and filling and maturity stage. The self-transfer probability is set in the model initialization. Higher, reflecting the persistence characteristics of the growth stage;

[0093] Step 2.4: Apply the Baum-Welch algorithm to optimize the HMM parameters and use the Viterbi algorithm to decode the most likely state sequence:

[0094] ;

[0095] The calculation method is:

[0096] ;

[0097] in, represents the most likely state sequence, For time point status, Represents a given latent variable sequence Time state sequence probability.

[0098] Transition points in a state sequence The key transition time points identified as growth phases are output as growth phase division results:

[0099] ;

[0100] in, Indicates identified growth stages, and In practical applications, the training process of the Baum-Welch algorithm is controlled by a maximum number of iterations of 100 and a convergence threshold of 0.001. For tests of japonica rice varieties in different seasons, the model can automatically identify the key physiological stage transition points with an accuracy of 92.5%, and detect the transition signal 1.67 days earlier than manual experience judgment on average, providing more sufficient preparation time for agricultural management.

[0101] Step 3: Growth trajectory deviation metric calculation

[0102] like Figure 4 As shown in the figure, a standard growth trajectory model based on statistical learning is constructed, and the degree of deviation between the current crop growth and the standard trajectory is calculated through multi-dimensional deviation measurement indicators. Specifically, it includes:

[0103] Step 3.1: Construct a standard growth trajectory model based on historical high-quality crop growth data. , collect the multi-dimensional time series features of this stage in the historical data ,in is the number of historical samples, and the functional data analysis method is used to fit the standard growth trajectory:

[0104] ;

[0105] in, For the The standard trajectory model of the growth stage, is the B-spline basis function, is the corresponding coefficient, is the residual term, is the number of basis functions; in the rice growth monitoring example, based on three years of historical data ( About 3000 plants), B-spline adopts 3rd order, and different numbers of nodes are set at different growth stages: 15 at tillering stage, 10 at panicle differentiation stage, 8 at heading stage, and 12 at filling and maturity stage. The node distribution density is adjusted according to the growth change rate at each stage, and the node density is higher in the faster changing area;

[0106] Step 3.2: Calculate the confidence interval boundaries for each growth indicator dimension to form a dynamic envelope:

[0107] ;

[0108] ;

[0109] in, Indicates Growth stage The standard trajectory of the dimension indicator, is the corresponding standard deviation function, is the significance level The Z score under and are the upper and lower envelopes respectively; in practical applications, an adaptive significance level is used , the value range is 0.01-0.05 at different growth stages and different indicators. Water stress sensitive indicators (such as leaf area and chlorophyll content) use smaller values ​​at sensitive stages (such as heading stage). value, improve the early warning sensitivity;

[0110] Step 3.3: Calculate the multi-dimensional deviation metric of the currently monitored crop growth trajectory from the standard trajectory:

[0111] ;

[0112] in, is a deviation measure in the form of Mahalanobis distance, For the The weight of the dimension indicator, is the currently monitored crop growth trajectory, Indicates Growth stage The standard trajectory of the dimension indicator, is the corresponding standard deviation function; in this embodiment, the weights of various indicators at different growth stages are determined by random forest feature importance analysis. For example, the weight of tiller number in the tillering stage is 0.35, and the weight of leaf area index is 0.25, which is relatively high, while the weight of grain filling rate in the filling stage is 0.40, and the weight of stem dry matter accumulation is 0.30, which is relatively high. The weights of various indicators at each stage are determined after big data analysis and pre-configured in the system;

[0113] Step 3.4: Calculate the time-accumulated deviation index to assess the severity of the continuous deviation:

[0114] ;

[0115] in, For time The cumulative deviation index at time The cumulative deviation degree at is the length of the integration time window, which indicates the length of the time window when calculating the cumulative deviation index; is the time decay factor, giving higher weight to recent deviations, Indicates that with the time difference As the value of the data increases, the impact of the deviation gradually decreases, emphasizing the importance of recent data; For time The deviation measure in the form of Mahalanobis distance at time The degree of deviation; is the current time, is the integral variable. In the rice monitoring application, different time windows are set according to different growth stages. , a shorter window (5-7 days) is set in the early growth stage, and a longer window (10-14 days) is set in the later stage, and the time decay factor , realize integral calculation through discretization;

[0116] Step 3.5: Construct a multidimensional deviation type identifier to classify the deviation patterns into different types of growth anomalies:

[0117] ;

[0118] in, For time The type of anomaly identified at For time The deviation measure in the form of Mahalanobis distance at indicates the degree of deviation between the current crop growth trajectory and the standard trajectory; For time The cumulative deviation index at time The cumulative deviation degree at is the rate of change of the deviation measure, indicating the deviation measure in time The classifier can use a decision tree or random forest model to output a set of deviation types ,in Indicates the time point when the deviation occurs. Indicates the type of deviation, Indicates the severity of the deviation. In practical applications, a random forest classifier (number of trees: 100, maximum depth: 8) is used, which is trained based on historically annotated anomaly cases. It can classify anomalies into 9 categories and 36 subcategories, including water stress (mild / moderate / severe), nutrient deficiency (nitrogen / phosphorus / potassium / trace elements), pest and disease risks (rice blast / sheath blight / rice planthoppers, etc.), and growth retardation (tillering inhibition / delayed panicle differentiation / poor filling). The accuracy rate reaches 87.5%, and the F1 score is 0.83, which is 26.5% higher than the traditional single indicator fixed threshold judgment method. It can generate targeted agricultural management suggestions based on different anomaly characteristics.

[0119] Step 4: Dynamic Threshold Warning Algorithm

[0120] like Figure 5 As shown in the figure, a dynamic threshold model based on reinforcement learning is applied to achieve early warning of abnormal crop growth and reduce the false positive warning rate. Specifically, it includes:

[0121] Step 4.1: Build a multi-level threshold model and set differentiated warning threshold systems for different deviation types and growth stages:

[0122] ;

[0123] in, Indicates that in the growth stage , Deviation type The next The warning threshold adopts a multi-level progressive structure. Corresponding to mild, moderate and severe warning levels respectively; in the rice monitoring system, the initial threshold is set according to the statistical analysis of historical data. For example, the three-level thresholds of water stress type in the panicle differentiation period are (mild), (moderate) and (Serious), these initial thresholds are dynamically adjusted and optimized through the subsequent reinforcement learning process;

[0124] Step 4.2: Build a reinforcement learning model for dynamic threshold adjustment based on historical warning data and expert feedback. Define the state space is the combination of the current growth stage, deviation measure and historical warning effect, action space is the direction and magnitude of the threshold adjustment, the reward function Based on the accuracy and timeliness of early warning:

[0125] ;

[0126] in, Indicates in status Take action The reward value, is the weight coefficient of each indicator, corresponding to the accuracy (Precision), recall rate (Recall), timeliness (Timeliness) and false alarm rate (FalseAlarm) of early warning respectively; for the application of japonica rice varieties in Jiangsu Province, set , giving priority to early warning timeliness and accuracy. State space Contains 36-dimensional feature vectors (including 6 growth stages, 9 types of anomalies, current deviation metrics, historical threshold effects, etc.), action space It is a 15-dimensional discrete space, corresponding to 5 adjustment operations for each of the 3-level thresholds (large increase, small increase, unchanged, small decrease, large decrease);

[0127] Step 4.3: Use double Q-learning algorithm to train threshold adjustment strategy , optimize the Q-value function through the experience replay mechanism:

[0128] ;

[0129] in, For the status Take action The Q value, is the learning rate, For time The immediate reward at Calculated, is the discount factor, For time The state of being For the status Select the action The maximum Q value of the output is the threshold adjustment strategy after optimization In the actual system, the model uses a three-layer neural network (with 128, 64, and 32 nodes respectively) to parameterize the Q function, set the experience playback buffer size to 10,000, the batch size to 64, and the learning rate to , discount factor , training adopts - Greedy strategy, exploration rate The linear decay from 0.9 to 0.1, and the target network update cycle used in the training process is 200 steps. These hyperparameter configurations make the model converge fastest and have the most stable performance in the rice growth abnormality warning scenario;

[0130] Step 4.4: Apply dynamic thresholds to the currently monitored crop growth trajectory for early warning trigger judgment:

[0131] ;

[0132] in, is the deviation measure, is the cumulative deviation index, is the deviation type, is the warning function, output value Indicates the triggered warning level. Indicates that no warning is triggered. is the corresponding cumulative deviation index threshold; the actual system adds an anti-shake mechanism, requiring that the trigger conditions be met for three consecutive days before an alert is generated. At the same time, emergency warning logic is set for extremely severe conditions. The highest level of warning is triggered immediately to ensure timely response to sudden and extreme abnormal situations;

[0133] Step 4.5: Apply the warning priority sorting algorithm to merge and sort multiple warnings triggered at the same time to generate the final warning information set:

[0134] ;

[0135] in, For warning time, Identifier of the individual crop to which it is associated (if applicable), is the spatial location information, is the exception type, For the warning level, The urgency level is calculated based on the length of the time window from the occurrence of an anomaly to the realization of potential harm. In the actual implementation of the system, the warning priority sorting algorithm adopts a weighted scoring mechanism, and the scoring function is:

[0136] ;

[0137] in, is the estimated percentage of affected area, is the diffusion risk coefficient, and the weight parameter is:

[0138] ;

[0139] The warnings generated by the system every day are arranged in descending order according to this score, and clustered and merged (multiple warnings of the same type with geographical proximity are merged into a single warning but the affected area information is retained), and finally transmitted to the mobile terminal of the agronomic manager. The system automatically generates a list of recommended intervention measures for each warning to help decision makers respond quickly. In the application of the Suqian Rice Base, this algorithm has shortened the agricultural intervention response time after the warning from an average of 1.8 days to 0.8 days, improving the practicality of the system.

[0140] Step 5: Construction of multi-dimensional clock-style visualization interface

[0141] like Figure 6 As shown in the figure, a multi-dimensional visualization interface based on the clock metaphor is constructed to intuitively display crop growth stages and early warning information, thereby improving the efficiency of information transmission. Specifically, it includes:

[0142] Step 5.1: Build a clock-style growth stage visualization module to map the entire crop growth period to a circular clock interface, with each growth stage represented by a different color and angle interval:

[0143] ;

[0144] in, Indicates The starting angle of each growth stage, and Indicates The start and end time of each growth phase, is the total number of growth stages. The current growth status is intuitively indicated by the position of the clock pointer and the corresponding logo; in the implementation of the rice monitoring system, HTML5's Canvas and SVG technologies are used to build an interactive clock interface. The six growth stages use green gradient colors (hue range 120°-150°, saturation 80%-95%, brightness from light to dark), and the angle span is proportional to the actual duration of each stage. For example, the tillering period (25 days) occupies an arc of 60°, and the panicle differentiation period (20 days) occupies an arc of 48°; the interface displays the current development progress through animation effects, and starts flashing 3 days before the key transition point to indicate that the transition is about to begin;

[0145] Step 5.2: Design a multi-layer nested information expression structure, which includes the growth stage area, standard growth trajectory area, current growth indicator area, warning information area, etc. from the outside to the inside, forming a clearly hierarchical visual layout:

[0146] ;

[0147] in, Indicates The radial position of the layer information, is the base radius, The system uses responsive design principles to automatically adjust the resolution of different screen sizes (from 1366×768 to 4K). and Parameters to ensure the best visual effect; radial gradient background is used to distinguish different information layers, and 15% transparency transition zone is used between adjacent layers to enhance the sense of hierarchy; wavy line graph is used in the standard growth trajectory area to indicate the normal trend range, and dynamic bar graph is used in the current growth index area to show the comparison between the real-time value of the main growth parameters and the normal range;

[0148] Step 5.3: Develop interactive warning information visualization components to visualize warning information Mapped to the clock interface in a visual coding manner, the warning type, level and urgency are expressed using visual attributes such as color, shape, size and flashing frequency. The urgency is calculated using the following formula to calculate the warning flashing frequency:

[0149] ;

[0150] in, is the flashing frequency, is the basic frequency, and To adjust the parameters, and are the urgency and level of the warning, is the base of the natural logarithm, approximately equal to 2.71828; in the actual system, set , , , ensuring that the flashing frequencies between different warning levels are clearly distinguishable but will not cause visual fatigue; the warning information adopts the following coding rules: water stress warnings use blue icons, nutrient deficiency warnings use green icons, pest and disease risk warnings use red icons, and growth retardation warnings use yellow icons. The warning level is encoded by the icon size (diameter 16px, 20px, 24px) and border thickness (1px, 2px, 3px), and the urgency is expressed by the flashing frequency and icon opacity (0.7, 0.85, 1.0);

[0151] Step 5.4: Construct a multi-scale warning information display interface, support multi-level warning information drill-down from the plot level to the individual level, and summarize and display the warning information at different scales by clustering:

[0152] ;

[0153] in, Indicated in scale The clustering results of early warning information under Representing the plot level, regional level and individual level scales respectively; the system implementation uses a three-level warning map to achieve multi-scale display: the satellite base map is superimposed with a heat map to display the plot-level warning distribution, and it supports zooming in to the regional level to view the warning density and type distribution, and then zooming in to the single plant level to view which specific plants are affected and to what extent; the clustering algorithm uses a variant of the DBSCAN algorithm based on density and geographic location, and automatically adjusts the clustering parameters at different zoom scales ( and ), ensure that the visual expression is clear and not overloaded; in conjunction with the clock interface, clicking the warning area on the map can highlight the corresponding stage and warning on the clock interface;

[0154] Step 5.5: Implement multi-dimensional data linkage analysis functions, support cross-view data selection, screening and correlation analysis, and build a complete visual analysis link. Design an adaptive layout algorithm to keep the interface readable and interactive on different devices:

[0155] ;

[0156] in, represents the optimized layout function, Represents the set of data items to be displayed. represents the display space constraint, For data items The importance weight of It is a function for evaluating layout quality. The system is developed based on the React framework and uses adaptive grid layout technology to establish data binding relationships between views, realize the linkage of clock view, map view, graph view and list view; add a detailed expansion panel, and when clicking on a warning, automatically expand the information panel containing detailed descriptions, historical warning records of similar warnings, recommended intervention measures and effect estimates; optimize touch interaction for mobile devices, support gesture zooming, sliding to view historical data and other operations; in the application of the system in the Suqian rice base, the average time for agronomists to view warning information through mobile terminals was reduced from 2.8 minutes to 45 seconds, the information cognition efficiency was improved by 73.2%, and the user satisfaction score reached 4.7 points (out of 5 points).

[0157] The visualization analysis method based on the agricultural big data model provided in this embodiment has the following technical effects:

[0158] 1) It realizes automatic identification of crop growth stages and precise positioning of key physiological transition points, with an identification accuracy rate of 92.5%, an improvement of 18.3% compared with the manual experience judgment of the existing technology, and the positioning time accuracy is improved to ±1.2 days;

[0159] 2) Through the multi-scale feature extraction method combining recursive neural network and wavelet transform, it is possible to simultaneously capture the dynamic characteristics of growth at different time scales, improve the feature expression ability by 28.7%, and increase the sensitivity to short-term fluctuations and long-term trends by 24.5% and 31.2% respectively;

[0160] 3) The unsupervised segmentation method of growth stage based on variational autoencoder and hidden Markov model has good generalization ability for unseen crop varieties and growth conditions. The model adaptability is 35.6% higher than the traditional method, and the stability under different environmental conditions is improved by 42.3%;

[0161] 4) The constructed multi-dimensional deviation metric and dynamic threshold warning algorithm can issue warnings in the early stages of growth abnormalities, identifying potential problems an average of 6.8 days in advance, with a warning accuracy rate of 87.3%, an increase of 26.5% over the fixed threshold method, and a reduction of the false positive rate by 31.7%;

[0162] 5) The multi-dimensional visualization interface built based on the clock metaphor enables agronomists to intuitively understand crop growth stages and early warning information, improving information cognition efficiency by 43.2%, shortening the average decision response time by 57.4%, and increasing user satisfaction scores by 4.2 points (out of 5).

[0163] This implementation effectively solves the technical problems in the prior art that crop growth monitoring systems are difficult to accurately identify the key physiological stages of plant growth and development and lack the ability to provide early warning of growth abnormalities. It provides a more accurate and timely decision-making support tool for agricultural production management and has significant technical innovation and practical value.

[0164] The specific application of this implementation method in a rice planting base is as follows:

[0165] The test was conducted in a rice planting base covering an area of ​​120 mu. Three different varieties of rice (japonica rice, indica rice, and hybrid rice) were selected for monitoring. A standard management area (control area) and a test area were set up for each variety. The following data collection equipment was deployed in the test area:

[0166] Multispectral imaging equipment: collect multispectral images of the field every 3 days, including visible light, near infrared and red edge bands;

[0167] IoT sensor network: A 32-node sensor network is deployed to collect environmental parameters such as soil temperature and humidity, air temperature and humidity, and light intensity every hour;

[0168] Fixed-position HD camera array: 8 HD cameras are installed at the boundaries of the experimental area to capture images of rice populations at regular intervals every day.

[0169] The test cycle covers the entire growth period of rice (from transplanting to harvesting), a total of 125 days.

[0170] Data preprocessing and model training phase:

[0171] The rice growth data of the past three years were collected, including standard characteristic data of different varieties at different growth stages, typical abnormal cases and growth stage demarcation points marked by experts. Based on these data, the growth stage recognition model and early warning threshold model were trained, and the performance of the model training was evaluated using the 5-fold cross-validation method.

[0172] System deployment phase:

[0173] Data processing and model calculation modules are deployed on the server side, a visualization interface is deployed on the cloud, and mobile terminal devices are provided for on-site agronomists. The system adopts a layered architecture, with the data collection layer responsible for raw data acquisition, the edge computing layer responsible for preliminary data processing, the cloud service layer responsible for model calculation and data storage, and the application layer responsible for visualization and user interaction.

[0174] Operation monitoring phase:

[0175] The system automatically processes the collected data, identifies the rice growth stage, calculates the deviation of the growth trajectory, triggers an early warning and displays it on a visual interface. Agronomists receive early warning information in real time through mobile terminals and take corresponding farming operations according to the system's recommendations.

[0176] Application effect analysis

[0177] Growth stage identification effect:

[0178] The system automatically identifies the six main growth stages of rice: seedling stage, tillering stage, panicle differentiation stage, heading stage, flowering stage and grain filling and maturity stage. By comparing with the traditional manual judgment method, the results are shown in Table 1:

[0179] Table 1

[0180]

[0181] The results show that the recognition accuracy of the system at each growth stage is higher than that of traditional manual judgment, with an average increase of 15.6 percentage points, especially in the spike differentiation stage which is difficult to observe with the naked eye. At the same time, the system can predict the transition time of key growth stages in advance, and detect the growth stage transition signal 1.67 days in advance on average.

[0182] The growth abnormality warning effect is shown in Table 2:

[0183] Table 2

[0184]

[0185] Compared with traditional regular inspections, the system detects potential problems an average of 6.3 days in advance, providing ample response time for agricultural intervention.

[0186] The effect of agricultural decision support is shown in Table 3:

[0187] Based on the system warning, agronomists implemented corresponding agricultural operation interventions, and the results showed:

[0188] Table 3

[0189]

[0190] The user experience evaluation is shown in Table 4:

[0191] A satisfaction survey was conducted on 10 agronomists using the system, and the results showed:

[0192] Table 4

[0193]

[0194] The economic benefit analysis is shown in Table 5:

[0195] According to the actual output data of the test area, compared with the control area, the economic benefits of adopting this system are as follows:

[0196] Direct economic benefits: Saving irrigation water costs: 18.3% per mu, equivalent to 56.7 yuan per mu Reducing fertilizer input: 12.6% per mu, equivalent to 38.4 yuan per mu Reducing pesticide use: 26.7% per mu, equivalent to 42.3 yuan per mu Increased yield income: 7.2% per mu, equivalent to 218.6 yuan per mu Premium brought by quality improvement: 4.8% per mu, equivalent to 85.2 yuan per mu;

[0197] Indirect economic benefits: Reduce labor input: reduce labor costs by about 65.4 yuan per mu Reduce environmental pollution: reduce the loss of pesticides and fertilizers, reduce management costs Improve the scientific nature of planting decisions: reduce trial and error costs and risk losses;

[0198] After comprehensive calculation, after adopting this system, the comprehensive economic benefits of rice planting per mu increased by about 506.6 yuan, and the return on investment reached 285%.

[0199] Table 5

[0200]

[0201] The results in Table 5 comprehensively verify the significant effects of this implementation in multiple key technical dimensions. In terms of individual identification and tracking, the performance improvement is more obvious, especially under complex lighting conditions, with an accuracy increase of 42.7%; in terms of three-dimensional structural parameter measurement, the plant height and leaf area measurement accuracy increased by 72.0% and 70.1% respectively, reaching millimeter-level measurement accuracy; in terms of growth dynamics prediction, the prediction error was reduced by 48.7%; in terms of system application efficiency, the completion time of the analysis task was reduced by 90.8%, greatly improving efficiency. The improvement of these technical indicators fully demonstrates the significant technical progress of this implementation in the field of three-dimensional visualization analysis of individual crop growth dynamics.

Claims

1. A visualization analysis method based on agricultural big data model, characterized in that: The following steps are involved: A multi-scale feature extraction algorithm combining recursive neural network and wavelet transform is used to process crop growth time series data and generate representation vectors of multi-time scale features. A deep learning architecture based on variational autoencoders and hidden Markov models performs unsupervised segmentation of the extracted multi-scale time series features to identify key transition points in the growth stage; Construct a standard growth trajectory model and calculate the degree of deviation of current crop growth from the standard trajectory through multi-dimensional deviation metrics; Apply a dynamic threshold model based on reinforcement learning to make early warning trigger judgments and achieve early warning of growth abnormalities; Build a multi-dimensional visualization interface based on the clock metaphor to intuitively display growth stages and warning information.

2. The visualization analysis method based on the agricultural big data model according to claim 1 is characterized in that: The multi-scale feature extraction algorithm for processing crop growth time series data includes: For the original time series dataset: ； Apply multi-level wavelet transform to decompose into different frequency components ,in, , indicating that at a point in time Collected dimensional growth parameter, where Indicates Level of detail factor, Indicates The level approximation coefficient captures the variation patterns on different time scales; Construct a multi-branch bidirectional long short-term memory network to process wavelet coefficients of different scales; The attention fusion module is used to calculate the importance weights of representations at different scales. , and generate a comprehensive multi-scale temporal feature representation vector .

3. The visualization analysis method based on the agricultural big data model according to claim 1 is characterized in that: The deep learning architecture based on variational autoencoder and hidden Markov model includes: Multi-scale time series features Input the variational autoencoder and map it to the latent space through the encoder network to obtain the latent variables The mean and variance ; The latent variables are transformed through the decoder network Reconstruct into original feature representation; Constructing Hidden Markov Models for Latent Variables The time series is divided into states; The Baum-Welch algorithm is used to optimize the HMM parameters, and the Viterbi algorithm is used to decode the most likely state sequence.

4. The visualization analysis method based on the agricultural big data model according to claim 1 is characterized in that: The method of constructing a standard growth trajectory model based on statistical learning and calculating the degree of deviation of the current crop growth from the standard trajectory through a multi-dimensional deviation metric includes: Based on the historical growth data of high-quality crops, a functional data analysis method is used to fit the standard growth trajectory: ； in, For the The standard trajectory model of the growth stage, is the B-spline basis function, is the corresponding coefficient, is the residual term, is the number of basis functions; For each growth indicator dimension, the confidence interval boundaries are calculated to form a dynamic envelope; Calculate the multidimensional deviation metric between the currently monitored crop growth trajectory and the standard trajectory ; Calculate the time cumulative deviation index Assess the severity of ongoing deviations.

5. The visualization analysis method based on the agricultural big data model according to claim 1 is characterized in that: The application of the dynamic threshold model based on reinforcement learning to perform early warning trigger judgment includes: Construct a multi-level threshold model: ； in, Indicates that in the growth stage , Deviation type The next Differentiated warning threshold systems are set for different deviation types and growth stages; Defining the state space is the combination of the current growth stage, deviation measure and historical warning effect, action space The direction and magnitude of the threshold adjustment; Using double Q-learning algorithm to train threshold adjustment strategy ; Based on the calculated deviation metric, cumulative deviation index and deviation type, dynamic thresholds are applied to make early warning trigger judgments.

6. The visualization analysis method based on the agricultural big data model according to claim 1 is characterized in that: The construction of a multi-dimensional visualization interface based on the clock metaphor includes: Construct a clock-style growth stage visualization module to map the entire crop growth period onto a circular clock interface, with each growth stage represented by a different color and angle interval; Design a multi-layer nested information expression structure, which includes the growth stage area, standard growth trajectory area, current growth indicator area, and early warning information area from the outside to the inside; Develop an interactive warning information visualization component to map the warning information onto the clock interface in a visually encoded manner.

7. The visualization analysis method based on the agricultural big data model according to claim 4 is characterized in that: The multidimensional deviation metric is calculated by the following formula: ； in, is a deviation measure in the form of Mahalanobis distance, For the The weight of the dimension indicator, is the currently monitored crop growth trajectory, Indicates Growth stage The standard trajectory of the dimension indicator, is the corresponding standard deviation function.

8. The visualization analysis method based on the agricultural big data model according to claim 5 is characterized in that: The double Q-learning algorithm trains the threshold adjustment strategy through the following formula: ； in, For the status Take action The Q value, is the learning rate, For time The immediate reward at Calculated, is the discount factor, For time The state of being For the status Select the action The maximum Q value of the output is the threshold adjustment strategy after optimization ; ； in, Indicates in status Take action The reward value, are the weight coefficients of each indicator, corresponding to the accuracy, recall, timeliness and false alarm rate of the warning respectively.

9. The visualization analysis method based on the agricultural big data model according to claim 6 is characterized in that: The interactive warning information visualization component uses the following formula to calculate the warning flashing frequency: ； in, is the flashing frequency, is the basic frequency, and To adjust the parameters, and are the urgency and level of the warning, is the base of natural logarithms.

10. A visualization analysis system based on agricultural big data model, characterized in that: A visualization analysis method based on an agricultural big data model for executing any one of claims 1 to 9, comprising a data acquisition module, a multi-scale feature extraction module, a growth stage identification module, a deviation metric calculation module, a dynamic threshold warning module and a visualization interface module, wherein data flow and processing links are formed between the modules, the data acquisition module acquires crop growth time series data, the multi-scale feature extraction module and the growth stage identification module respectively implement feature extraction and stage division, the deviation metric calculation module and the dynamic threshold warning module collaborate to complete anomaly detection and warning triggering, and the visualization interface module integrates analysis results and provides an interactive operation interface.

Images