Deep learning crop growth state identification method for intelligent agriculture and forestry environment

Abstract

The invention relates to the technical field of agricultural crop growth state recognition, in particular to a deep learning crop growth state recognition method for a smart agriculture and forestry environment, which comprises the following steps: receiving multi-source crop growth data input in a smart agriculture and forestry scene, preprocessing and standardizing the input data, and obtaining a multi-source crop growth state recognition result; the problems of data missing and heterogeneity are solved through a data completion mechanism, and a complete and unified crop growth data set and an output-based standardized data set are obtained. The crop growth data are integrated and optimized by constructing a multi-source data input and standardized processing mechanism, a reliable data basis is provided for subsequent accurate recognition, crop growth state recognition is performed through a non-destructive data acquisition and analysis mode, crop growth is prevented from being damaged, the method is suitable for crop whole growth period detection, and the crop growth state recognition efficiency is improved. The identification efficiency is improved by optimizing a model structure and a data processing flow, and the requirements of large-scale agriculture and forestry scene single-batch rapid processing and single-plant real-time identification are met.

Description

Technical Field

[0001] This invention relates to the field of agricultural crop growth status recognition technology, specifically a deep learning method for crop growth status recognition in smart agroforestry environments. Background Technology

[0002] In the fields of smart agriculture and forestry and precision agriculture, crop growth status identification is a core prerequisite for achieving scientific planting and precise regulation. Its accuracy and efficiency directly affect the scientific nature and timeliness of production decisions such as irrigation, fertilization, and pest and disease control. With the development of large-scale and intensive agricultural and forestry planting, higher requirements are placed on the accuracy, real-time performance, comprehensiveness, and scenario adaptability of crop growth status identification.

[0003] Currently, traditional methods for identifying crop growth status mainly rely on manual visual inspection, conventional instrument sampling and testing, empirical environmental correlation judgment, and destructive sampling laboratory testing. Manual visual inspection depends on the operator's planting experience and subjective judgment. Different personnel have different judgment standards, resulting in large errors in the identification results. Moreover, it requires on-site observation of each field and each plant, which is extremely inefficient and cannot meet the needs of full coverage detection in large-scale agricultural and forestry scenarios. In addition, it can only identify obvious external morphological abnormalities, and it is difficult to capture early physiological stress signals. The identification results are lagging and may miss the best time for intervention.

[0004] Conventional instrument sampling and testing requires multiple cumbersome steps, including sample collection, instrument calibration, on-site testing, or laboratory pretreatment, which is time-consuming and cannot achieve real-time rapid identification. Furthermore, the results of sampling and testing only reflect local conditions and are difficult to represent the uniformity of crop growth across an entire farmland. The limited testing coverage can lead to biased production decisions. Additionally, specialized instruments are mostly single-indicator testing devices, unable to acquire multi-dimensional data simultaneously, resulting in a single identification dimension. Empirical environmental correlation judgments rely on indirect inferences based on local planting experience; this experience is highly regional and crop-specific, with extremely poor universality and inability to adapt to different climates. The identification of climate, soil, and crop varieties is not feasible, and there is a lack of scientific testing data and quantitative analysis basis. The accuracy of judgment is extremely low, the anti-interference ability is weak, and the identification results are prone to failure in the face of sudden situations such as extreme weather and new pests and diseases. Destructive sampling laboratory testing can cause irreversible damage to crop growth, and it is especially unsuitable for fruit and cash crops during their growth period. In addition, the testing cycle is long, the response speed is slow, and it cannot meet the needs of timely regulation. At the same time, the testing cost is high and the process is complicated, so it is only suitable for small-scale scientific research experiments and it is difficult to promote it to large-scale farmland production. Moreover, it can only obtain static index data and cannot achieve dynamic continuous monitoring.

[0005] In summary, traditional crop growth status identification methods generally suffer from drawbacks such as strong subjectivity, low identification accuracy, poor operational efficiency, limited coverage, insufficient real-time performance, poor scene adaptability, easy damage to crops, and difficulty in large-scale application. They cannot meet the needs of large-scale, precise, and real-time crop growth status identification in smart agriculture and forestry environments.

[0006] Based on this, the present invention provides a deep learning method for identifying crop growth status in smart agroforestry environments to solve the aforementioned technical problems. Summary of the Invention

[0007] The purpose of this invention is to provide a deep learning-based crop growth status recognition method for smart agroforestry environments, thereby solving the problems mentioned in the background.

[0008] To achieve the above objectives, the present invention provides the following technical solution: This invention proposes a deep learning-based method for identifying crop growth status in smart agroforestry environments, comprising the following steps: S1. Receive multi-source crop growth data input in smart agriculture and forestry scenarios, preprocess and standardize the input data, handle data missing and heterogeneous issues through data completion mechanism, and obtain a complete and unified crop growth dataset; S2. Based on the standardized dataset output by S1, a deep learning model is used to extract multimodal features, including crop morphological features, physiological and biochemical features, environmental correlation features, and stress anomaly features, to generate a structured feature set. S3. Using the structured feature set generated in S2 as input, construct a multimodal fusion growth state identification model. Through feature weighted fusion and hierarchical reasoning, achieve comprehensive identification of crop growth stage, health level, stress type and yield potential. S4. Combining the recognition model output of S3 with the characteristics of smart agriculture and forestry scenarios, establish an agricultural and forestry environment adaptation system to achieve accurate adaptation of the model to different crop varieties, terrain conditions, climate types and planting patterns, and generate scenario-based recognition parameter configurations. S5. Using the scenario-based configuration of S4, the recognition results of S3 are verified for accuracy, real-time performance and robustness. If they fail, the feature extraction parameters of S2 and the model fusion weights of S3 are adjusted in reverse based on the verification results. If they pass, multi-dimensional recognition results and API interfaces are output. S6. Record the key parameters and performance indicators throughout the entire process from S1 to S5. Combine the verification feedback from S5 with the actual feedback from agricultural and forestry production to continuously optimize the data completion rules of S1, the feature weight algorithm of S2, the model structure of S3, and the scenario adaptation strategy of S4, so as to achieve a dynamic iterative closed loop of the technical solution.

[0009] Preferably, the implementation process of step S1 is as follows: It provides a multi-source input interface for drone aerial images, ground fixed camera images, multispectral sensor data, and environmental sensor data, and performs noise reduction, enhancement, and registration and alignment preprocessing on the input data in sequence. Core growth characteristics are selected using a crop feature weighting algorithm, as shown in equation (1): ; In the formula, For the first In the crop data sample, the first The weight of each growth feature is assigned. The higher the weight value, the greater the likelihood that the feature is a core element to be identified. The threshold is set to 0.7. If the value exceeds the threshold, it is determined to be a core feature. For the first In the data sample, the first The word frequencies of each growth characteristic are shown in equation (2): ; In the formula, For the first The feature in the first Frequency of occurrence in each sample For the first The total number of features in a sample reflects the importance of the features in the current sample; For the first The inverse sample frequency of each growth feature is given by equation (3): ; In the formula, The total number of crop data samples in the field of smart agriculture and forestry, including the first... The number of samples for each growth characteristic reflects the universality of the characteristic in the agricultural and forestry fields; the lower the universality, the greater the universality. The higher the value, the more likely it is to be a unique growth characteristic of the crop; This is the matching coefficient for agricultural and forestry features, with a value ranging from 0.8 to 1.2. If the feature is a core feature unique to agriculture and forestry, then... ,otherwise The weighting is used to strengthen the unique characteristics of agriculture and forestry; This is the correlation coefficient between growth status and crop growth status, ranging from 0.9 to 1.1. If the feature is directly correlated with the crop growth status, then... ,otherwise , used to strengthen the weights of features related to growth state; For data samples lacking core features, the default information is automatically completed based on the agricultural and forestry knowledge base optimized by S6 iteration and the feature distribution of similar samples, ensuring the integrity of the dataset. The completion rules are continuously updated with the iteration of S6.

[0010] Preferably, the implementation process of step S2 is as follows: Based on the feature weights calculated by S1, features are selected as core identification features, and the morphological feature set, physiological feature set, environmental feature set and stress feature set are determined by feature type classification. Physiological relationships between features are inferred by calculating the confidence level of growth feature associations. The final output is a structured feature set in JSON format containing core feature types, feature parameters, and physiological relationships. It clearly defines the numerical range, weight ratio, and association confidence of each feature. This set is directly used as the core input for building the S3 model.

[0011] Preferably, the implementation process of step S3 is as follows: A multimodal growth state recognition model is constructed using a CNN-Transformer fusion architecture. The CNN module uses a ResNet50 network to extract the spatial dimension information of image morphological features and stress features in S2, while the Transformer module extracts the temporal correlation information of multispectral physiological features and environmental features in S2. Feature association confidence based on S2 calculation Feature fusion weights are generated by combining the contribution of feature recognition. The model training process uses a labeled dataset of agricultural and forestry crop growth status for supervised training, and the loss function is the cross-entropy loss function; the optimizer is the Adam optimizer, with the initial learning rate set to 0.001 and dynamically adjusted through a learning rate decay strategy; The model output includes crop growth stage, health level, stress type, predicted yield potential, and quantitative value of stress impact, achieving a comprehensive qualitative and quantitative identification of growth status. The output serves as the basis for S4 scenario adaptation.

[0012] Preferably, the implementation process of step S4 is as follows: To achieve crop variety adaptation, the corresponding recognition parameter threshold is automatically matched according to the input crop type. The matching process refers to the feature type and numerical range in S2. To achieve terrain condition adaptation, the image acquisition angle correction coefficient and environmental data weight ratio are adjusted. In mountainous scenes, the influence factor of terrain slope on growth status is enhanced. In greenhouse scenes, the weight of temperature, humidity and light inside the greenhouse is strengthened. The parameters are adjusted based on S3 recognition error feedback optimization. To achieve climate type adaptation, optimize the calibration parameters for identifying growth stages based on seasonal changes, enhance the identification weight of water features in arid climate zones, and improve the detection sensitivity of flood stress and disease features in rainy climate zones, the calibration parameters are linked with the feature weight calculation of S2. The model configuration is optimized by calculating the scene adaptation coefficient, as shown in equation (4): ; In the formula, This is the scene adaptation coefficient, with a value ranging from 0 to 1. The crop variety matching degree is calculated based on the feature similarity between the input crop and the benchmark database, with a value ranging from 0 to 1. The similarity calculation is shown in equation (5): ; In the formula, To input crop feature values, The feature values ​​of the benchmark library; The terrain condition matching degree is calculated based on the degree of adaptation between the planting terrain and scene parameters, and the value ranges from 0 to 1. Climate type matching degree is calculated based on the fit between regional climate data and model climate parameters, with a value ranging from 0 to 1. Apply the corresponding scenario configuration directly. Based on the interpolation optimization of parameters from neighboring scenes, personalized scene adaptation configuration information is generated, which directly affects the model inference process of S3.

[0013] Preferably, the implementation process of step S5 is as follows: The accuracy of the growth state recognition results output by S3 is verified, and the recognition accuracy, recall, and F1 score are calculated. The verification pass criteria were set as follows: growth stage identification accuracy ≥ 95%, health level identification accuracy ≥ 93%, and stress type identification accuracy ≥ 90%. Real-time verification was performed by statistically analyzing the processing time of a single batch of data and the identification time of a single crop. The requirement was that the processing time of a single batch of data should be ≤10 seconds and the identification time of a single crop should be ≤0.5 seconds to meet the real-time detection needs of large-scale agricultural and forestry scenarios. The real-time performance indicators were fed back to the feature extraction engine of S7 for optimization. Robustness verification was conducted by simulating extreme conditions such as changes in illumination, occlusion interference, and sensor noise, and the fluctuation range of the recognition results was calculated. A growth state correlation verification mechanism is constructed. Based on the feature physiological correlation relationship of S2, the logical consistency between growth stage and physiological characteristics and the causal correlation between stress type and environmental data are checked. When there is a contradiction, the parameter adjustment process is triggered to optimize the model inference logic of S3 and the scenario adaptation coefficient of S4 in reverse. After successful verification, the identification model parameters and scene configuration information are stored in a distributed database, and an Excel-formatted identification report, a visualized growth status map, a one-click deployment model file, and a RESTful API interface are output, supporting integration with smart agricultural and forestry management platforms and irrigation and fertilization control systems. Error data and abnormal cases generated during the verification process serve as core samples for S6 iterative optimization.

[0014] Preferably, the deep learning feature extraction engine used in step S2 is implemented as follows: EfficientNetV2 and ViT-Llama2 were selected as the base models. The models were fine-tuned based on the labeled dataset in the field of smart agriculture and forestry. The fine-tuning process used the scene adaptation parameters of S4 to divide the training subset. The goal was to achieve an accuracy of ≥96% for morphological feature extraction, ≥94% for physiological feature extraction, and ≥93% for stress feature extraction. The extraction accuracy was calculated based on the accuracy verification method of S5. The model quantization and distillation techniques are used to optimize performance, compressing the model volume to less than 40% of the original volume, while ensuring that the recognition accuracy decreases by ≤2%. The optimized model performance directly improves the real-time performance of S1 data processing and S3 model inference. A collaborative mechanism between the main model and sub-models is established. The main model is used for global feature extraction, while the sub-models focus on feature refinement for different crop types. The accuracy of recognition in complex scenarios is improved by result fusion. The fusion result refers to the feature fusion weighting algorithm of S3. We import a dictionary of growth characteristics and physiological association rules from the agricultural and forestry fields, and regularly collect growth data of different production areas and different crops through the S6 iterative process to update the dictionary and rules, continuously improving the targeting and accuracy of feature extraction. The updated rules then feed back into the feature classification and association confidence calculation in S2.

[0015] Preferably, the implementation process of the scene adaptation engine in step S4 is as follows: Configure four-level adaptation rules for crop varieties, terrain conditions, climate types, and planting patterns. The rules include triggering conditions, adaptation parameters, and priority levels 1 to 10. The priority of core crops and core scenarios is set to level 1 to 3. The triggering results of the rules directly affect the adjustment of the model parameters in S3. It has a built-in best practice library for smart agriculture and forestry planting, which covers the growth cycle standards, physiological parameter thresholds and stress response mechanisms of crops in different production areas. It is automatically applied to the calibration of identification parameters. The calibration process combines the characteristic value range of S2 with the identification result error of S3. To handle special planting scenarios, a feature separation algorithm is used for compound crop scenarios, indoor environmental factor weights are optimized for facility cultivation scenarios, and natural stress feature recognition is strengthened for organic planting scenarios. It provides rule templates and visual configuration tools, supports users to customize crop types and scene parameters, rules support hot loading and change log recording, and take effect without restarting the system. Custom rules can be incorporated into the basic adaptation rule library after being verified and optimized through the S6 iteration process.

[0016] Preferably, the output and integration module implementation process in step S5 is as follows: Generate model deployment artifacts and optimize deployment configurations based on S4 scene adaptation coefficients; Generate application integration artifacts, where the functional logic of the integration code corresponds one-to-one with the identification output fields of S3; Based on the identification results of S3 and the scenario adaptation configuration of S4, precision agricultural regulation suggestions are generated, including irrigation amount calculation, fertilizer formula calculation based on the difference between nutrient accumulation and soil nutrient content in S2, and matching the stress type identified by S3 with the pest and disease control schemes in the best practice library. Generate model test cases, including unit test cases, integration test cases, and performance test cases, and support exporting to JUnit and PyTest formats. The verification standards for test cases are based on S5's accuracy, real-time performance, and robustness metrics. It enables synchronized management of model and application versions, supports multi-condition querying of identification records and model parameters by crop type, planting area, and time range, and can export the query results to an Excel format analysis report. The report data provides quantitative basis for the iterative optimization of S6.

[0017] Preferably, the dynamic iterative optimization implementation process of step S6 is as follows: Collect feedback from agricultural and forestry production users, including suggestions for correcting identification results, needs for scenario adaptation and optimization, and feedback on the control effect. Establish a feedback scoring mechanism. Feedback with a score of ≥4 is considered valid feedback. Valid feedback is associated with the optimization points of the corresponding steps S1-S5 according to its type. After every 20 batches of recognition results are generated, based on the effective feedback and the recognition performance index of S5, the feature weight calculation coefficient of S2, the model fusion weight ratio of S3, and the scene adaptation coefficient weight of S4 are calibrated to optimize the decision logic of the rule engine. Regularly import growth data of newly added crop varieties and crop growth cases under extreme climates to expand the feature set of S2 and the scenario adaptation range of S4. After the expanded features and scenarios are verified for effectiveness through the S5 verification process, they are incorporated into the basic dataset and adaptation rule base. Establish a model performance degradation monitoring mechanism, and statistically analyze the recognition accuracy of S5 in batches. When the recognition accuracy drops by more than 3% for three consecutive batches, the model retraining process is automatically triggered. The model parameters are updated based on the latest dataset, including the error samples accumulated by S5 and the new samples. The retraining process uses the training algorithm and target index of S3. The model recognition accuracy is improved through continuous iteration, and the parameters and rules after iteration are updated synchronously to each step from S1 to S5.

[0018] Compared with the prior art, the beneficial effects of the present invention are: This invention integrates and optimizes crop growth data by constructing a multi-source data input and standardized processing mechanism, improving data integrity and consistency and providing a reliable data foundation for subsequent accurate identification. It employs a deep learning model for multimodal feature extraction, comprehensively capturing multi-dimensional features such as crop morphology, physiological and biochemical properties, environmental correlations, and stress anomalies, overcoming the limitations of traditional methods with their single identification dimension. By constructing a CNN-Transformer fusion architecture-based multimodal growth state identification model for feature weighted fusion and hierarchical inference, it achieves comprehensive and accurate identification of crop growth stages, health levels, stress types, and yield potential, reducing subjective errors and improving identification accuracy. Furthermore, by establishing a four-level scene adaptation system for parameter matching across different crop varieties, terrain conditions, climate types, and planting patterns, it enhances the method's scene adaptability, overcoming the regional and specific limitations of traditional methods. The verification mechanism and parameter reverse adjustment strategy are used to verify and optimize the recognition results, ensuring their reliability and stability and guaranteeing recognition accuracy in complex agricultural and forestry environments. A dynamic iterative optimization mechanism based on feedback data is established to continuously update model parameters and rules, improving the method's adaptability and evolutionary capabilities to meet the dynamic changes in smart agricultural and forestry environments. Standardized deployment products and integration interfaces enable rapid implementation of the recognition model, achieving seamless integration with smart agricultural and forestry management platforms and control systems, providing efficient technical support for precision agricultural production decisions. Non-destructive data collection and analysis methods are used to identify crop growth status, avoiding damage to crop growth and making it suitable for monitoring the entire crop growth period. Optimized model structure and data processing flow improve recognition efficiency, meeting the needs of rapid single-batch processing and real-time single-plant identification in large-scale agricultural and forestry scenarios, thus improving the timeliness of production control. Attached Figure Description

[0019] Figure 1 This is a flowchart of the deep learning crop growth status recognition method for smart agroforestry environments according to the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] For examples, please refer to Figure 1In practical applications, this invention proposes a deep learning-based crop growth status recognition method for smart agroforestry environments, specifically including the following steps: S1. Receive multi-source crop growth data input in smart agriculture and forestry scenarios, preprocess and standardize the input data, handle data missing and heterogeneous issues through data completion mechanism, and obtain a complete and unified crop growth dataset; It should also be noted that the implementation process of step S1 is as follows: It provides a multi-source input interface for drone aerial images, ground fixed camera images, multispectral sensor data, and environmental sensor data, and performs noise reduction, enhancement, and registration and alignment preprocessing on the input data in sequence. Core growth characteristics are selected using a crop feature weighting algorithm, as shown in equation (1): ; In the formula, For the first In the crop data sample, the first The weight of each growth feature is assigned. The higher the weight value, the greater the likelihood that the feature is a core element to be identified. The threshold is set to 0.7. If the value exceeds the threshold, it is determined to be a core feature. For the first In the data sample, the first The word frequencies of each growth characteristic are shown in equation (2): ; In the formula, For the first The feature in the first Frequency of occurrence in each sample For the first The total number of features in a sample reflects the importance of the features in the current sample; For the first The inverse sample frequency of each growth feature is given by equation (3): ; In the formula, This represents the total number of crop data samples in the field of smart agriculture and forestry. For including the first The number of samples for each growth characteristic reflects the universality of the characteristic in the agricultural and forestry fields; the lower the universality, the greater the universality. The higher the value, the more likely it is to be a unique growth characteristic of the crop; This is the matching coefficient for agricultural and forestry features, ranging from 0.8 to 1.2. If the feature is a core feature unique to agriculture and forestry, such as spectral reflectance, leaf texture, or soil nutrient correlation features, then... ,otherwise The weighting is used to strengthen the unique characteristics of agriculture and forestry; The growth status correlation coefficient ranges from 0.9 to 1.1. If the characteristic is directly related to crop health status, stress type, growth status, chlorophyll content, and pest / disease spot characteristics, then... ,otherwise , used to strengthen the weights of features related to growth state; For data samples lacking core features, the default information is automatically completed based on the agricultural and forestry knowledge base optimized by S6 iteration and the feature distribution of similar samples, ensuring the integrity of the dataset. The completion rules are continuously updated with the iteration of S6.

[0022] In this step, the dataset serves as the basic input for subsequent feature extraction, and its data integrity and standardization directly determine the accuracy of feature extraction. S2. Based on the standardized dataset output by S1, a deep learning model is used to extract multimodal features, including crop morphological features, physiological and biochemical features, environmental correlation features, and stress anomaly features, to generate a structured feature set. It should also be noted that the implementation process of step S2 is as follows: Feature weights calculated based on S1 ,filter The features are used as core identification features, and the morphological feature set, physiological feature set, environmental feature set and stress feature set are determined by feature type classification; The morphological feature set includes plant height, leaf area, stem diameter, number of leaves, fruit size, and plant compactness, which are extracted through image segmentation and pixel statistics algorithms; The set of physiological characteristics includes chlorophyll content, water content, nutrient accumulation, and photosynthetic efficiency, which are extracted using a multispectral data inversion algorithm. The set of environmental characteristics includes soil temperature and humidity, soil nutrient content, light intensity, air temperature and humidity, and cumulative precipitation, which are obtained after calibration using environmental sensor data. The stress feature set includes leaf yellowing area, number of disease and pest spots, degree of wilting, and deformity rate, which are extracted through an abnormal area detection algorithm; The physiological relationships between features are inferred by calculating the confidence level of growth feature associations. The algorithm formula is as follows: ; In the formula, Features With features The association confidence level, with a value ranging from 0 to 1. Determined to be a strong association, It was determined to be a weak association. Determined to be unrelated; Features With features The number of times a data sample co-occurs (i.e., the number of samples that appear at the same time); Features With features The physiological semantic association degree is calculated based on the semantic similarity score of the pre-trained language model in the agricultural and forestry field, with a value range of 0 to 1; Features Total number of occurrences in all data samples Features The total number of occurrences in all data samples is used to normalize the co-occurrence count and avoid interference from high-frequency features in the results; The final output is a structured feature set in JSON format containing core feature types, feature parameters, and physiological relationships. It clearly defines the numerical range, weight ratio, and association confidence of each feature. This set is directly used as the core input for building the S3 model.

[0023] It should also be noted that the deep learning feature extraction engine used in step S2 is implemented as follows: EfficientNetV2 and ViT-Llama2, with semantic understanding accuracy ≥92% and inference speed ≤0.3 seconds / sample, were selected as the base models. The performance of the models directly affects the feature extraction efficiency and accuracy of S2. The model was fine-tuned based on a labeled dataset in the field of smart agriculture and forestry, including 100,000+ crop image samples, 50,000+ multispectral data samples, and 30,000+ environmental sensor data samples. The fine-tuning process used the scene adaptation parameters of S4 to divide the training subset. The goal was to achieve an accuracy of ≥96% for morphological feature extraction, ≥94% for physiological feature extraction, and ≥93% for stress feature extraction. The extraction accuracy was calculated based on the accuracy verification method of S5. The model quantization and distillation techniques are used to optimize performance, compressing the model volume to less than 40% of the original volume, while ensuring that the recognition accuracy decreases by ≤2%. The optimized model performance directly improves the real-time performance of S1 data processing and S3 model inference. A collaborative mechanism between the main model and sub-models is established. The main model is used for global feature extraction, while the sub-models focus on feature refinement for different crop types. The accuracy of recognition in complex scenarios is improved by result fusion. The fusion result refers to the feature fusion weighting algorithm of S3. We import a dictionary of growth characteristics and physiological association rules from the agricultural and forestry fields, and regularly collect growth data of different production areas and different crops through the S6 iterative process to update the dictionary and rules, continuously improving the targeting and accuracy of feature extraction. The updated rules then feed back into the feature classification and association confidence calculation in S2.

[0024] Through this step, the feature set carries feature weights, type classifications, and correlation information, providing core support for subsequent model construction; S3. Using the structured feature set generated in S2 as input, construct a multimodal fusion growth state identification model. Through feature weighted fusion and hierarchical reasoning, achieve comprehensive identification of crop growth stage, health level, stress type and yield potential. It should also be noted that the implementation process of step S3 is as follows: A multimodal growth state recognition model is constructed using a CNN-Transformer fusion architecture. The CNN module uses a ResNet50 network to extract the spatial dimension information of image morphological features and stress features in S2, while the Transformer module extracts the temporal correlation information of multispectral physiological features and environmental features in S2. Feature association confidence based on S2 calculation The feature fusion weights are generated by combining the contribution of feature recognition. The algorithm formula is as follows: ; In the formula, The weights for feature fusion range from 0 to 1. Features The recognition accuracy rate, i.e., the correct recognition rate when the feature is used alone for recognition, ranges from 0 to 1, and is calculated as follows: ; in Features Identify the correct number of samples. Features Identify the number of missed samples; The model training process uses a labeled dataset of agricultural and forestry crop growth status for supervised training, and the loss function is the cross-entropy loss function, the mathematical expression of which is: ; In the formula, The total number of training samples, To identify the number of categories, For the first The sample belongs to the first The actual label of the class (0 or 1). For the first The sample belongs to the first The predicted probability of a class; The optimizer uses the Adam optimizer, with an initial learning rate of 0.001, which is dynamically adjusted using a learning rate decay strategy. The decay formula is as follows: ; In the formula, For the first Learning rate during round training The initial learning rate, The attenuation coefficient; The model output includes crop growth stages: sowing period, seedling stage, jointing stage, flowering stage, grain filling stage, and maturity stage; Health level: Excellent, Good, Very Good, Average, Poor, Poor; Types of stress: drought stress, flood stress, nutrient deficiency stress, pest and disease stress, and pesticide damage stress; The predicted yield potential and the quantitative value of the degree of stress impact enable a comprehensive qualitative and quantitative identification of the growth status, and the output results serve as the basis for S4 scenario adaptation.

[0025] Through this step, the model integrates weights and inference logic, which are dynamically adjusted based on the feature association confidence in S2. S4. Combining the recognition model output of S3 with the characteristics of smart agriculture and forestry scenarios, establish an agricultural and forestry environment adaptation system to achieve accurate adaptation of the model to different crop varieties, terrain conditions, climate types and planting patterns, and generate scenario-based recognition parameter configurations. It should also be noted that the implementation process of step S4 is as follows: To achieve crop variety adaptation, a built-in benchmark library of growth characteristics of common agricultural and forestry crops such as rice, wheat, corn, cotton, vegetables, and fruit trees is used. The corresponding recognition parameter threshold is automatically matched according to the input crop type. The matching process refers to the feature type and numerical range in S2. To achieve terrain adaptation, the image acquisition angle correction coefficient and environmental data weight ratio are adjusted for different terrains and planting scenarios such as plains, mountains, hills, and greenhouses. In mountain scenes, the influence factor of terrain slope on growth status is enhanced, and in greenhouse scenes, the weight of temperature, humidity and light inside the greenhouse is strengthened. The parameters are adjusted based on S3 recognition error feedback optimization. To achieve climate type adaptation, based on tropical, subtropical, temperate, and cold temperate climate zones, the calibration parameters for identifying growth stages by seasonal changes are optimized. The identification weight of water characteristics is increased in arid climate zones, and the detection sensitivity of flood stress and disease characteristics is enhanced in rainy climate zones. The calibration parameters are linked with the feature weight calculation of S2. The model configuration is optimized by calculating the scene adaptation coefficient, as shown in equation (4): ; In the formula, This is the scene adaptation coefficient, with a value ranging from 0 to 1. The crop variety matching degree is calculated based on the feature similarity between the input crop and the benchmark database, with a value ranging from 0 to 1. The similarity calculation is shown in equation (5): ; In the formula, To input crop feature values, The feature values ​​of the benchmark library; The terrain condition matching degree is calculated based on the degree of adaptation between the planting terrain and scene parameters, and the value ranges from 0 to 1. Climate type matching degree is calculated based on the fit between regional climate data and model climate parameters, with a value ranging from 0 to 1. Apply the corresponding scenario configuration directly. Based on the interpolation optimization of parameters from neighboring scenes, personalized scene adaptation configuration information is generated, which directly affects the model inference process of S3.

[0026] It should also be noted that the implementation process of the scene adaptation engine in step S4 is as follows: Configure four-level adaptation rules for crop varieties, terrain conditions, climate types, and planting patterns. The rules include triggering conditions, adaptation parameters, and priority levels 1 to 10. The priority of core crops and core scenarios is set to level 1 to 3. The triggering results of the rules directly affect the adjustment of the model parameters in S3. It has a built-in best practice library for smart agriculture and forestry planting, which covers the growth cycle standards, physiological parameter thresholds and stress response mechanisms of crops in different production areas. It is automatically applied to the calibration of identification parameters. The calibration process combines the characteristic value range of S2 with the identification result error of S3. To handle special planting scenarios, including intercropping, facility cultivation, and organic farming, a feature separation algorithm is used for compound crop scenarios based on S2 feature type classification. For facility cultivation scenarios with fusion weight calculation associated with S3, the weight of indoor environmental factors is optimized. For organic farming scenarios, the identification of natural stress features is strengthened, and the stress feature weight coefficient of S2 is adjusted. It provides rule templates and visual configuration tools, supports users to customize crop types and scene parameters, rules support hot loading and change log recording, and take effect without restarting the system. Custom rules can be incorporated into the basic adaptation rule library after being verified and optimized through the S6 iteration process.

[0027] Through this step, the configuration acts in reverse on the S3 model inference process, optimizing the recognition accuracy in different scenarios; S5. Using the scenario-based configuration of S4, the recognition results of S3 are verified for accuracy, real-time performance and robustness. If they fail, the feature extraction parameters of S2 and the model fusion weights of S3 are adjusted in reverse based on the verification results. If they pass, multi-dimensional recognition results and API interfaces are output. It should also be noted that the implementation process of step S5 is as follows: The accuracy of the growth state recognition results output by S3 is verified by calculating the recognition precision, recall, and F1 score. The precision formula is as follows: ; The recall formula is: ; The formula for F1 score is: ; The verification pass criteria were set as follows: growth stage identification accuracy ≥ 95%, health level identification accuracy ≥ 93%, and stress type identification accuracy ≥ 90%. Real-time verification was performed by statistically analyzing the processing time of a single batch of data and the identification time of a single crop. The requirement was that the processing time of a single batch of data should be ≤10 seconds and the identification time of a single crop should be ≤0.5 seconds to meet the real-time detection needs of large-scale agricultural and forestry scenarios. The real-time performance indicators were fed back to the feature extraction engine of S7 for optimization. Robustness verification was performed by simulating extreme conditions such as changes in illumination, occlusion interference, and sensor noise. The fluctuation range of the recognition results was calculated, and the formula for the fluctuation range is as follows: ; In the formula, The accuracy rate under normal conditions. To ensure recognition accuracy under extreme conditions, the fluctuation range should be ≤5% to guarantee recognition stability in complex environments; A growth state correlation verification mechanism is constructed. Based on the feature physiological correlation relationship of S2, the logical consistency between growth stage and physiological characteristics and the causal correlation between stress type and environmental data are checked. When there is a contradiction, the parameter adjustment process is triggered to optimize the model inference logic of S3 and the scenario adaptation coefficient of S4 in reverse. After successful verification, the recognition model parameters and scene configuration information are stored in a distributed database, and an Excel-format recognition report, a visualized growth status map, a one-click deployment model file, and a RESTful API interface are output, supporting integration with smart agricultural and forestry management platforms and irrigation and fertilization control systems. Error data and abnormal cases generated during the verification process serve as core samples for S6 iterative optimization.

[0028] It should also be noted that the implementation process of the output and integration module in step S5 is as follows: Generate model deployment artifacts, including inference engine scripts, model weight files, and configuration manifests. Supports Docker containerized deployment and cloud-edge collaborative deployment. Deployment configuration is optimized based on S4 scenario adaptation coefficients. Generate application integration artifacts, including data access layer code, API interface code, and front-end visualization components. It supports integration with mainstream development frameworks such as Spring Boot, Flask, Django, and FastAPI. The functional logic of the integration code corresponds one-to-one with the recognition output fields of S3. Based on the recognition results of S3 and the scenario adaptation configuration of S4, precision agriculture regulation suggestions are generated, including irrigation amount calculation, with the following formula: ; In the formula, For crop coefficients, For reference, crop evapotranspiration, For the planting area, the fertilizer formula is calculated based on the difference between the nutrient accumulation and soil nutrient content in S2, and the stress type identified in S3 is matched with the pest and disease control plan in the best practice library. Generate model test cases, including unit test cases, integration test cases, and performance test cases, and support exporting to JUnit and PyTest formats. The verification standards for test cases are based on S5's accuracy, real-time performance, and robustness metrics. It enables synchronized management of model and application versions, supports multi-condition querying of identification records and model parameters by crop type, planting area, and time range, and can export the query results to an Excel format analysis report. The report data provides quantitative basis for the iterative optimization of S6.

[0029] This step validates the data and serves as the basis for model iteration. S6. Record the key parameters and performance indicators throughout the entire process from S1 to S5, and combine the verification feedback from S5 with actual feedback from agricultural and forestry production.

[0030] It should also be noted that the dynamic iterative optimization implementation process of step S6 is as follows: Collect feedback from agricultural and forestry production users, including suggestions for correcting identification results, needs for scenario adaptation and optimization, and feedback on the control effect. Establish a feedback scoring mechanism. Feedback with a score of ≥4 is considered valid feedback. Valid feedback is associated with the optimization points of the corresponding steps S1-S5 according to its type. After every 20 batches of recognition results are generated, the feature weight calculation coefficients of S2 are calibrated based on the effective feedback and the recognition performance index of S5. The model fusion weight ratio of S3 and the scene adaptation coefficient weight of S4 are used to optimize the decision logic of the rule engine. The calibration formula is as follows: ; In the formula, For the new parameters, For the original parameters, To adjust the coefficient, For feedback optimization factors; Regularly import growth data of newly added crop varieties and crop growth cases under extreme climates to expand the feature set of S2 and the scenario adaptation range of S4. After the expanded features and scenarios are verified for effectiveness through the S5 verification process, they are incorporated into the basic dataset and adaptation rule base. Establish a model performance degradation monitoring mechanism, and statistically analyze the recognition accuracy of S5 in batches. When the recognition accuracy drops by more than 3% for three consecutive batches, the model retraining process is automatically triggered. The model parameters are updated based on the latest dataset, including the error samples accumulated by S5 and the new samples. The retraining process uses the training algorithm and target index of S3. Through continuous iteration, the model recognition accuracy is improved by ≥1% every quarter, and the scene adaptability is expanded to more than 5 crop types every six months. The parameters and rules after iteration are updated synchronously to each step of S1-S5, forming a complete technical closed loop to meet the dynamic change needs of smart agriculture and forestry environment.

[0031] Through this step, we continuously optimize the data completion rules in S1, the feature weighting algorithm in S2, the model structure in S3, and the scenario adaptation strategy in S4, thereby achieving a dynamic iterative closed loop for the technical solution.

[0032] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0033] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A deep learning-based crop growth status recognition method for smart agroforestry environments, characterized in that, Includes the following steps: S1. Receive multi-source crop growth data input in smart agriculture and forestry scenarios, preprocess and standardize the input data, handle data missing and heterogeneous issues through data completion mechanism, and obtain a complete and unified crop growth dataset; S2. Based on the standardized dataset output by S1, a deep learning model is used to extract multimodal features, including crop morphological features, physiological and biochemical features, environmental correlation features, and stress anomaly features, to generate a structured feature set. S3. Using the structured feature set generated in S2 as input, construct a multimodal fusion growth state recognition model, and perform feature weighted fusion and hierarchical reasoning. S4. Combining the recognition model output of S3 with the characteristics of smart agriculture and forestry scenarios, establish an agricultural and forestry environment adaptation system to achieve accurate adaptation of the model to different crop varieties, terrain conditions, climate types and planting patterns, and generate scenario-based recognition parameter configurations. S5. Using the scenario-based configuration of S4, the recognition results of S3 are verified for accuracy, real-time performance and robustness. If they fail, the feature extraction parameters of S2 and the model fusion weights of S3 are adjusted in reverse based on the verification results. If they pass, multi-dimensional recognition results and API interfaces are output. S6. Record the key parameters and performance indicators throughout the entire process from S1 to S5. Combine the verification feedback from S5 with the actual feedback from agricultural and forestry production to continuously optimize the data completion rules of S1, the feature weight algorithm of S2, the model structure of S3, and the scenario adaptation strategy of S4.

2. The deep learning-based crop growth status recognition method for smart agroforestry environments according to claim 1, characterized in that, The implementation process of step S1 is as follows: It provides a multi-source input interface for drone aerial images, ground fixed camera images, multispectral sensor data, and environmental sensor data, and performs noise reduction, enhancement, and registration and alignment preprocessing on the input data in sequence. Core growth characteristics are selected using a crop feature weighting algorithm, as shown in equation (1): ； In the formula, For the first In the crop data sample, the first The weight of each growth feature is assigned. The higher the weight value, the greater the likelihood that the feature is a core element to be identified. The threshold is set to 0.

7. If the value exceeds the threshold, it is determined to be a core feature. For the first In the data sample, the first The word frequencies of each growth characteristic are shown in equation (2): ； In the formula, For the first The feature in the first Frequency of occurrence in each sample For the first The total number of features in a sample reflects the importance of the features in the current sample; For the first The inverse sample frequency of each growth feature is given by equation (3): ； In the formula, This represents the total number of crop data samples in the field of smart agriculture and forestry. For including the first The number of samples for each growth characteristic reflects the universality of the characteristic in the agricultural and forestry fields; the lower the universality, the greater the universality. The higher the value, the more likely it is to be a unique growth characteristic of the crop; This is the matching coefficient for agricultural and forestry features, with a value ranging from 0.8 to 1.

2. If the feature is a core feature unique to agriculture and forestry, then... ,otherwise The weighting is used to strengthen the unique characteristics of agriculture and forestry; This is the correlation coefficient between growth status and crop growth status, ranging from 0.9 to 1.

1. If the feature is directly correlated with the crop growth status, then... ,otherwise , used to strengthen the weights of features related to growth state; For data samples lacking core features, the default information is automatically completed based on the agricultural and forestry knowledge base optimized by S6 iteration and the feature distribution of similar samples, ensuring the integrity of the dataset. The completion rules are continuously updated with the iteration of S6.

3. The deep learning-based crop growth status recognition method for smart agroforestry environments according to claim 2, characterized in that, The implementation process of step S2 is as follows: Based on the feature weights calculated by S1, features are selected as core identification features, and the morphological feature set, physiological feature set, environmental feature set and stress feature set are determined by feature type classification. Physiological relationships between features are inferred by calculating the confidence level of growth feature associations. The final output is a structured feature set in JSON format containing core feature types, feature parameters, and physiological relationships. It clearly defines the numerical range, weight ratio, and association confidence of each feature. This set is directly used as the core input for building the S3 model.

4. The deep learning-based crop growth status recognition method for smart agroforestry environments according to claim 3, characterized in that, The implementation process of step S3 is as follows: A multimodal growth state recognition model is constructed using a CNN-Transformer fusion architecture. The CNN module uses a ResNet50 network to extract the spatial dimension information of image morphological features and stress features in S2, while the Transformer module extracts the temporal correlation information of multispectral physiological features and environmental features in S2. Feature association confidence based on S2 calculation Feature fusion weights are generated by combining the contribution of feature recognition. The model training process uses a labeled dataset of agricultural and forestry crop growth status for supervised training, and the loss function is the cross-entropy loss function; the optimizer is the Adam optimizer, with the initial learning rate set to 0.001 and dynamically adjusted through a learning rate decay strategy; The model output includes crop growth stage, health level, stress type, predicted yield potential, and quantitative value of stress impact, achieving a comprehensive qualitative and quantitative identification of growth status. The output serves as the basis for S4 scenario adaptation.

5. The deep learning-based crop growth status recognition method for smart agroforestry environments according to claim 4, characterized in that, The implementation process of step S4 is as follows: To achieve crop variety adaptation, the corresponding recognition parameter threshold is automatically matched according to the input crop type. The matching process refers to the feature type and numerical range in S2. To achieve terrain condition adaptation, the image acquisition angle correction coefficient and environmental data weight ratio are adjusted. In mountainous scenes, the influence factor of terrain slope on growth status is enhanced. In greenhouse scenes, the weight of temperature, humidity and light inside the greenhouse is strengthened. The parameters are adjusted based on S3 recognition error feedback optimization. To achieve climate type adaptation, optimize the calibration parameters for identifying growth stages based on seasonal changes, enhance the identification weight of water features in arid climate zones, and improve the detection sensitivity of flood stress and disease features in rainy climate zones, the calibration parameters are linked with the feature weight calculation of S2. The model configuration is optimized by calculating the scene adaptation coefficient, as shown in equation (4): ； In the formula, This is the scene adaptation coefficient, with a value ranging from 0 to 1. The crop variety matching degree is calculated based on the feature similarity between the input crop and the benchmark database, with a value ranging from 0 to 1. The similarity calculation is shown in equation (5): ； In the formula, To input crop feature values, The feature values ​​of the benchmark library; The terrain condition matching degree is calculated based on the degree of adaptation between the planting terrain and scene parameters, and the value ranges from 0 to 1. Climate type matching degree is calculated based on the fit between regional climate data and model climate parameters, with a value ranging from 0 to 1. The corresponding scenario configuration can be applied directly. Based on the interpolation optimization of parameters from neighboring scenes, personalized scene adaptation configuration information is generated, which directly affects the model inference process of S3.

6. The deep learning-based crop growth status recognition method for smart agroforestry environments according to claim 5, characterized in that, The implementation process of step S5 is as follows: The accuracy of the growth state recognition results output by S3 is verified, and the recognition accuracy, recall rate, and F1 score are calculated. The verification pass criteria were set as follows: growth stage identification accuracy ≥ 95%, health level identification accuracy ≥ 93%, and stress type identification accuracy ≥ 90%. Real-time verification was performed by statistically analyzing the processing time of a single batch of data and the identification time of a single crop. The requirement was that the processing time of a single batch of data should be ≤10 seconds and the identification time of a single crop should be ≤0.5 seconds to meet the real-time detection needs of large-scale agricultural and forestry scenarios. The real-time performance indicators were fed back to the feature extraction engine of S7 for optimization. Robustness verification was conducted by simulating extreme conditions such as changes in illumination, occlusion interference, and sensor noise, and the fluctuation range of the recognition results was calculated. A growth state correlation verification mechanism is constructed. Based on the feature physiological correlation relationship of S2, the logical consistency between growth stage and physiological characteristics and the causal correlation between stress type and environmental data are checked. When there is a contradiction, the parameter adjustment process is triggered to optimize the model inference logic of S3 and the scenario adaptation coefficient of S4 in reverse. After successful verification, the identification model parameters and scene configuration information are stored in a distributed database, and an Excel-formatted identification report, a visualized growth status map, a one-click deployment model file, and a RESTful API interface are output, supporting integration with smart agricultural and forestry management platforms and irrigation and fertilization control systems. Error data and abnormal cases generated during the verification process serve as core samples for S6 iterative optimization.

7. The deep learning-based crop growth status recognition method for smart agroforestry environments according to claim 6, characterized in that, The deep learning feature extraction engine used in step S2 is implemented as follows: EfficientNetV2 and ViT-Llama2 were selected as the base models. The models were fine-tuned based on the labeled dataset in the field of smart agriculture and forestry. The fine-tuning process used the scene adaptation parameters of S4 to divide the training subset. The goal was to achieve an accuracy of ≥96% for morphological feature extraction, ≥94% for physiological feature extraction, and ≥93% for stress feature extraction. The extraction accuracy was calculated based on the accuracy verification method of S5. The model quantization and distillation techniques are used to optimize performance, compressing the model volume to less than 40% of the original volume, while ensuring that the recognition accuracy decreases by ≤2%. The optimized model performance directly improves the real-time performance of S1 data processing and S3 model inference. A collaborative mechanism between the main model and sub-models is established. The main model is used for global feature extraction, while the sub-models focus on feature refinement for different crop types. The accuracy of recognition in complex scenarios is improved by result fusion. The fusion result refers to the feature fusion weighting algorithm of S3. We import a dictionary of growth characteristics and physiological association rules from the agricultural and forestry fields, and regularly collect growth data of different production areas and different crops through the S6 iterative process to update the dictionary and rules, continuously improving the targeting and accuracy of feature extraction. The updated rules then feed back into the feature classification and association confidence calculation in S2.

8. The deep learning-based crop growth status recognition method for smart agroforestry environments according to claim 7, characterized in that, The implementation process of the scene adaptation engine in step S4 is as follows: Configure four-level adaptation rules for crop varieties, terrain conditions, climate types, and planting patterns. The rules include triggering conditions, adaptation parameters, and priority levels 1 to 10. The priority of core crops and core scenarios is set to level 1 to 3. The triggering results of the rules directly affect the adjustment of the model parameters in S3. It has a built-in best practice library for smart agriculture and forestry planting, which covers the growth cycle standards, physiological parameter thresholds and stress response mechanisms of crops in different production areas. It is automatically applied to the calibration of identification parameters. The calibration process combines the characteristic value range of S2 with the identification result error of S3. To handle special planting scenarios, a feature separation algorithm is used for compound crop scenarios, indoor environmental factor weights are optimized for facility cultivation scenarios, and natural stress feature identification is strengthened for organic planting scenarios. It provides rule templates and visual configuration tools, supports users to customize crop types and scene parameters, rules support hot loading and change log recording, and can take effect without restarting the system. Custom rules can be incorporated into the basic adaptation rule library after being verified and optimized through the S6 iteration process.

9. A deep learning method for identifying crop growth status in a smart agroforestry environment according to claim 8, characterized in that, The implementation process of the output and integration module in step S5 is as follows: Generate model deployment artifacts and optimize deployment configurations based on S4 scene adaptation coefficients; Generate application integration artifacts, where the functional logic of the integration code corresponds one-to-one with the identification output fields of S3; Based on the identification results of S3 and the scenario adaptation configuration of S4, precision agricultural regulation suggestions are generated, including irrigation amount calculation, fertilizer formula calculation based on the difference between nutrient accumulation and soil nutrient content in S2, and matching the stress type identified by S3 with the pest and disease control schemes in the best practice library. Generate model test cases, including unit test cases, integration test cases, and performance test cases, and support exporting to JUnit and PyTest formats. The verification standards for test cases are based on S5's accuracy, real-time performance, and robustness metrics. It enables synchronized management of model and application versions, supports multi-condition querying of identification records and model parameters by crop type, planting area, and time range, and can export the query results to an Excel format analysis report. The report data provides quantitative basis for the iterative optimization of S6.

10. A deep learning method for identifying crop growth status in a smart agroforestry environment according to claim 9, characterized in that, The dynamic iterative optimization implementation process of step S6 is as follows: Collect feedback from agricultural and forestry production users, including suggestions for correcting identification results, needs for scenario adaptation and optimization, and feedback on the control effect. Establish a feedback scoring mechanism. Feedback with a score of ≥4 is considered valid feedback. Valid feedback is associated with the optimization points of the corresponding steps S1-S5 according to its type. After every 20 batches of recognition results are generated, based on the effective feedback and the recognition performance index of S5, the feature weight calculation coefficient of S2, the model fusion weight ratio of S3, and the scene adaptation coefficient weight of S4 are calibrated to optimize the decision logic of the rule engine. Regularly import growth data of newly added crop varieties and crop growth cases under extreme climates to expand the feature set of S2 and the scenario adaptation range of S4. After the expanded features and scenarios are verified for effectiveness through the S5 verification process, they are incorporated into the basic dataset and adaptation rule base. Establish a model performance degradation monitoring mechanism, and statistically analyze the recognition accuracy of S5 in batches. When the recognition accuracy drops by more than 3% for three consecutive batches, the model retraining process is automatically triggered. The model parameters are updated based on the latest dataset, including the error samples accumulated by S5 and the new samples. The retraining process uses the training algorithm and target index of S3. The model recognition accuracy is improved through continuous iteration, and the parameters and rules after iteration are updated synchronously to each step from S1 to S5.

Images