A monitoring system and method for diseases of mixed beans

Abstract

This invention relates to the field of monitoring and control technology, specifically disclosing a monitoring system and method for diseases of mixed beans, including a UAV remote sensing acquisition module, an image preprocessing and feature extraction module, an expert teaching and path recording module, a trajectory replay and servo execution module, a gradient ensemble classification module, a dynamic abnormal sample mining and adaptive reinforcement learning module, and a visualization and control scheduling module. This invention achieves spatial localization, hierarchical identification, and closed-loop scheduling of control tasks for CBB lesions in red kidney beans at different stages (early, middle, and late) through hyperspectral multi-resolution data acquisition, expert interactive path standardization script generation and automatic reproduction, priority capture of dynamic abnormal samples and multi-terminal expert remote collaborative verification, feature screening and gradient ensemble modeling, and intelligent generation of variable prescription maps.

Description

Technical Field

[0001] This invention relates to the field of monitoring and control technology, and in particular to a monitoring system and method for diseases of miscellaneous beans. Background Technology

[0002] Existing methods for monitoring diseases in mixed beans, especially remote sensing methods for monitoring common bacterial blight (CBB, caused by Xanthomonas aeruginosa) in red kidney beans, often rely on static thresholds or end-to-end deep learning models based on hyperspectral or multispectral imagery for lesion identification. These methods lack real-time recording and replay capabilities for expert demonstrations of key operations such as multispectral-texture feature extraction window selection, threshold parameter setting, and model fine-tuning. Furthermore, they fail to establish a closed-loop feedback mechanism between the recorded operation sequence, extraction window position, and parameter settings to automatically update membership functions and fuzzy rule bases. Therefore, when faced with spectral-texture feature drift caused by differences in different mixed bean varieties, growth stages, or spatial resolution adjustments, these methods are ineffective. The system cannot reproduce the optimal identification process through historical teaching scripts, resulting in the misjudgment or omission of atypical lesions that are water-soaked, have blurred boundaries, and are small (<0.5mm) in the early stages. This leads to a significant decrease in classification confidence and generalization ability, high retraining costs, and difficulty in meeting the needs of continuous monitoring across multiple periods at the field level. Therefore, it is urgent to build an adaptive control framework based on a loop of operations that can be recorded and played back. This framework would capture in real time the operations performed by experts in lesion identification, such as feature extraction, threshold and parameter adjustment, and convert them into a playable script. The parameters would be dynamically updated in scenarios involving multiple varieties, multiple resolutions, and multiple growth stages, thereby improving the identification confidence and generalization performance of early atypical CBB lesions and supporting continuous monitoring across multiple periods at the field level. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide a monitoring system and method for diseases of mixed beans. This system acquires canopy spectral and spatial information through hyperspectral multi-resolution data acquisition using a pre-programmed control loop. The system records drag-and-drop identification trajectories, threshold parameters, and model fine-tuning operations in real time as standardized scripts. Trajectory replay and servo execution engines accurately reproduce the taught path through digital control commands. Abnormal sampling commands are triggered based on confidence levels and label conflicts, and the verification results are replayed to an online incremental training library. A hierarchical identification model is generated based on multi-scale feature fusion. The identification and hierarchical results are converted into a closed-loop scheduling script and drive precise spraying execution. This achieves high-precision spatial positioning, hierarchical identification, and automated closed-loop control of CBB lesions in red kidney beans at early, mid, and late stages.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A bean disease monitoring system includes a UAV remote sensing acquisition module, an image preprocessing and feature extraction module, an expert teaching and path recording module, a trajectory replay and servo execution module, a gradient ensemble classification module, a dynamic anomaly sample mining and adaptive reinforcement learning module, and a visualization and control scheduling module. The UAV remote sensing acquisition module digitally drives the UAV along a pre-programmed flight path to acquire hyperspectral and attitude data of the red kidney bean canopy and generate flight commands. The image preprocessing and feature extraction module performs radiometric correction and geometric registration, uses a sliding window to scan vegetation indices and gray-level co-occurrence matrices to extract sensitive features of lesions and locate their coordinates, and the expert teaching and path recording module uses a human-machine collaborative approach to drag the lesion identification trajectory on the feature map. The real-time recording of window position, threshold, and model parameter adjustment is used as a replay script. The trajectory replay and servo execution module calls the servo driver to perform servo positioning and scanning of the window or slide rail according to the script, so as to reproduce the diagnosis process in the new field. The gradient ensemble classification module takes multi-scale features and teaching trajectory data as input, outputs the severity of CBB disease in red kidney beans and feeds back the confidence level through a rule-based ensemble strategy. The dynamic abnormal sample mining and adaptive reinforcement learning module dynamically captures atypical samples based on confidence level and label conflict. After expert review, the samples are included in the incremental training set to update the membership function and fuzzy rules. The visualization and control scheduling module renders the results as disease hotspots and calls the agricultural machinery servo system to spray pesticides according to the pre-recorded spraying trajectory and sends back the log.

[0006] As a further embodiment of the system of the present invention, the UAV remote sensing acquisition module is equipped with a pushbroom hyperspectral camera with a spectral range of 400 to 1000 nanometers, a band spacing of 2.1 nanometers, and a spatial resolution adjustable between 0.015 meters and 0.2 meters. It also integrates an inertial measurement unit and a global navigation satellite system. Kalman filtering is used to achieve attitude-positioning data fusion and generate spectral-pose metadata with timestamps. Under normal operating conditions, the UAV remote sensing acquisition module automatically adjusts the acquisition step and flight speed of the pushbroom hyperspectral camera according to the spatial resolution requirements set by the expert teaching and path recording module. This enables continuous imaging of the path of lesions marked by the expert and the corresponding feature extraction area. When the dynamic anomaly sample mining and adaptive reinforcement learning module detects difficult areas or atypical lesion targets, it triggers a high-density acquisition mode, reducing the acquisition step to 0.015 meters, increasing the number of spectral samples per pixel to all available channels, and reducing the flight speed to 0.5 meters per second. All acquisition parameters and corresponding lesion trajectories are written into the metadata along with the timestamps and pose information.

[0007] As a further aspect of the system of this invention, after receiving the hyperspectral raw data, the image preprocessing and feature extraction module first automatically segments and removes soil and shadow pixels based on a normalized vegetation index threshold of 0.62. Then, the background-removed image is resampled to four spatial resolutions of 0.03m, 0.05m, 0.1m, and 0.15m using nearest neighbor interpolation. Based on expert annotations and path information for the early, middle, and late stages of CBB disease, the optimal resolution consistent with the teaching trajectory is automatically selected for feature extraction. When working with the dynamic anomaly sample mining and adaptive reinforcement learning modules to push out difficult areas, the image preprocessing and feature extraction module automatically switches to the minimum resolution and maximum sliding window according to the specific morphology and boundary features of the identified atypical lesions, and simultaneously extracts multi-scale vegetation indices and gray-level co-occurrence matrix texture features. All feature extraction processes are associated with and stored with the expert teaching trajectory and dynamic sample positioning data.

[0008] As a further embodiment of the system of the present invention, the expert teaching and path recording module supports plant protection experts to draw identification paths on the pre-processed feature map using interactive methods such as dragging, box selection, or multi-point drawing, according to the actual lesion distribution. It records the pixel coordinates, feature window size, band channel, expert interpretation timestamp, operation sequence, and servo position feedback of each trajectory node in real time. The feature window size ranges from 3×3 to 11×11. The above paths and parameters are encapsulated in a standardized structured data format to generate a reusable JSON / XML script. The reusable script is associated with expert identity, field identification, and disease period, and can be shared and called in different terminals and new tasks. The script structure also reserves an interface for linkage with the dynamic abnormal sample mining and adaptive reinforcement learning module.

[0009] As a further aspect of the system of this invention, during the model training phase, the gradient ensemble classification module first performs Pearson correlation coefficient analysis on the multi-resolution vegetation index and gray-level co-occurrence matrix texture features, in conjunction with the key areas marked in the expert teaching path and the atypical lesion samples collected by the dynamic anomaly sample mining and adaptive reinforcement learning module. Collinear redundant features with an absolute value greater than 0.8 are removed. Then, the importance score of each feature is calculated using the random forest algorithm, and only features with an importance greater than 1% are retained as the candidate set. Finally, regularization screening is performed using the minimum absolute shrinkage and selection operator to determine the set of non-zero weight features to be input into the gradient boosting tree model. The classification model parameters are dynamically adaptive according to the expert path adjustment and anomaly sample supplementation process.

[0010] As a further aspect of the system of this invention, the dynamic abnormal sample mining and adaptive reinforcement learning module continuously monitors the model inference confidence index of each sample during the disease identification and classification process, and compares the classification results with historical expert labels. If it is found that the confidence is lower than the set threshold, the discrimination result does not match the previous label, or atypical or easily misjudged samples are detected in the spatial distribution area of ​​new lesions, the sample is automatically marked as a difficult abnormal sample in real time and pushed to experts for review. After expert confirmation, it is synchronously archived into the incremental training set, triggering the model adaptive retraining mechanism to update the model weights in a timely manner.

[0011] As a further embodiment of the system of the present invention, after receiving the CBB level label of red kidney beans output by the gradient integrated classification module, the visualization and control scheduling module overlays the disease level spatial grid onto the geometrically corrected orthophoto mosaic. Based on the severity of lesions, expert teaching paths, and the distribution of dynamic abnormal samples, it automatically generates a variable application prescription map with spatial location information and classification thresholds, and outputs shape files and interchangeable tags for agricultural machinery operations that conform to the international variable spraying equipment control protocol. It accurately controls the type, dosage, and operation path of pesticides according to the distribution of diseases. At the same time, it displays the evolution trend of lesion hot zones, expert trajectory playback, and historical sampling information of difficult areas on the terminal in real time. All prescriptions and scheduling data are synchronously archived to the cloud.

[0012] A method for monitoring diseases in legumes, using the aforementioned legume disease monitoring system, includes the following steps:

[0013] Step 1, CNC remote sensing data acquisition: Drive the UAV flight path in a pre-programmed CNC manner according to the boundary of the work area. The push-broom hyperspectral camera and inertial navigation unit run synchronously to collect hyperspectral cube and attitude data in the 400–1000nm spectral band in a loop. Spatiotemporal registration is completed through Kalman filtering to generate a spectral-pose metadata sequence with timestamps.

[0014] Step 2, Multi-resolution feature scanning teaching: On the data after radiometric correction and geometric registration, soil / shade is automatically segmented with a preset threshold, and then resampled cyclically at four resolutions of 0.03m, 0.05m, 0.10m, and 0.15m. Under the guidance of experts, the vegetation index and gray-scale co-occurrence matrix feature extraction windows are dragged sequentially in a sliding window teaching method. The window position, band channel, and operation sequence are recorded in real time for each operation as a replayable script.

[0015] Step 3, Teaching script recording and storage: The dragging path, threshold adjustment and model fine-tuning operations of experts on multi-scale feature maps are archived in the form of scripted instruction sequences, which can be called across terminals through JSON / XML format and linked with the event triggering of the abnormal sample module;

[0016] Step 4, Trajectory Replay and Servo Loop Control: When working on new plots, the above teaching script is automatically loaded, and the XY slide rail or window servo positioning device is driven to perform the teaching path scan in sequence according to the instruction sequence, accurately reproducing the expert diagnosis and sampling operation;

[0017] Step 5, Abnormal Sample Triggering and Collaborative Review: Monitor the consistency between classification confidence and label throughout the process. If it is marked as a low confidence, conflict, or spatially new lesion event, an abnormal sampling instruction is automatically generated and sent to multiple expert clients for review. The review results are recorded back into the training library through the playback mechanism.

[0018] Step 6, Feature Selection - Gradient Integration Modeling: Based on the teaching path scanning and the feedback of the verification samples, Pearson coefficient, random forest importance and LASSO regularization are applied in sequence to perform iterative feature selection. Finally, the gradient boosting tree model is used to output CBB grade labels and provide feedback on the confidence level.

[0019] Step 7, Prescription Map Generation and Closed-Loop Scheduling: After the graded labels are rasterized and overlaid onto the orthophoto, a variable application prescription map script is automatically generated. The spraying servo system is driven to perform precise prevention and control according to the pre-recorded trajectory, and the execution log is sent back to the cloud, forming a complete adaptive control loop from remote sensing acquisition, teaching recording, replay execution, anomaly review to prevention and control scheduling.

[0020] As a further embodiment of the method of the present invention, in step 3, the spatial distribution and sample type of the lesion identification path drawn by the expert are automatically statistically analyzed. The path is checked to see if it covers each major plot unit, each growth period area and the labeled diverse sample types in the target area. If there are spatial or sample category omissions, a sampling prompt list is automatically generated and pushed to the expert. The expert selects to add a new sampling point and adds the new node to the path script. All nodes are accompanied by growth period, plot number, feature window parameters and operation time sequence labels.

[0021] It should be noted that in this invention, step 3 automatically statistically analyzes the spatial distribution and sample types of the lesion identification paths drawn by experts. Combined with plot units, reproductive period zones, and historical diversity sample labels, it achieves intelligent detection of coverage blind spots and category omissions, and automatically generates supplementary sampling suggestions, prompting experts to dynamically improve the path script. All newly added nodes accurately record the plot, developmental stage, window parameters, and operation sequence. This process overcomes the limitations of existing technologies that rely solely on subjective sampling based on expert experience, have fixed and singular paths, and lack global statistical and feedback mechanisms. It also overcomes the difficulty of conventional methods in covering large-scale plot heterogeneity and cross-period issues. The problems of uneven data collection and insufficient representativeness of sample types enable the path script to have a high degree of adaptability and diverse data collection capabilities when dynamically adapting to complex scenarios with different spaces, periods, and lesion patterns. This effectively improves the system's ability to collect holographic features of lesions and generalize models across fields, growth stages, and heterogeneous environments. This automatic feedback, supplementation, and holographic recording collaborative mechanism is difficult for conventional technologies to achieve in terms of closed-loop self-optimization at the path and data levels. However, this invention can continuously and dynamically optimize the diagnostic path and sample structure, realize regional intelligent perception and path knowledge accumulation, and demonstrate significant technical advantages and scalability in large-scale intelligent monitoring of soybean diseases.

[0022] As a further embodiment of the method of the present invention, in step 5, for atypical samples with confidence levels below a set threshold, label conflicts, and newly emerging lesion areas, the diversity index of the samples in the feature space is automatically calculated, and new or unclassified samples are automatically marked with high priority. These samples, along with the collection parameters and label information, are synchronously uploaded to the cloud expert database. Through remote expert distributed review and label correction, the reviewed samples are automatically added to the incremental training set and labeled with sample category, weight, and collection batch information.

[0023] It should be noted that in step 5 of this invention, atypical samples with confidence levels below a set threshold, label conflicts, and newly emerging lesion areas are automatically screened and prioritized using the feature space diversity index. These problematic samples, along with their collection parameters and label information, are then simultaneously uploaded to a cloud-based expert database. This enables remote, distributed expert review and label correction. The reviewed samples are automatically archived into the incremental training set with category, weight, and collection batch identifier. This differs from the conventional passive process that relies solely on static model training and offline manual sample review. It significantly improves the timely detection and precise classification of atypical, boundary, and novel variant lesion samples. Furthermore, through real-time data streaming, collaborative label calibration by multiple experts in different locations, and dynamic priority weight allocation, this invention further enhances the capabilities of this invention. This invention achieves multi-terminal parallelism and dynamic closed-loop optimization of data acquisition, label verification, and model training, effectively overcoming the shortcomings of existing technologies, such as missed detection of diverse samples, lagging model adaptation to new variants, and difficulty in global collaborative sharing of expert experience. It enables the invention to continuously self-evolve diagnostic boundaries and expand the disease knowledge system in heterogeneous environments across plots, multiple stages, and multiple lesions, ultimately greatly improving the adaptive capability, diagnostic depth, and generalization robustness of soybean disease monitoring. This distributed intelligent feedback and dynamic data closed-loop mechanism surpasses the limitations of conventional reliance on offline manual interpretation and static sample supplementation, fully releasing the technical advantages of expert group collaboration and big data intelligent screening, and providing an innovative solution for the identification and proactive control of high-confidence soybean diseases over large areas.

[0024] The technical effects of this invention's bean disease monitoring system and method are as follows: A UAV remote sensing acquisition module implements high-spectral multi-resolution spatial distribution and holographic coverage of diverse samples; an image preprocessing and feature extraction module achieves closed-loop archiving of acquisition parameters and spatiotemporal labels throughout the entire process; an expert teaching and path recording module records expert interactive lesion identification paths in real time and generates adaptive scripts; a trajectory replay and servo execution module accurately reuses the teaching scripts to drive diagnosis and sampling cycles; a dynamic abnormal sample mining and adaptive reinforcement learning module automatically prioritizes and archives atypical samples and links multi-terminal remote distributed label calibration; a gradient integrated classification module performs incremental self-training after multi-stage feature regularization screening; and a visualization and control scheduling module intelligently generates variable-based pesticide prescriptions and performs global traceability. This collaboratively constructs an integrated adaptive control loop of teaching-replay-training-scheduling, achieving real-time linkage of multiple stages including lesion identification, sample reinforcement, label normalization, model evolution, and intelligent output of control strategies. This significantly improves the high-precision spatial positioning, hierarchical identification, and closed-loop control capabilities of red kidney bean CBB at all stages (early, middle, and late). Attached Figure Description

[0025] Figure 1 This is a system block diagram of the present invention;

[0026] Figure 2This is a schematic diagram and a multi-resolution schematic diagram of the CBB infection stage of the present invention;

[0027] Figure 3 This is a graph showing the average light reflectance of red kidney beans infected with CBB at different stages as a function of wavelength.

[0028] Figure 4 This is a statistical analysis chart of various vegetation indices at different stages of CBB infection in red kidney beans according to the present invention;

[0029] Figure 5 This is a statistical analysis diagram of the texture characteristics of red kidney beans infected with CBB at different stages according to the present invention;

[0030] Figure 6 This is a pixel-based statistical chart showing the severity of CBB infection in red kidney beans at different stages, as presented in this invention.

[0031] Figure 7 This is a flowchart of the method proposed in this invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Example 1. As... Figure 1As shown, the present invention proposes a monitoring system for diseases of mixed beans, including a UAV remote sensing acquisition module, an image preprocessing and feature extraction module, an expert teaching and path recording module, a trajectory replay and servo execution module, a gradient ensemble classification module, a dynamic anomaly sample mining and adaptive reinforcement learning module, and a visualization and control scheduling module. The UAV remote sensing acquisition module digitally drives the acquisition of hyperspectral and attitude data of the red kidney bean canopy along a pre-programmed flight path and generates flight commands. The image preprocessing and feature extraction module completes radiometric correction and geometric registration, and uses a sliding window to scan vegetation indices and gray-level co-occurrence matrices to extract sensitive features of lesions and locate their coordinates. The expert teaching and path recording module uses a human-machine collaborative approach to drag and identify lesions on the feature map. The system identifies the trajectory and records the window position, threshold, and model parameters in real time to create a replay script. The trajectory replay and servo execution module calls the servo driver to perform servo positioning and scanning of the window or slide rail according to the script, in order to reproduce the diagnostic process in the new field. The gradient ensemble classification module takes multi-scale features and taught trajectory data as input, outputs the severity of CBB disease in red kidney beans and provides feedback on the confidence level through a rule-based ensemble strategy. The dynamic anomaly sample mining and adaptive reinforcement learning module dynamically captures atypical samples based on confidence level and label conflict. After expert review, these samples are included in the incremental training set to update the membership function and fuzzy rules. The visualization and control scheduling module renders the results as disease hotspots and calls the agricultural machinery servo system to spray pesticides according to the pre-recorded spray trajectory and transmits the logs back.

[0034] It should be noted that the UAV remote sensing acquisition module is equipped with a pushbroom hyperspectral camera with a spectral range of 400 to 1000 nanometers, a band spacing of 2.1 nanometers, and a spatial resolution adjustable between 0.015 meters and 0.2 meters. It also integrates an inertial measurement unit and a global navigation satellite system. Kalman filtering is used to achieve attitude-positioning data fusion and generate spectral-pose metadata with timestamps. Under normal operating conditions, the UAV remote sensing acquisition module automatically adjusts the acquisition step and flight speed of the pushbroom hyperspectral camera according to the spatial resolution requirements set by the expert teaching and path recording modules. This enables continuous imaging of the path of lesions marked by experts and the corresponding feature extraction areas. When the dynamic anomaly sample mining and adaptive reinforcement learning module detects difficult areas or atypical lesion targets, it triggers a high-density acquisition mode, reducing the acquisition step to 0.015 meters, increasing the number of spectral samples per pixel to all available channels, and reducing the flight speed to 0.5 meters per second. All acquisition parameters and corresponding lesion trajectories are written into the metadata along with the timestamps and pose information.

[0035] It should be noted that after receiving the original hyperspectral data, the image preprocessing and feature extraction module first automatically segments and removes soil and shadow pixels based on a normalized vegetation index threshold of 0.62. Then, the background-removed image is resampled to four spatial resolutions of 0.03m, 0.05m, 0.1m, and 0.15m using nearest neighbor interpolation. Based on expert annotations and path information for the early, middle, and late stages of CBB disease, the module automatically selects the optimal resolution consistent with the teaching trajectory for feature extraction. When working with the dynamic anomaly sample mining and adaptive reinforcement learning modules to push out difficult areas, the image preprocessing and feature extraction module automatically switches to the minimum resolution and maximum sliding window based on the specific morphology and boundary features of the identified atypical lesions, and simultaneously extracts multi-scale vegetation indices and gray-level co-occurrence matrix texture features. All feature extraction processes are stored in association with the expert teaching trajectory and dynamic sample positioning data.

[0036] like Figure 2 As shown, in this embodiment, the technical effects of the above-mentioned technical solution are illustrated through two experimental data: In Experiment 1, a drone (UAV) was used to capture hyperspectral images of red kidney bean plants on days 5, 10, 17, 23, 28, and 37 after CBB inoculation (e.g., Figure 2As shown in the upper row of pictures), the disease index (DI) of CBB infection was investigated simultaneously. Under the guidance of professional plant protection personnel, the overall occurrence of CBB infection was divided into four grades: healthy (DI ≤ 5%), mild (5 < DI ≤ 20%), moderate (20 < DI ≤ 40%), and severe (DI > 40%). The early, middle, and late infection stages corresponded to the development of CBB after inoculation. The six sets of data collected were divided into three infection stages for analysis, namely the early stage (5 days later, 10 days later), the middle stage (17 days later, 23 days later), and the late stage (28 days later, 37 days later). In the early infection stage, non - connected dark green water - soaked spots appeared on the leaves. In the middle infection stage, these water - soaked spots expanded into irregular brown lesions with a lemon - yellow border. In the late infection stage, the brown lesions gradually merged, causing the entire leaf to die or deform. Five survey points within six inoculated plots and two healthy areas were selected, with each area being 1 square meter (1 meter × 1 meter). There were a total of eight plots and 40 survey points. Experiment 1 collected a total of 240 sets of data, including 180 sets of data from CBB - infected red kidney bean plots and 60 sets of data from healthy plots. Experiment 2 was sampled three times on July 18, 2023, August 14, 2023, and July 31, 2022, corresponding to the early, middle, and late stages of CBB infection respectively. Experiment 2 was sown on May 15, 2022, and on June 4, 2023. The sowing time in 2022 was earlier, and continuous rainfall in mid - July accelerated the spread of the disease, resulting in CBB showing typical late - stage disease symptoms on July 31. For each experiment, 30 survey points were randomly selected to measure the DI value of CBB, and high - spectral images were obtained synchronously using drones. According to the method proposed in the present invention, as Figure 1 shown, the monitoring system for miscellaneous bean diseases and its core method proposed in the present invention make full use of the high - density remote sensing ability of the drone - mounted push - broom hyperspectral camera under the conditions of the full wavelength range from 400 to 1000 nm, 2.1 - nm interval, and adjustable resolution from 0.015 to 0.2 m (as Figure 2The image below shows that the system achieved continuous monitoring of the evolution of red kidney bean canopy lesions across multiple time and all stages. Data from Experiments 1 and 2 indicate that the system, through dynamic adaptive adjustment of flight path, acquisition step, and spectral channels, combined with expert-led instruction paths and intelligent dynamic abnormal sample push, achieved holographic hyperspectral imaging coverage of various lesion morphologies, including early-stage (dark green water-soaked spots), mid-stage (irregular brown lesions with yellow halos), and late-stage (large-area patch fusion and leaf death) CBB lesion morphology. The image preprocessing and feature extraction module automatically removed background using a normalized vegetation index threshold and supported multi-resolution synchronous sampling. Combined with expert paths, priority switching to high-resolution acquisition for key windows and difficult areas, it ensured accurate capture and identification of small, blurred-boundary, and highly heterogeneous lesion signals. The system supports real-time scripted trajectory reuse and batch sample acquisition. Combined with gradient ensemble model classification and expert review, it achieved high-confidence disease grading and discrimination under different stages, spatial plots, and growth years. Based on the comprehensive verification of 240 plot samples and multi-temporal and multi-location control data, this invention has achieved intelligent adaptive and fully automatic closed-loop optimization across resolution and time and space in key aspects such as early weak signal detection of lesions, extraction of complex spatial distribution features, and quantitative grading of disease status. This significantly improves the accuracy, timeliness, and scalability of field dynamic monitoring of soybean diseases.

[0037] like Figure 3 As shown, Figure 3 As shown, the method for monitoring diseases of mixed beans proposed in this invention achieves significant differentiation between healthy and diseased plants on the hyperspectral reflectance curves at different infection stages of common bacterial blight (CBB) in red kidney beans. Specifically, regardless of the early, middle, or late infection stages, healthy red kidney bean leaves (green curve) and CBB-infected leaves (orange curve) show obvious differences in average spectral reflectance in the 700nm to 900nm red-edge and near-infrared bands. Moreover, this contrast becomes more prominent as the infection progresses, and the reflectance of diseased leaves continuously decreases throughout the visible to near-infrared region. The above phenomena fully demonstrate that, through multi-temporal, staged hyperspectral remote sensing and dynamic feature extraction, this invention can accurately separate diseased plants from healthy plants and provide early warning when CBB disease in red kidney beans is still in its early, hidden stage. This improves the model's spectral typing sensitivity and dynamic monitoring capability for the entire CBB disease progression process, and solves the technical problems of insufficient resolution at different disease stages in the early, middle, and late stages and difficulty in quantifying and identifying heterogeneous lesions in conventional methods. It has excellent full-cycle, full-process, and full-resolution disease diagnosis and grading effects.

[0038] It should be noted that the expert teaching and path recording module allows plant protection experts to draw recognition paths on the pre-processed feature map using interactive methods such as dragging, selecting, or multi-point drawing, according to the actual distribution of lesions. It records the pixel coordinates of each trajectory node, the size of the feature window used, the band channel, the expert interpretation timestamp, the operation sequence, and the servo position feedback in real time. The feature window size used ranges from 3×3 to 11×11. The above paths and parameters are encapsulated in a standardized structured data format to generate a reusable JSON / XML script. The reusable script is associated with the expert identity, field identifier, and disease period, and can be shared and called in different terminals and new tasks. The script structure also reserves an interface for linkage with the dynamic anomaly sample mining and adaptive reinforcement learning modules.

[0039] This invention, through an expert teaching and path recording module, digitizes and standardizes the entire diagnostic path of plant protection experts on feature maps, including key elements such as pixel coordinates, window parameters, band channels, operation sequence, and interpretation time. This generates a reusable structured script, achieving a high degree of restoration of expert knowledge paths and rapid migration and invocation across terminals and task scenarios. It also supports seamless linkage with dynamic anomaly sample mining and adaptive reinforcement learning modules, facilitating subsequent sample supplementation, dynamic path optimization, and continuous improvement of feature acquisition schemes. This breaks through the problems of traditional manual annotation relying on expert subjective experience, non-repeatable paths, and difficulty in standardizing diagnostic results, significantly improving the system's intelligent collaboration, diagnostic consistency, and automation, generalization, and sustainable evolution capabilities in large-scale operation environments.

[0040] It should be noted that during the model training phase, the gradient ensemble classification module first performs Pearson correlation coefficient analysis on the texture features of multi-resolution vegetation index and gray-level co-occurrence matrix, combined with key areas marked in the expert teaching path and atypical lesion samples collected by the dynamic anomaly sample mining and adaptive reinforcement learning module. Collinear redundant features with an absolute value greater than 0.8 are removed. Then, the importance score of each feature is calculated using the random forest algorithm, and only features with an importance greater than 1% are retained as the candidate set. Finally, regularization screening is performed using the minimum absolute shrinkage and selection operator to determine the set of non-zero weight features to be input into the gradient boosting tree model. The classification model parameters are dynamically adaptive according to the expert path adjustment and anomaly sample supplementation process.

[0041] like Figure 4 and Figure 5 As shown, the typical technical effects of the bean disease monitoring system and method proposed in this invention in the extraction and screening of multidimensional information such as "vegetation index features" and "gray-level symbiotic matrix texture features" are demonstrated. Figure 4(Vegetation Index Section) The upper part (a) shows the correlation coefficient distribution between different vegetation indices (such as AR1, NDVI, Clrededge, NormRRE, etc.) and CBB disease of red kidney bean in different infection stages of early, middle and late stages. The lower part (b) shows the weight distribution of each vegetation index in the final discriminant model after multi-stage feature screening and integrated modeling. Figure 5 (Texture Features) The upper part (a) also shows the correlation distribution between the texture features (such as DIS, CON, ENT, MEA, VAR, etc.) of each gray-level co-occurrence matrix under different principal components (PC1, PC2, PC3) and the disease progression stage. The lower part (b) gives the importance weight of each texture feature in the model.

[0042] The above results demonstrate that this invention can automatically extract and evaluate the response performance of hundreds of spectral indices and texture features by fusing hyperspectral-multi-resolution remote sensing with expert teaching for different disease progression stages. Based on Pearson correlation, multi-stage screening, and gradient ensemble regularization, it achieves fine-grained screening and weight allocation of highly correlated and sensitive features (such as vegetation indices like NormRRE, PRI, and Clrededge, and texture features like DIS, CON, and MEA), effectively filtering redundant features and improving the model's generalization ability and lesion grading accuracy. This automated feature screening and dynamic weight allocation capability significantly differs from traditional discrimination methods that rely on manual experience or static thresholds for single features, demonstrating the innovation and technical effectiveness of this invention in large-scale multimodal intelligent disease identification and adaptive model optimization.

[0043] It should be noted that the dynamic anomaly sample mining and adaptive reinforcement learning module continuously monitors the model inference confidence index of each sample during the disease identification and classification process, and compares the classification results with historical expert labels. If the confidence is found to be lower than the set threshold, the judgment result does not match the previous label, or atypical or easily misjudged samples are detected in the spatial distribution area of ​​new lesions, the module automatically marks such samples as difficult anomalies in real time and pushes them to experts for review. After expert confirmation, the samples are synchronously archived into the incremental training set, triggering the model adaptive retraining mechanism to update the model weights in a timely manner.

[0044] like Figure 6As shown, this invention demonstrates the automatic identification and classification mapping results of CBB disease severity spatial distribution across multiple areas (Area a, b, c, d) in a red kidney bean field, based on UAV remote sensing and hyperspectral analysis. Different colors correspond to healthy, mild, moderate, and severe lesions, achieving precise spatial positioning and dynamic monitoring of disease severity evolution across a large field. Combined with the dynamic anomaly sample mining and adaptive reinforcement learning modules of this invention, the system can automatically identify problematic samples with low inference confidence, conflicting labels, or newly emerging lesions during disease identification and classification. These samples are promptly submitted for expert review, and the reviewed data is incorporated into the incremental training set to drive dynamic optimization of model weights. Therefore, when facing disease areas with complex spatial distribution, strong lesion heterogeneity, or sudden new variations, the system can continuously improve identification accuracy and robustness, ensuring continuous, dynamic, and refined quantification of disease distribution across different infection stages and multiple field plots. Compared with traditional static interpretation and single-marking schemes, this invention realizes real-time closed-loop self-learning and hierarchical intelligent evolution of disease monitoring, which greatly enhances the intelligence level and adaptability of spatial diagnosis, zoned prevention and control and large-scale application of diseases in soybean fields.

[0045] It should be noted that after receiving the CBB level label of red kidney beans output by the gradient integrated classification module, the visualization and control scheduling module overlays the disease level spatial raster onto the geometrically corrected orthophoto mosaic. Based on the severity of lesions, expert teaching paths, and the distribution of dynamic abnormal samples, it automatically generates a variable application prescription map with spatial location information and classification thresholds, and outputs shape files and interchangeable tags for agricultural machinery operations that conform to the international variable spraying equipment control protocol. It can accurately control the type, dosage, and operation path of pesticides according to the distribution of diseases. At the same time, it displays the evolution trend of lesion hot zones, expert trajectory playback, and historical sampling information of difficult areas on the terminal in real time. All prescriptions and scheduling data are synchronously archived to the cloud.

[0046] This invention utilizes a visualization and control scheduling module to overlay the CBB disease level labels of red kidney beans output by the gradient integrated classification module with orthophoto mosaics, achieving a spatial and visual representation of lesion grading. Based on lesion severity, expert teaching paths, and the distribution of dynamic abnormal samples, it automatically generates variable application prescription maps with precise spatial location information and grading thresholds. It can output standardized operation shape files and interchangeable tags for agricultural machinery, enabling seamless integration with variable spraying equipment and precise control of pesticide type, dosage, and application path. This module also supports real-time dynamic display of lesion hotspot evolution, expert trajectory reproduction, and sampling history of difficult areas on the terminal. All prescription and scheduling data are uniformly archived in the cloud, ensuring traceability throughout the entire operation process and closed-loop management of control decisions. This zonal visualization and automatic prescription generation based on multi-source heterogeneous data fusion overcomes the bottlenecks of traditional disease control, such as extensive zoning, wasteful pesticide use, difficulty in tracing the source, and slow response. It greatly improves the accuracy, intelligence, and scalability of control operations, providing a solid technical foundation for green disease control and digital agriculture in large-scale mixed bean fields.

[0047] Example 2. The difference between Example 2 and Example 1 is that this example introduces a method for monitoring diseases in legumes.

[0048] like Figure 7 As shown, the present invention proposes a method for monitoring diseases of mixed beans, which utilizes the aforementioned mixed bean disease monitoring system and includes the following steps:

[0049] Step 1, CNC remote sensing data acquisition: Drive the UAV flight path in a pre-programmed CNC manner according to the boundary of the work area. The push-broom hyperspectral camera and inertial navigation unit run synchronously to collect hyperspectral cube and attitude data in the 400–1000nm spectral band in a loop. Spatiotemporal registration is completed through Kalman filtering to generate a spectral-pose metadata sequence with timestamps.

[0050] Step 2, Multi-resolution feature scanning teaching: On the data after radiometric correction and geometric registration, soil / shade is automatically segmented with a preset threshold, and then resampled cyclically at four resolutions of 0.03m, 0.05m, 0.10m, and 0.15m. Under the guidance of experts, the vegetation index and gray-scale co-occurrence matrix feature extraction windows are dragged sequentially in a sliding window teaching method. The window position, band channel, and operation sequence are recorded in real time for each operation as a replayable script.

[0051] Step 3, Teaching script recording and storage: The dragging path, threshold adjustment and model fine-tuning operations of experts on multi-scale feature maps are archived in the form of scripted instruction sequences, which can be called across terminals through JSON / XML format and linked with the event triggering of the abnormal sample module;

[0052] Step 4, Trajectory Replay and Servo Loop Control: When working on new plots, the above teaching script is automatically loaded, and the XY slide rail or window servo positioning device is driven to perform the teaching path scan in sequence according to the instruction sequence, accurately reproducing the expert diagnosis and sampling operation;

[0053] Step 5, Abnormal Sample Triggering and Collaborative Review: Monitor the consistency between classification confidence and label throughout the process. If it is marked as a low confidence, conflict, or spatially new lesion event, an abnormal sampling instruction is automatically generated and sent to multiple expert clients for review. The review results are recorded back into the training library through the playback mechanism.

[0054] Step 6, Feature Selection - Gradient Integration Modeling: Based on the teaching path scanning and the feedback of the verification samples, Pearson coefficient, random forest importance and LASSO regularization are applied in sequence to perform iterative feature selection. Finally, the gradient boosting tree model is used to output CBB grade labels and provide feedback on the confidence level.

[0055] Step 7, Prescription Map Generation and Closed-Loop Scheduling: After the graded labels are rasterized and overlaid onto the orthophoto, a variable application prescription map script is automatically generated. The spraying servo system is driven to perform precise prevention and control according to the pre-recorded trajectory, and the execution log is sent back to the cloud, forming a complete adaptive control loop from remote sensing acquisition, teaching recording, replay execution, anomaly review to prevention and control scheduling.

[0056] It should be noted that in step 3, the spatial distribution and sample types of the lesion identification path drawn by the expert are automatically statistically analyzed. The path is checked to see if it covers each major plot unit, each growth period area and the labeled diverse sample types within the target area. If there are spatial or sample category omissions, a sampling prompt list is automatically generated and pushed to the expert. The expert selects to add a new sampling point and adds the new node to the path script. All nodes are accompanied by growth period, plot number, feature window parameters and operation time sequence labels.

[0057] By automatically statistically analyzing the spatial distribution and sample types of lesion identification paths drawn by experts within the target area, the system can detect in real time whether the paths have covered all key plot units, reproductive period areas, and diverse sample types. When spatial or categorical omissions are detected, the system automatically generates and pushes sampling suggestions to guide experts in supplementing sampling, achieving dynamic improvement and scientific expansion of the path script. All newly added sampling nodes accurately record the reproductive period, plot number, feature window parameters, and operation sequence, forming a structured diagnostic knowledge data chain throughout the entire process. This mechanism significantly improves the comprehensiveness and representativeness of lesion sample collection, ensuring that features from different spaces, periods, and lesion types are effectively collected and archived. It overcomes the shortcomings of traditional manual sampling, such as easy omissions, non-reusable paths, and incomplete diagnostic information, providing a solid data foundation for subsequent model training, feature extraction, and cross-regional generalization. This greatly enhances the system's capabilities in intelligent collection, path standardization, and consistency of diagnostic results.

[0058] It should be noted that in step 5, for atypical samples with confidence levels below the set threshold, label conflicts, and newly emerging lesion areas, the diversity index of the samples in the feature space is automatically calculated. New or unclassified samples are automatically marked with high priority. These samples, along with the collection parameters and label information, are uploaded to the cloud expert database. Remote experts conduct distributed review and label correction. The reviewed samples are then automatically added to the incremental training set and labeled with sample category, weight, and collection batch information.

[0059] By automatically calculating the feature space diversity index for atypical samples with low confidence, label conflicts, and newly emerging lesion areas, the system can intelligently identify and prioritize the labeling of novel or unclassified difficult samples. Its collected parameters and label information are synchronized in real time to a cloud-based expert database, supporting remote review and label correction by experts in multiple locations. Confirmed samples are automatically archived into the incremental training set, with added category, weight, and batch information. This process achieves fully automated and standardized management of the entire process from difficult sample discovery, expert review, sample archiving to model retraining. It greatly improves the timely capture and high-quality label acquisition of novel, boundary, and variant lesion types, ensuring the integrity and representativeness of the model training data structure. This provides a continuous data-driven foundation for dynamic model optimization and generalization in complex scenarios, overcoming the limitations of traditional manual screening, offline verification, and static samples that are difficult to adapt to changes in new diseases. It significantly improves the intelligent evolution capability and accurate diagnostic level of the disease monitoring system.

[0060] Based on the content described in Examples 1 and 2, this invention achieves real-time linkage across multiple stages, including lesion identification, sample reinforcement, label normalization, model evolution, and intelligent output of control strategies. This is achieved through holographic coverage of spatial distribution and diverse samples, adaptive script reuse of expert paths, automatic priority archiving and remote distributed label calibration of dynamic atypical samples, closed-loop archiving of data from the entire process of collection parameters and spatiotemporal labels, multi-stage feature regularization screening and incremental self-training of models, automatic generation of variable application prescriptions, and global traceability. Unlike traditional schemes that rely on static collection, single-point labeling, offline training, and fixed-value control, this invention effectively improves the intelligent perception, classification, sample structure self-increase, and disease distribution prediction capabilities of common bacterial blight in mixed beans across fields, growth stages, complex spatial heterogeneity, and novel variant environments. It realizes end-to-end closed-loop self-evolution of data collection, path knowledge, diagnostic models, and application decisions, significantly enhancing the overall system performance of early-stage CBB micro-spot identification in red kidney beans, remote expert collaboration, and forward-looking variable control. This provides innovative and highly adaptable technical support for intelligent monitoring and green precision control of diseases in large-scale mixed bean crops.

[0061] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0062] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A monitoring system for diseases in mixed beans, characterized in that, The system includes a UAV remote sensing acquisition module, an image preprocessing and feature extraction module, an expert teaching and path recording module, a trajectory replay and servo execution module, a gradient ensemble classification module, a dynamic anomaly sample mining and adaptive reinforcement learning module, and a visualization and control scheduling module. The UAV remote sensing acquisition module, which digitally drives the UAV along a pre-programmed flight path, acquires hyperspectral and attitude data of the red kidney bean canopy and generates flight commands. The image preprocessing and feature extraction module performs radiometric correction and geometric registration, and uses a sliding window to scan vegetation indices and gray-level co-occurrence matrices to extract sensitive features of lesions and locate their coordinates. The expert teaching and path recording module uses a human-machine collaborative approach to drag the lesion recognition trajectory on the feature map and records the window in real time. The position, threshold, and model parameters are adjusted to form a replay script. The trajectory replay and servo execution module calls the servo driver to perform servo positioning and scanning of the window or slide rail according to the script, so as to reproduce the diagnosis process in the new field. The gradient ensemble classification module takes multi-scale features and teaching trajectory data as input, outputs the severity of CBB disease in red kidney beans and feeds back the confidence level through a rule-based ensemble strategy. The dynamic abnormal sample mining and adaptive reinforcement learning module dynamically captures atypical samples based on confidence level and label conflict. After expert review, the samples are included in the incremental training set to update the membership function and fuzzy rules. The visualization and control scheduling module renders the results as disease hotspots and calls the agricultural machinery servo system to spray pesticides according to the pre-recorded spraying trajectory and sends back the log. After receiving the hyperspectral raw data, the image preprocessing and feature extraction module first automatically segments and removes soil and shadow pixels based on a normalized vegetation index threshold of 0.

62. Then, the background-removed image is resampled to four spatial resolutions of 0.03m, 0.05m, 0.1m, and 0.15m using nearest neighbor interpolation. Based on expert annotations and path information for the early, middle, and late stages of CBB disease, the module automatically selects the optimal resolution consistent with the teaching trajectory for feature extraction. When working with the dynamic anomaly sample mining and adaptive reinforcement learning modules to push out difficult areas, the image preprocessing and feature extraction module automatically switches to the minimum resolution and maximum sliding window according to the specific morphology and boundary features of the identified atypical lesions, and simultaneously extracts multi-scale vegetation indices and gray-level co-occurrence matrix texture features. All feature extraction processes are stored in association with the expert teaching trajectory and dynamic sample positioning data. During the model training phase, the gradient ensemble classification module first performs Pearson correlation coefficient analysis on the texture features of multi-resolution vegetation index and gray-level co-occurrence matrix, combined with key areas marked in the expert teaching path and atypical lesion samples collected by the dynamic anomaly sample mining and adaptive reinforcement learning module. Collinear redundant features with an absolute value greater than 0.8 are removed. Then, the importance score of each feature is calculated using the random forest algorithm, and only features with an importance greater than 1% are retained as the candidate set. Finally, regularization is performed using minimum absolute shrinkage and selection operators to determine the set of non-zero weight features to be input into the gradient boosting tree model. The parameters of the classification module are dynamically adaptive according to the expert path adjustment and the anomaly sample supplementation process. During the disease identification and classification process, the dynamic anomaly sample mining and adaptive reinforcement learning module continuously monitors the model inference confidence index of each sample and compares the classification results with historical expert labels. If the confidence is found to be lower than the set threshold, the discrimination result does not match the previous label, or atypical or easily misjudged samples are detected in the spatial distribution area of ​​new lesions, the sample is automatically marked as a difficult anomaly sample in real time and pushed to experts for review. After expert confirmation, it is synchronously archived into the incremental training set, triggering the model adaptive retraining mechanism to update the model weights in a timely manner.

2. The bean disease monitoring system according to claim 1, characterized in that, The UAV remote sensing acquisition module is equipped with a pushbroom hyperspectral camera with a spectral range of 400 to 1000 nanometers, a band spacing of 2.1 nanometers, and an adjustable spatial resolution between 0.015 meters and 0.2 meters. It also integrates an inertial measurement unit and a global navigation satellite system. Kalman filtering is used to achieve attitude-positioning data fusion and generate spectral-pose metadata with timestamps. Under normal operating conditions, the UAV remote sensing acquisition module automatically adjusts the acquisition step and flight speed of the pushbroom hyperspectral camera according to the spatial resolution requirements set by the expert teaching and path recording modules. This enables continuous imaging of the path of lesions marked by experts and the corresponding feature extraction areas. When the dynamic anomaly sample mining and adaptive reinforcement learning module detects difficult areas or atypical lesion targets, it triggers a high-density acquisition mode, reducing the acquisition step to 0.015 meters, increasing the number of spectral samples per pixel to all available channels, and reducing the flight speed to 0.5 meters per second. All acquisition parameters and corresponding lesion trajectories are written into the metadata along with the timestamps and pose information.

3. The bean disease monitoring system according to claim 1, characterized in that, The expert teaching and path recording module allows plant protection experts to draw identification paths on the pre-processed feature map using interactive methods such as dragging, selecting, or multi-point drawing, according to the actual lesion distribution. It records the pixel coordinates, feature window size, band channels, expert interpretation timestamp, operation sequence, and servo position feedback of each trajectory node in real time. The feature window size ranges from 3×3 to 11×11. The above paths and parameters are encapsulated in a standardized structured data format to generate a reusable JSON / XML script. The reusable script is associated with expert identity, field identification, and disease stage, and can be shared and called in different terminals and new tasks. The script structure also reserves an interface for linkage with the dynamic abnormal sample mining and adaptive reinforcement learning modules.

4. The bean disease monitoring system according to claim 1, characterized in that, After receiving the CBB level labels for red kidney beans from the gradient integrated classification module, the visualization and control scheduling module overlays the disease level spatial raster onto the geometrically corrected orthophoto mosaic. Based on the severity of lesions, expert teaching paths, and the distribution of dynamic abnormal samples, it automatically generates a variable application prescription map with spatial location information and classification thresholds. It also outputs shape files and interchangeable tags for agricultural machinery operations that conform to the international variable spraying equipment control protocol, enabling precise control of pesticide type, dosage, and operation path according to disease distribution. At the same time, it displays the evolution trend of lesion hot zones, expert trajectory playback, and historical sampling information of difficult areas on the terminal in real time. All prescriptions and scheduling data are synchronously archived to the cloud.

5. A method for monitoring diseases of mixed beans, using the mixed bean disease monitoring system according to any one of claims 1-4, characterized in that, Includes the following steps: Step 1, CNC remote sensing data acquisition: Drive the UAV flight path in a pre-programmed CNC manner according to the boundary of the work area. The push-broom hyperspectral camera and inertial navigation unit run synchronously to collect hyperspectral cube and attitude data in the 400–1000nm spectral band in a loop. Spatiotemporal registration is completed through Kalman filtering to generate a spectral-pose metadata sequence with timestamps. Step 2, Multi-resolution Feature Scanning Teaching: On the data after radiometric correction and geometric registration, soil / shade is automatically segmented with a preset threshold, and then resampled cyclically at four resolutions: 0.03m, 0.05m, 0.10m, and 0.15m. Under the guidance of experts, the vegetation index and gray-scale co-occurrence matrix feature extraction windows are dragged sequentially in a sliding window teaching method. The window position, band channel, and operation sequence are recorded in real time for each operation as a replayable script. Step 3, Teaching script recording and storage: The dragging path, threshold adjustment and model fine-tuning operations of experts on multi-scale feature maps are archived in the form of scripted instruction sequences, which can be called across terminals through JSON / XML format and linked with the event triggering of the abnormal sample module; Step 4, Trajectory Replay and Servo Loop Control: When working on new plots, the above teaching script is automatically loaded, and the XY slide rail or window servo positioning device is driven to perform the teaching path scan in sequence according to the instruction sequence, accurately reproducing the expert diagnosis and sampling operation; Step 5, Abnormal Sample Triggering and Collaborative Review: Monitor the consistency between classification confidence and label throughout the process. If it is marked as a low confidence, conflict, or spatially new lesion event, an abnormal sampling instruction is automatically generated and sent to multiple expert clients for review. The review results are recorded back into the training library through the playback mechanism. Step 6, Feature Selection - Gradient Integration Modeling: Based on the teaching path scanning and the feedback of the verification samples, Pearson coefficient, random forest importance and LASSO regularization are applied in sequence to perform iterative feature selection. Finally, the gradient boosting tree model is used to output CBB grade labels and provide feedback on the confidence level. Step 7, Prescription Map Generation and Closed-Loop Scheduling: After the graded labels are rasterized and overlaid onto the orthophoto, a variable application prescription map script is automatically generated. The spraying servo system is driven to perform precise prevention and control according to the pre-recorded trajectory, and the execution log is sent back to the cloud, forming a complete adaptive control loop from remote sensing acquisition, teaching recording, replay execution, anomaly review to prevention and control scheduling.

6. The method for monitoring diseases of legumes according to claim 5, characterized in that, In step 3, the spatial distribution and sample types of the lesion identification path drawn by the expert are automatically statistically analyzed. The path is checked to see if it covers each major plot unit, each growth period area and the labeled diverse sample types within the target area. If there are spatial or sample category omissions, a sampling prompt list is automatically generated and pushed to the expert. The expert selects to add a new sampling point and adds the new node to the path script. All nodes are accompanied by growth period, plot number, feature window parameters and operation time sequence labels.

7. A method for monitoring diseases of legumes according to claim 5, characterized in that, In step 5, for atypical samples with confidence levels below a set threshold, label conflicts, and newly emerging lesion areas, the diversity index of the samples in the feature space is automatically calculated. New or unclassified samples are automatically marked with high priority. These samples, along with the collection parameters and label information, are uploaded to the cloud expert database. Remote experts conduct distributed review and label correction. The reviewed samples are then automatically added to the incremental training set and labeled with sample category, weight, and collection batch information.

Images