A machine vision-based system and method for controlling camphor tree chlorosis.

Abstract

This invention relates to the field of disease control technology, specifically disclosing a system and method for controlling camphor tree chlorosis based on machine vision. This system addresses problems such as multi-spectral imaging misalignment, inconsistent correction, low segmentation accuracy, and lack of a closed-loop control mechanism. The system includes an imaging component, an edge computing module, a pseudo-label adaptive refinement module, a segmentation module, a fusion processing module, a multi-classification model, a decision map module, and an intelligent spraying device. The method includes synchronous imaging, initial labeling, pseudo-label refinement, lesion segmentation, feature fusion, state diagnosis, path planning, precise spraying, and closed-loop control. This technical solution achieves multi-spectral synchronization and consistency through fiber optic hardware triggering and closed-loop correction. It generates weak labels in real time and refines them through convolutional iteration of leaf vein skeleton maps. Multi-scale U-Net combined with leaf edge curvature CRF pixel-level segmentation is used. REP / GPRI / WBI spectral features are dynamically selected, and LightGBM is alternately trained to achieve four-level diagnosis, path planning closed-loop control, and spraying updates.

Description

Technical Field

[0001] This invention relates to the field of disease control technology, and in particular to a system and method for controlling camphor tree chlorosis based on machine vision. Background Technology

[0002] In plantations in southern my country, iron deficiency chlorosis is prevalent among important greening tree species such as camphor trees due to the alkalinity of forest soils and long-term short-rotation and high-intensity extensive management practices. Previous studies have used VNIR hyperspectral imaging in the 350-2500 nm band combined with the OPLS-DA discriminant model or the rhizosphere growth-promoting bacterium Rahnella aquati ilis JZ-GX1 to secrete various organic acids and synthesize isohydroxamic acid-type deferoxamine to lower soil pH, release VOCs to induce the expression of genes such as AtAHA2 / AtFRO2 / AtIRT1, and regulate the rhizosphere microbial community to restore chlorophyll content, SPAD value, and antioxidant enzyme activity, significantly promoting the regreening of chlorotic seedlings and enhancing root growth. However, this study did not establish a classification and labeling system based on airborne and ground-based multispectral / hyperspectral images to automatically extract and distinguish the effects of four treatments: healthy control group, untreated chlorotic group, FeSO4-treated recovery group, and JZ-GX1-treated recovery group. It also did not extract and differentiate the effects of different treatments. Image-level key features highly correlated with PAD values, such as red edge displacement, ExG / NDVI green index, SWIR water absorption characteristics, Gabor / LBP texture energy, leaf vein structure intensity, and leaf margin roughness, are used to achieve early sensitive detection, multi-level typing, and precise localization of camphor chlorosis. However, it is difficult to link biological / chemical control strategies with UAV inspections, GIS positioning, and intelligent fertilization / bacterial application equipment to form a closed-loop control process of identification—segmentation—multi-level discrimination—decision—precise execution—effect feedback—model update. Therefore, there is an urgent need to propose a camphor chlorosis control system and method based on machine vision. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide a system and method for the prevention and control of camphor tree chlorosis based on machine vision. By simultaneously performing multi-spectral imaging, generating weak labels with YOLOv5+CAM through pruning quantization and refining leaf vein superpixel images through convolution, performing precise segmentation of leaf margins using multi-scale U-Net+CRF, screening features of REP / GPRI / WBI, and performing multi-classification diagnosis and GPS closed-loop spraying, the system achieves early, efficient, sensitive detection and precise closed-loop control of camphor tree chlorosis.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a machine vision-based system for the prevention and control of camphor tree chlorosis, comprising an imaging component deployed on a drone and a ground platform, including visible light and red-edge hyperspectral and SWIR narrow-band multispectral imaging modules, an edge computing module, a pseudo-label adaptive refinement module, a segmentation module, a fusion processing module, a multi-classification model, a decision map module, and an intelligent spraying device; the imaging component is used to simultaneously acquire 350-1000nm visible light and red-edge hyperspectral and SWIR moisture absorption data and automatically perform white and dark board correction; the edge computing module is used to run YOLOv5 to locate the ROI of camphor tree leaves and generate initial Grad-CAM pseudo-labels through a classification network; the pseudo-label adaptive refinement module performs pixel-level loss cycling... The feedback identification module identifies areas of concentrated error and generates enhanced pseudo-labels; the segmentation module performs pixel-level segmentation of yellowing lesions on leaf ROIs based on the enhanced pseudo-labels; the fusion processing module fuses the lesion mask with the multispectral image and extracts key features such as red edge displacement, ExG, NDVI, SWIR water absorption, Gabor and LBP texture, leaf vein grayscale, and leaf edge roughness; the multi-classification model uses LightGBM to classify the extracted features into four levels of labels: healthy, untreated yellowing, FeSO4 treated, and JZ-GX1 inoculant applied; the decision map module generates precise fertilization / inoculant application paths based on classification labels and GIS positioning mapping; after the intelligent spraying equipment performs fertilization / inoculant application, it feeds back SPAD and chlorophyll data to the edge computing module and the fusion processing module.

[0005] As a further embodiment of the present invention, the imaging component includes a hyperspectral camera with visible light and red edge in the range of 350-1000nm, a SWIR narrow-band camera with center wavelengths of 980nm, 1450nm and 1940nm respectively, and also includes a ring LED fill light array, a red edge laser irradiation source and built-in reference white plate and dark plate. Each camera is coaxially aligned in the same optical window through a preset optical path. The transmission and reflection paths are synchronized and phase-locked through an optical fiber trigger line and a hardware clock, and the synchronous acquisition of the reference white plate, dark plate and images of each band is completed within the same exposure cycle.

[0006] As a further aspect of the present invention, in the imaging assembly, the preset optical path through which each camera passes includes: a first dichroic beam splitter, a multi-wedge prism array, a second dichroic beam splitter, a third dichroic beam splitter, and a narrow-band filter are sequentially arranged behind the optical window; the first dichroic beam splitter transmits visible light and the red-edge band of 350-1000nm and reflects long-wavelength light >1000nm; the reflected beam is shaped by the multi-wedge prism array and enters the second dichroic beam splitter, which transmits the 1400nm±10nm band and reflects the remaining >1350nm band; the reflected beam from the second dichroic beam splitter enters the third dichroic beam splitter, which reflects the 1940nm±10nm band and transmits the 1000nm-1500nm residual band; all transmitted bands pass through corresponding narrow-band filters and sequentially enter SWIR cameras with center wavelengths of 980nm, 1450nm, and 1940nm.

[0007] As a further aspect of the present invention, the edge computing module is a heterogeneous SoC platform integrating NPU and GPU, loading a YOLOv5 detection and classification model and Grad-CAM branch that has been pruned and quantized by INT8. The platform uses hardware DMA to zero-copy map the acquired camphor tree image frames to a unified on-chip memory. The NPU and GPU perform YOLOv5 leaf ROI localization and classification intermediate feature extraction in parallel, and simultaneously trigger the Grad-CAM branch to directly read the classification head weights and intermediate feature maps to generate an initial pseudo-label heatmap. Finally, the on-chip bilinear interpolation accelerator upsamples the heatmap to the original ROI pixel size.

[0008] As a further aspect of the present invention, the pseudo-label adaptive refinement module includes a leaf vein skeleton extraction unit, a weighted superpixel adjacency unit constructed based on the reflection gradient of the 690nm-730nm red-edge band, and a two-layer graph convolutional propagation unit; the leaf vein skeleton extraction unit performs morphological refinement on the leaf ROI to obtain the topology of the primary and secondary leaf veins; the weighted superpixel adjacency unit constructs an adjacency matrix with superpixels as nodes and reflection gradient and spatial proximity as weights; the graph convolutional propagation unit performs two rounds of propagation and update of the initial weak label confidence along the leaf vein topology to generate enhanced pseudo-labels.

[0009] As a further aspect of the present invention, the segmentation module uses U-Net network units with void ratios of 2, 4, and 8 respectively for three scales of typical camphor tree yellowing patches with diameters of 100px, 150px, and 200px to extract multi-scale lesion features, and performs pixel-level optimization of the segmentation results using a conditional random field with leaf edge curvature similarity as the edge weight.

[0010] As a further aspect of the present invention, in the fusion processing module, the three indices with correlation coefficients ≥0.75—Red Edge Displacement (REP), Ratio Green Peak Index (GPRI), and Water Stress Index (WBI)—are dynamically screened based on the pre-calculated correlation of camphor tree yellowing disease, and the screening results are used as the input feature set of the LightGBM model.

[0011] A machine vision-based method for controlling camphor tree chlorosis, employing the aforementioned machine vision-based camphor tree chlorosis control system, includes:

[0012] Step 1, Image Acquisition: Using airborne and ground-based multispectral imaging units, images of 400nm-1000nm visible light, 690nm-730nm red edge, and 1450nm-1750nm shortwave infrared are acquired simultaneously at 30fps, and white and dark plate corrections are completed.

[0013] Step 2, Initial Labeling: Perform YOLOv5 detection in the edge computing module to locate the leaf ROI with IoU≥0.5 and call Grad-CAM to generate initial weak labels;

[0014] Step 3, Pseudo-label refinement: The pseudo-label refinement module constructs a weighted adjacency graph based on the reflection gradient of the red-edge band and the superpixel topology of the leaf vein skeleton. It then propagates the weak label confidence along the leaf vein topology through a two-layer graph convolutional network and outputs enhanced pseudo-labels after two rounds of iteration.

[0015] Step 4, Lesion segmentation: The segmentation module extracts lesion features using a three-scale U-Net model with void ratios of 2, 4, and 8, corresponding to 100px, 150px, and 200px respectively, and optimizes the segmentation results at the pixel level by combining the leaf edge curvature conditional random field.

[0016] Step 5, Feature Extraction: The fusion processing module fuses the enhanced pseudo-label mask with the multispectral image, and extracts red edge shift, REP, GPRI, WBI spectral indices filtered by SPAD correlation ≥0.75, as well as Gabor / LBP texture energy, leaf vein grayscale and leaf edge roughness features.

[0017] Step 6, Status Diagnosis: The diagnosis module uses an alternating training strategy to optimize the LightGBM model on the training and validation sets, achieving healthy, untreated yellowing, chemical recovery after FeSO4 spraying (0.5g / strain), and JZ-GX1 bacterial solution (10... 6 CFU / mL biological recovery classification (quadruple level);

[0018] Step 7, Path Planning: The path planning module generates a refined spraying path with a node spacing of 0.5m based on the classification tags and GPS positioning;

[0019] Step 8, Precision spraying: The spraying subsystem applies fertilizer or microbial agent along the generation path, with a spraying amount error of ≤±5%;

[0020] Step 9, Closed-loop feedback: After spraying, the SPAD value and chlorophyll content are collected in real time and transmitted back to the edge computing module and fusion processing module to complete the closed-loop update.

[0021] As a further aspect of the present invention, in step 3, the primary vein superpixel nodes with a red-edge reflection gradient ≥ 0.15 at 690nm to 730nm are assigned an initial weight of 0.80 to 1.00, and the secondary vein nodes with a reflection gradient < 0.15 are assigned an initial weight of 0.50 to 0.70; the residual connection coefficient in the two-layer graph convolution propagation is taken as 0.10 to 0.30, and the normalization coefficient is taken as 0.80 to 1.00.

[0022] As a further aspect of the present invention, in step 4, the number of intermediate convolutional layers in the three scales of the U-Net network is expanded from the original 3 layers to 5 layers; in the post-processing stage of the conditional random field, the edge weight is 0.70 when the difference in leaf edge curvature is ≤0.10, and otherwise the edge weight is 0.30.

[0023] As a further aspect of the present invention, in step 5, the fusion processing module sets the initial gating threshold for REP, GPRI, and WBI to 0.75 respectively. When the deviation between the real-time SPAD value and the target SPAD value is ≥2%, the threshold is adjusted linearly within the range [0.70, 0.80].

[0024] The technical advantages of the machine vision-based system and method for controlling camphor tree chlorosis in this invention are as follows:

[0025] This invention achieves early sensitive detection, multi-level typing and precise positioning, and intelligent closed-loop control of iron deficiency chlorosis in camphor trees by integrating airborne and ground-based multispectral / hyperspectral imaging, initial weak labels using YOLOv5+Grad-CAM, two-layer graph convolutional iterative refinement of pseudo-labels based on 690-730nm red edge reflection gradient and leaf vein skeleton superpixel topology, multi-scale U-Net joint conditional random field segmentation, dynamic spectral index and texture feature fusion extraction, alternating training of LightGBM four-level classification, GPS fine spraying path planning and online adaptive closed-loop feedback of spraying amount and SPAD real-time correction. This effectively improves the monitoring response speed, typing accuracy and treatment precision. Attached Figure Description

[0026] Figure 1 This is a system diagram of the present invention;

[0027] Figure 2 This is a network architecture diagram of the edge computing module of the present invention;

[0028] Figure 3 This is a comparison image of the initial pseudo-label edge and the refined pseudo-label edge of the present invention;

[0029] Figure 4 Comparison of single-scale threshold segmentation of the lesion area of ​​camphor tree leaves with the segmentation optimized by the U-Net+CRF of this invention;

[0030] Figure 5 This is a bar chart comparing the model performance indicators before and after dynamic feature filtering in this invention.

[0031] Figure 6 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

[0034] like Figure 1 As shown, this invention proposes a machine vision-based system for controlling camphor tree chlorosis, comprising an imaging component deployed on a drone and a ground platform, featuring visible light and red-edge hyperspectral and SWIR narrow-band multispectral imaging, an edge computing module, a pseudo-label adaptive refinement module, a segmentation module, a fusion processing module, a multi-classification model, a decision map module, and an intelligent spraying device. The imaging component simultaneously acquires 350–1000 nm visible light and red-edge hyperspectral and SWIR moisture absorption data and automatically performs white and dark board correction. The edge computing module runs YOLOv5 to locate the ROI of camphor tree leaves and generates initial Grad-CAM pseudo-labels through a classification network. The pseudo-label adaptive refinement module uses pixel-level loss feedback to identify errors. The system focuses on key regions and generates enhanced pseudo-labels; the segmentation module performs pixel-level segmentation of yellowing lesions on leaf ROIs based on the enhanced pseudo-labels; the fusion processing module fuses the lesion mask with the multispectral image and extracts key features such as red edge displacement, ExG, NDVI, SWIR water absorption, Gabor and LBP texture, leaf vein grayscale, and leaf margin roughness; the multi-classification model uses LightGBM to classify the extracted features into four levels of labels: healthy, untreated yellowing, FeSO4 treated, and JZ-GX1 inoculant applied; the decision map module generates precise fertilization / inoculant application paths based on classification labels and GIS positioning mapping; after the intelligent spraying equipment performs fertilization / inoculant application, it feeds back SPAD and chlorophyll data to the edge computing module and the fusion processing module.

[0035] In the field of camphor tree chlorosis control, the light intensity varies greatly, and the reflectivity of camphor tree leaves fluctuates significantly with time and location. A single sensor cannot simultaneously capture visible light, red edge, and SWIR information. A multi-spectral imaging approach combining airborne and ground-based sensors enables millisecond-level synchronous acquisition and correction of 50–1000nm visible light / red edge and 980nm–1940nm SWIR multi-spectral bands. This eliminates spatiotemporal misalignment and light source interference at the data source, significantly improving the sensitivity to early chlorosis signals in young leaves. For machine vision-based camphor tree chlorosis identification, traditional offline annotation methods are time-consuming and unsuitable for real-time annotation requirements during large-area inspections. However, on-chip YOLOv5+Grad-CAM generates initial pseudo-labels in real time, and edge NPU / GPU collaboration achieves instant ROI localization and classification heatmap output with IOU ≥ 0.5, reducing data annotation latency to <50ms and ensuring the continuous online operation capability of the entire system. Especially for camphor chlorosis, the initial weak Grad-CAM labels are prone to misclassification or omission at leaf veins and leaf margins, affecting subsequent segmentation accuracy. A combination of leaf vein skeleton superpixels and a two-layer GCN iterative refinement of pseudo-labels is employed. Based on the 690–730nm red-edge gradient, primary and secondary leaf vein skeletons are extracted and weighted superpixel maps are constructed. These superpixel maps are then propagated along the veins through two-layer convolution and iteratively updated, ultimately improving the IoU of pseudo-labels and reducing the misclassification rate in detailed areas. Camphor chlorosis lesion diameters vary greatly (100–200px), making it difficult for single-scale models to handle both large and small lesions. The serrated leaf margins make it difficult for traditional CRF to remove false lesions. A three-scale U-Net with parallel porosity of 2 / 4 / 8, combined with CRF post-processing based on leaf margin curvature, improves the Dice coefficient of the lesions and reduces the false positive rate. When using machine vision to identify camphor tree chlorosis, simple spectral or texture features are insufficiently correlated with SPAD (spider vein disease), resulting in weak generalization ability of the diagnostic model. Integrating multi-dimensional features such as REP, ExG, NDVI, SWIR indices, Gabor / LBP texture, leaf vein grayscale, and leaf margin roughness significantly improved the average correlation coefficient with SPAD to 0.82 (from 0.65), greatly enhancing the model's ability to map physiological states. Furthermore, in the identification and control of camphor tree chlorosis, data imbalance and the scarcity of recovery group (chemical / biological) samples easily lead to overfitting or underfitting of the classifier. Alternating parameter tuning strategies between training and validation sets improved the average F1 score for all four classes, with a particularly improved recall rate for the recovery group. In the prevention and control of camphor tree chlorosis, conventional manual or simple GPS path spraying has low precision, poor fertilizer and bacteria utilization rate and no feedback mechanism. Based on grid-based path planning with a node spacing of 0.5m, the spraying coverage rate is ≥95%, the spraying volume error is controlled within ±5%, and combined with SPAD real-time feedback, the application amount is dynamically corrected, the amount of chemical fertilizer used is reduced and the revegetation speed is improved.

[0036] It should be noted that the imaging components include a hyperspectral camera with visible light and red edge in the range of 350–1000 nm, a SWIR narrow-band camera with center wavelengths of 980 nm, 1450 nm and 1940 nm, a ring LED fill light array, a red edge laser irradiation source and built-in reference white plate and dark plate. Each camera is coaxially aligned in the same optical window through a preset optical path. The transmission and reflection paths are synchronized and phase-locked through fiber optic trigger lines and hardware clocks, and the synchronous acquisition of reference white plate, dark plate and images of each band is completed within the same exposure cycle.

[0037] Existing technologies have significant limitations in terms of synchronization accuracy, correction consistency, spectral coverage, illumination uniformity, and image alignment for multispectral imaging. This invention, however, optimizes the spatiotemporal consistency and spectral integrity of multispectral data through features such as hardware-triggered synchronization, closed-loop hardware correction, visible / red edge SWIR broadband coverage, a ring-shaped LED illumination array, and coaxial optical alignment. Table 1 compares the technical features of existing technologies with those of this invention.

[0038] Table 1 Comparison of the Imaging Components in the Existing Technology and the Invention

[0039]

[0040] To address the technical challenges of early-stage, sensitive detection of iron deficiency chlorosis in plantation camphor trees, including temporal misalignment in multi-spectral imaging, inconsistent calibration, limited spectral coverage, uneven illumination, and low image alignment accuracy, this invention employs a collaborative deployment of 350–1000 nm visible / red-edge hyperspectral and 980 / 1450 / 1940 nm SWIR multispectral cameras on a UAV and ground platform. Timing synchronization errors are reduced to ≤1ms through fiber optic hardware triggering and hardware clock synchronization; hardware closed-loop white-panel correction improves calibration accuracy to ±2%; spectral coverage is expanded to 350–1000 nm and three SWIR bands, increasing spectral information by more than three times; a ring-shaped LED illumination array controls brightness variation to ≤5%; and coaxial optical path alignment reduces alignment errors to ≤1px. This comprehensive and high-precision approach ensures the spatiotemporal consistency and spectral integrity of multi-spectral data, providing a solid foundation for the accurate operation of advanced algorithms such as precise real-time weak label generation based on intelligent NPU / GPU, iterative refinement of graph convolutional pseudo-labels, and multi-scale segmentation.

[0041] It should be noted that in the imaging assembly, the preset optical paths for each camera include: a first dichroic beam splitter, a multi-wedge prism array, a second dichroic beam splitter, a third dichroic beam splitter, and a narrow-band filter are sequentially arranged behind the optical window; the first dichroic beam splitter transmits visible light (350–1000 nm) and the red-edge band and reflects long-wavelength light (>1000 nm); the reflected beam is shaped by the multi-wedge prism array and enters the second dichroic beam splitter, which transmits light (1400 nm ± 10 nm) and reflects the remaining light (>1350 nm); the reflected beam from the second dichroic beam splitter enters the third dichroic beam splitter, which reflects light (1940 nm ± 10 nm) and transmits light (1000 nm–1500 nm); all transmitted wavelengths pass through corresponding narrow-band filters and sequentially enter SWIR cameras with center wavelengths of 980 nm, 1450 nm, and 1940 nm.

[0042] Based on actual test data of each beam splitter and filter, Table 2 lists the transmission efficiency, cutoff band suppression, and insertion loss indices of each component in the optical path of this invention at the target wavelength:

[0043] Table 2 Performance Test Records of Optical Path Components

[0044] Optical components Target band (nm) Transmission efficiency (%) Cutoff band suppression (dB) Insertion loss (dB) First dichroic beam splitter 350–1000 94.5 28 0.6 Multi-wedge prism array Reflection > 1000 92 30 0.7 Second dichroic mirror 1400±10 93.5 32 0.5 Third dichroic beam splitter 1940±10 92.8 31 0.6 Narrow band filter 980 / 1450 / 1940 96.2 35 0.4

[0045] The data in the table clearly shows that:

[0046] The transmission efficiency is greater than 92% (average ≈93.8%), which is about 3% higher than the typical 90% transmission efficiency of traditional single beam splitters, ensuring that more target wavelength light energy reaches each SWIR camera.

[0047] The cutoff band suppression reaches 28–35dB, which is at least 12% higher than the traditional ≈25dB, effectively isolating non-target wavelength light and reducing crosstalk;

[0048] Insertion loss is controlled within 0.4–0.7 dB, which is about 30–60% lower than the traditional insertion loss below 1.0 dB, further improving the system's optical path efficiency.

[0049] The aforementioned performance improvements directly address the three major bottlenecks of low spectroscopic efficiency, insufficient out-of-band suppression, and high insertion loss: high transmission ensures the intensity of red-edge and SWIR signals, strong suppression reduces stray light interference, and low insertion loss improves the overall signal-to-noise ratio, thus providing fundamental support for early sensitive detection and accurate feature extraction of camphor tree chlorosis.

[0050] It should be noted that, as Figure 2As shown, the edge computing module is a heterogeneous SoC platform integrating NPU and GPU, loading a YOLOv5 detection and classification model and Grad-CAM branch that has been pruned and quantized with INT8. The platform uses hardware DMA to zero-copy map the acquired camphor tree image frames to a unified on-chip memory. The NPU and GPU perform YOLOv5 leaf ROI localization and classification intermediate feature extraction in parallel, and simultaneously trigger the Grad-CAM branch to directly read the classification head weights and intermediate feature maps to generate an initial pseudo-label heatmap. Finally, the on-chip bilinear interpolation accelerator upsamples the heatmap to the original ROI pixel size.

[0051] like Figure 2 As shown, the edge computing module integrates an NPU and GPU on a heterogeneous SoC platform. Through network pruning and INT8 quantization, it compresses the original FP32 YOLOv5 model into a lightweight version, achieving a 3x speedup for single detection inference compared to conventional methods, while reducing power consumption from 20W to 5W. Hardware DMA zero-copy is used to directly map the acquired camphor tree image frames to unified on-chip memory, reducing memory transfer latency by approximately 50%. The NPU and GPU execute YOLOv5 leaf ROI localization and intermediate feature extraction in parallel, simultaneously triggering the Grad-CAM branch to directly read the classification head weights and feature maps, generating an initial pseudo-label heatmap, reducing latency to ≤5ms, more than 10 times faster than traditional offline Grad-CAM. Finally, on-chip bilinear interpolation... The accelerator completes upsampling from the heatmap to the original ROI pixel size, reducing the upsampling latency from 20ms to ≤2ms. The overall end-to-end inference latency is reduced from the traditional ≈200ms to ≤30ms, and the real-time frame rate is increased to 30fps. This solution improves the overall battery life by more than 20%, and ensures the accuracy of subsequent pseudo-label iteration and multi-level classification through real-time annotation and precise segmentation, thereby accelerating the early warning and precise treatment of camphor tree chlorosis. At the same time, the overall module power consumption is reduced by 75%, and the single inspection flight time is increased from 30 minutes to more than 36 minutes, further meeting the needs of large-scale inspection of plantations.

[0052] It should be noted that the pseudo-label adaptive refinement module includes a leaf vein skeleton extraction unit, a weighted superpixel adjacency unit constructed based on the reflection gradient of the 690nm-730nm red-edge band, and a two-layer graph convolutional propagation unit. The leaf vein skeleton extraction unit performs morphological refinement on the leaf ROI to obtain the topology of the primary and secondary leaf veins. The weighted superpixel adjacency unit constructs an adjacency matrix with superpixels as nodes and reflection gradient and spatial proximity as weights. The graph convolutional propagation unit performs two rounds of propagation and update of the initial weak label confidence along the leaf vein topology to generate enhanced pseudo-labels.

[0053] like Figure 3As shown, the left image displays the pseudo-label edges (red) obtained by direct thresholding based on the initial Grad-CAM weak labels. Obvious breaks and missed areas are visible at the leaf veins (dark green stripes). The right image shows the edges obtained by iteratively refining the pseudo-labels after adding leaf vein skeleton superpixel mapping and propagating along the topological two-round graph convolution. It can be observed that the continuity of the lesion edges is significantly enhanced in the leaf vein intersection areas, effectively filling in the original breaks. Compared with conventional direct thresholding or single-frame CRF-based segmentation methods, this technique overcomes the problem of pseudo-label breaks caused by leaf vein occlusion in the machine vision recognition and control of camphor chlorosis in the image processing. It utilizes the 690–730nm red-edge reflection gradient to construct a weighted superpixel map and combines it with leaf vein skeleton topological information for graph convolution propagation. This achieves continuous and complete delineation of chlorosis lesions on complex leaf vein networks, thus providing more reliable input for subsequent accurate segmentation and multi-level diagnosis.

[0054] It should be noted that the segmentation module uses U-Net network units with void ratios of 2, 4, and 8 for three scales of typical camphor tree yellowing patches with diameters of 100px, 150px, and 200px, respectively, to extract multi-scale lesion features. The segmentation results are then optimized at the pixel level using a conditional random field with leaf edge curvature similarity as the edge weight.

[0055] like Figure 4 As shown, compared with the conventional single-scale threshold segmentation method, the above-mentioned technical solution adopted in the segmentation module of this invention, on the left is "before processing: single-scale threshold segmentation", where it can be seen that the red boundary is discontinuous in places with different patch sizes and frequently misses segmentation at the intersection of leaf edge and leaf vein; on the right is "after processing: multi-scale U-Net+CRF optimization", where the red boundary is continuously extracted for three typical patches of 100px, 150px, and 200px, and remains intact in the serrated areas of leaf edge and leaf vein occlusion areas, with significantly reduced noise. This comparison fully demonstrates that this technical means, through multi-scale porosity U-Net network and pixel-level optimization based on leaf edge curvature conditional random field, effectively improves the segmentation coherence and accuracy, and overcomes the limitations of conventional single-scale methods in segmenting yellow spots of different sizes and leaf vein interference.

[0056] It should be noted that in the fusion processing module, based on the pre-calculated correlation of camphor yellowing disease, three indices with correlation coefficients ≥0.75—Red Edge Displacement (REP), Ratio Green Peak Index (GPRI), and Water Stress Index (WBI)—are dynamically screened, and the screening results are used as the input feature set of the LightGBM model.

[0057] like Figure 5As shown in the bar chart, the comparison illustrates the differences in key performance indicators before and after the dynamic spectral index screening: classification accuracy increased from 84.2% to 92.5%, an improvement of 8.3 percentage points; F1 score increased from 0.81 to 0.89, an improvement of 0.08; and single-frame inference latency was reduced from approximately 15ms to approximately 6ms, a reduction of 60%.

[0058] These data trends intuitively demonstrate that this technology significantly optimizes the model's feature input by dynamically screening three spectral indices—REP, GPRI, and WBI—that are highly correlated with camphor tree yellowing disease. This allows LightGBM to greatly improve operational efficiency while ensuring high-precision diagnosis, meeting the needs of real-time online prevention and control by drones.

[0059] Example 2

[0060] The difference between Embodiment 2 and Embodiment 1 is that this embodiment introduces a method for preventing and controlling camphor tree chlorosis based on machine vision.

[0061] like Figure 6 As shown, the present invention proposes a machine vision-based method for controlling camphor tree chlorosis, which applies a machine vision-based camphor tree chlorosis control system as shown in Example 1, including:

[0062] Step 1, Image Acquisition: Using airborne and ground-based multispectral imaging units, images of 400nm-1000nm visible light, 690nm-730nm red edge, and 1450nm-1750nm shortwave infrared are acquired simultaneously at 30fps, and white and dark plate corrections are completed.

[0063] Step 2, Initial Labeling: Perform YOLOv5 detection in the edge computing module to locate the leaf ROI with IoU≥0.5 and call Grad-CAM to generate initial weak labels;

[0064] Step 3, Pseudo-label refinement: The pseudo-label refinement module constructs a weighted adjacency graph based on the reflection gradient of the red-edge band and the superpixel topology of the leaf vein skeleton. It then propagates the weak label confidence along the leaf vein topology through a two-layer graph convolutional network and outputs enhanced pseudo-labels after two rounds of iteration.

[0065] Step 4, Lesion segmentation: The segmentation module extracts lesion features using a three-scale U-Net model with void ratios of 2, 4, and 8, corresponding to 100px, 150px, and 200px respectively, and optimizes the segmentation results at the pixel level by combining the leaf edge curvature conditional random field.

[0066] Step 5, Feature Extraction: The fusion processing module fuses the enhanced pseudo-label mask with the multispectral image, and extracts red edge shift, REP, GPRI, WBI spectral indices filtered by SPAD correlation ≥0.75, as well as Gabor / LBP texture energy, leaf vein grayscale and leaf edge roughness features.

[0067] Step 6, Status Diagnosis: The diagnosis module uses an alternating training strategy to optimize the LightGBM model on the training and validation sets, achieving healthy, untreated yellowing, chemical recovery after FeSO4 spraying (0.5g / strain), and JZ-GX1 bacterial solution (10... 6 CFU / mL biological recovery classification (quadruple level);

[0068] Step 7, Path Planning: The path planning module generates a refined spraying path with a node spacing of 0.5m based on the classification tags and GPS positioning;

[0069] Step 8, Precision spraying: The spraying subsystem applies fertilizer or microbial agent along the generation path, with a spraying amount error of ≤±5%;

[0070] Step 9, Closed-loop feedback: After spraying, the SPAD value and chlorophyll content are collected in real time and transmitted back to the edge computing module and fusion processing module to complete the closed-loop update.

[0071] This method employs multi-spectral (400–1000nm visible / red edge + 1450–1750nm SWIR) simultaneous imaging of camphor tree leaves at 30fps, real-time weak label generation based on pruned and quantized YOLOv5+Grad-CAM, iterative refinement using double-layer graph convolution combining 690–730nm red edge gradient and primary and secondary vein skeleton superpixel topology, pixel-level lesion segmentation using multi-scale U-Net with porosity of 2 / 4 / 8 combined with leaf margin curvature conditional random field, dynamic screening and fusion of multi-dimensional features such as REP / GPRI / WBI spectral indices and Gabor / LBP texture, vein grayscale, and leaf margin roughness, and alternating optimization of LightGBM for healthy-untreated-chemically restored conditions on training and validation sets. - The system features four-level classification for biological restoration, GPS-based 0.5m gridded spraying path planning and ±5% spraying amount control, and SPAD-chlorophyll real-time closed-loop feedback. It has independently achieved sub-pixel-level sensitive capture of early red edge displacement and water stress signals in camphor iron deficiency chlorosis. The classification accuracy has been improved from 84% to over 92%, the recall rate has been increased by 12%, the real-time diagnosis frame rate has reached 30fps, the amount of chemical / biological intervention has been reduced by ≥20%, and the flight time for a single inspection has been extended to over 36 minutes. Furthermore, the system has significantly improved the control efficiency, diagnostic accuracy, and cost-effectiveness of large-area plantation inspections through online adaptive updates.

[0072] In the method proposed in this invention, in step 3, the primary vein superpixel nodes with a red-edge reflection gradient ≥ 0.15 at 690nm to 730nm are assigned an initial weight of 0.80 to 1.00, and the secondary vein nodes with a reflection gradient < 0.15 are assigned an initial weight of 0.50 to 0.70; the residual connection coefficient in the two-layer graph convolution propagation is taken as 0.10 to 0.30, and the normalization coefficient is taken as 0.80 to 1.00.

[0073] In step 3, the initial weights of the primary vein superpixels are set to 0.80–1.00 and the secondary vein weights to 0.50–0.70 for the 690nm–730nm red-edge reflection gradient. Residual connectivity coefficients of 0.10–0.30 and normalization coefficients of 0.80–1.00 are used in the two-layer graph convolutional network, respectively. This allows the false label confidence to propagate efficiently along the real vein topology, significantly enhancing the connectivity and integrity of lesions in the vein intersection and occlusion areas, suppressing false boundaries, and thus greatly improving the overlap between the false label and the actual lesion shape. This provides a solid foundation for subsequent pixel-level fine segmentation and accurate diagnosis.

[0074] In the method proposed in this invention, in step 4, the number of intermediate convolutional layers in the three scales of the U-Net network is expanded from the original 3 layers to 5 layers; in the post-processing stage of the conditional random field, the edge weight is 0.70 when the difference in leaf edge curvature is ≤0.10, and otherwise the edge weight is 0.30.

[0075] By expanding the intermediate bottleneck module of the U-Net network at three scales from three layers to five layers, the network can extract the subtle textures and edge features of camphor tree yellowing lesions more deeply at different resolutions. In the post-processing stage of the conditional random field, a high edge weight (0.70) is assigned when the leaf edge curvature difference is ≤0.10 and a low edge weight (0.30) is assigned when the curvature difference is >0.10. This maintains the continuous smoothness of the segmentation boundary in areas with small curvature changes and enhances the sharpness of the segmentation edge in high curvature areas. Ultimately, the Dice coefficient of lesion segmentation is increased by about 12%, and the false positive rate is reduced by about 15%, which significantly improves the segmentation accuracy and robustness of yellowing lesions of different shapes and sizes.

[0076] In the method proposed in this invention, in step 5, the fusion processing module sets the initial gating threshold of REP, GPRI, and WBI to 0.75 respectively. When the deviation between the real-time SPAD value and the target SPAD value is ≥2%, the threshold is adjusted linearly within the range [0.70, 0.80].

[0077] To address the issue that fluctuations in chlorophyll content caused by iron deficiency chlorosis in camphor trees affect the reliable extraction of red edge displacement, green index, and water stress signals, an initial gating threshold of 0.75 was set for REP, GPRI, and WBI in the fusion processing module. When the deviation between the real-time SPAD value and the preset target SPAD value was ≥2%, this threshold was linearly adjusted within the range of 0.70 to 0.80 to eliminate abnormal spectral interference caused by leaf vein shading, lesion differences, and changes in light and humidity. This ensured that the final spectral characteristics were highly correlated with chlorophyll content, significantly improving the accuracy and stability of the LightGBM model for grading and diagnosing camphor tree chlorosis.

[0078] In summary, the technical solutions of the present invention described in Embodiments 1 and 2 achieve ±2% correction accuracy and ≤1ms timing synchronization through the following methods in Embodiment 1: airborne and ground-based multispectral / hyperspectral and SWIR collaborative imaging, ring LED supplementary lighting array, and hardware closed-loop white background correction; heterogeneous NPU / GPU zero-copy parallel execution of pruned quantization YOLOv5+Grad-CAM weak label generation and on-chip bilinear interpolation upsampling; pseudo-label adaptive refinement module constructs leaf vein skeleton superpixel weighted map using 690–730nm red-edge reflection gradient and iteratively updates and strengthens pseudo-labels through two-layer graph convolution; multi-scale U-Net with a porosity of 2 / 4 / 8 combined with a leaf edge curvature conditional random field pixel-level segmentation module that takes into account 100px / 150px / 200px lesions and removes noise; and fusion processing module dynamically gates REP / GPRI / WBI spectral indices and Gabor / LBP texture, leaf vein grayscale, and leaf edge roughness features (initial threshold 0.75). When the SPAD deviation is ≥2%, it is linearly adjusted within 0.70–0.80. The diagnostic module alternately trains LightGBM to achieve high-precision classification of four levels: healthy / untreated / FeSO4 chemical recovery / JZ-GX1 biological recovery. The path planning module generates a gridded spraying path with a node spacing of 0.5m based on GPS. The precision spraying subsystem controls the spraying amount error to ≤±5% and provides real-time SPAD / chlorophyll closed-loop feedback. Combined with the first dichroic spectroscope and multi-wedge prism array in Example 2, The high-efficiency optical path constructed with columns and narrowband filters achieves an average transmission efficiency of ≥92%, out-of-band suppression of ≥28dB, and insertion loss of ≤0.7dB. This improves classification accuracy from 84% to ≥92%, F1 score from 0.81 to 0.89, reduces single-frame diagnostic latency to ≤30ms, reduces power consumption by 75%, reduces false positive rate by ≥20%, and extends single-inspection flight time by more than 20%. This significantly enhances the early sensitive detection, multi-level classification, and precise closed-loop control capabilities for camphor yellowing disease.

[0079] 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.

[0080] 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 machine vision-based system for controlling camphor tree chlorosis, characterized in that, The system includes an imaging component capable of simultaneously acquiring visible light, red-edge hyperspectral, and SWIR narrow-band multispectral data; an edge computing module; a pseudo-label adaptive refinement module; a segmentation module; a fusion processing module; a multi-classification model; a decision map module; and an intelligent spraying device. The imaging component is used to simultaneously acquire 350–1000 nm visible light, red-edge hyperspectral, and SWIR moisture absorption data and automatically perform white and dark board correction. The edge computing module is used to run YOLOv5 to locate the ROI of camphor tree leaves and generate initial Grad-CAM pseudo-labels through a classification network. The pseudo-label adaptive refinement module identifies error concentration areas through pixel-level loss cyclic feedback and generates enhanced pseudo-labels. The segmentation module performs pixel-level segmentation of yellowing lesions on leaf ROIs based on enhanced pseudo-labels; the fusion processing module fuses lesion masks with multispectral images and extracts key features such as red edge displacement, ExG, NDVI, SWIR water absorption, Gabor and LBP textures, leaf vein grayscale, and leaf margin roughness; the multi-classification model uses LightGBM to classify the extracted features into four levels of labels: healthy, untreated yellowing, FeSO4 treated, and JZ-GX1 inoculant applied; the decision map module generates precise fertilization / inoculant application paths based on classification labels and GIS positioning mapping; after the intelligent spraying equipment performs fertilization / inoculant application, it feeds back SPAD and chlorophyll data to the edge computing module and the fusion processing module.

2. The system for controlling camphor tree chlorosis based on machine vision according to claim 1, characterized in that, The imaging components include a hyperspectral camera with visible light and red edge in the range of 350–1000 nm, a SWIR narrow-band camera with center wavelengths of 980 nm, 1450 nm and 1940 nm, a ring LED fill light array, a red edge laser irradiation source and built-in reference white plate and dark plate. Each camera is coaxially aligned in the same optical window through a preset optical path. The transmission and reflection paths are synchronized and phase-locked through fiber optic trigger lines and hardware clocks, and the synchronous acquisition of reference white plate, dark plate and images of each band is completed within the same exposure cycle.

3. The system for controlling camphor tree chlorosis based on machine vision according to claim 2, characterized in that, In the imaging assembly, the preset optical paths for each camera include: a first dichroic beam splitter, a multi-wedge prism array, a second dichroic beam splitter, a third dichroic beam splitter, and a narrow-band filter arranged sequentially behind the optical window; the first dichroic beam splitter transmits visible light (350–1000 nm) and the red-edge band and reflects long-wavelength light (>1000 nm); the reflected beam is shaped by the multi-wedge prism array and enters the second dichroic beam splitter, which transmits light (1400 nm ± 10 nm) and reflects the remaining light (>1350 nm); the reflected beam from the second dichroic beam splitter enters the third dichroic beam splitter, which reflects light (1940 nm ± 10 nm); all transmitted wavelengths pass through corresponding narrow-band filters and then sequentially enter SWIR cameras with center wavelengths of 980 nm, 1450 nm, and 1940 nm.

4. The system for controlling camphor tree chlorosis based on machine vision according to claim 3, characterized in that, The edge computing module is a heterogeneous SoC platform integrating NPU and GPU, loading a YOLOv5 detection and classification model and Grad-CAM branch that has been pruned and quantized with INT8. The platform uses hardware DMA to zero-copy map the acquired camphor tree image frames to a unified on-chip memory. The NPU and GPU perform YOLOv5 leaf ROI localization and classification intermediate feature extraction in parallel, and simultaneously trigger the Grad-CAM branch to directly read the classification head weights and intermediate feature maps to generate an initial pseudo-label heatmap. Finally, the on-chip bilinear interpolation accelerator upsamples the heatmap to the original ROI pixel size.

5. The system for controlling camphor tree chlorosis based on machine vision according to claim 1, characterized in that, The pseudo-label adaptive refinement module includes a leaf vein skeleton extraction unit, a weighted superpixel adjacency unit constructed based on the reflection gradient of the 690nm-730nm red-edge band, and a two-layer graph convolutional propagation unit. The leaf vein skeleton extraction unit performs morphological refinement on the leaf ROI to obtain the topology of the primary and secondary leaf veins. The weighted superpixel adjacency unit constructs an adjacency matrix with superpixels as nodes and reflection gradient and spatial proximity as weights. The graph convolutional propagation unit performs two rounds of propagation and update of the initial weak label confidence along the leaf vein topology to generate enhanced pseudo-labels.

6. The system for controlling camphor tree chlorosis based on machine vision according to claim 1, characterized in that, The segmentation module uses U-Net network units with void ratios of 2, 4, and 8 for three scales of typical camphor tree yellowing patches with diameters of 100px, 150px, and 200px, respectively, to extract multi-scale lesion features. The segmentation results are then optimized at the pixel level using a conditional random field with leaf edge curvature similarity as the edge weight.

7. The system for controlling camphor tree chlorosis based on machine vision according to claim 1, characterized in that, In the fusion processing module, based on the pre-calculated correlation of camphor chlorosis, three indices with correlation coefficients ≥0.75—Red Edge Displacement (REP), Ratio Green Peak Index (GPRI), and Water Stress Index (WBI)—are dynamically screened, and the screening results are used as the input feature set of the LightGBM model.

8. A method for controlling camphor tree chlorosis based on machine vision, using the camphor tree chlorosis control system based on machine vision as described in any one of claims 1-7, characterized in that, include: Step 1, Image Acquisition: Using airborne and ground-based multispectral imaging units, images of 400nm-1000nm visible light, 690nm-730nm red edge, and 1450nm-1750nm shortwave infrared are acquired simultaneously at 30fps, and white and dark plate corrections are completed. Step 2, Initial Labeling: Perform YOLOv5 detection in the edge computing module to locate the leaf ROI with IoU≥0.5 and call Grad-CAM to generate initial weak labels; Step 3, Pseudo-label refinement: The pseudo-label refinement module constructs a weighted adjacency graph based on the reflection gradient of the red-edge band and the superpixel topology of the leaf vein skeleton. It then propagates the weak label confidence along the leaf vein topology through a two-layer graph convolutional network and outputs enhanced pseudo-labels after two rounds of iteration. Step 4, Lesion segmentation: The segmentation module extracts lesion features using a three-scale U-Net model with void ratios of 2, 4, and 8, corresponding to 100px, 150px, and 200px respectively, and optimizes the segmentation results at the pixel level by combining the leaf edge curvature conditional random field. Step 5, Feature Extraction: The fusion processing module fuses the enhanced pseudo-label mask with the multispectral image, and extracts red edge shift, REP, GPRI, WBI spectral indices filtered by SPAD correlation ≥0.75, as well as Gabor / LBP texture energy, leaf vein grayscale and leaf edge roughness features. Step 6, Status Diagnosis: The diagnosis module uses an alternating training strategy to optimize the LightGBM model on the training and validation sets, achieving healthy, untreated yellowing, chemical recovery after FeSO4 spraying (0.5g / strain), and JZ-GX1 bacterial solution (10... 6 CFU / mL biological recovery classification (quadruple level); Step 7, Path Planning: The path planning module generates a refined spraying path with a node spacing of 0.5m based on the classification tags and GPS positioning; Step 8, Precision Spraying: The spraying subsystem applies fertilizer or microbial agent along the generation path, with a spraying amount error of ≤±5%; Step 9, Closed-loop feedback: After spraying, the SPAD value and chlorophyll content are collected in real time and transmitted back to the edge computing module and fusion processing module to complete the closed-loop update.

9. A method for controlling camphor tree chlorosis based on machine vision according to claim 8, characterized in that, In step 3, the primary vein superpixel nodes with a red-edge reflection gradient ≥ 0.15 in the 690nm–730nm range are assigned an initial weight of 0.80–1.00, and the secondary vein nodes with a reflection gradient < 0.15 are assigned an initial weight of 0.50–0.

70. The residual connection coefficients in the two-layer graph convolutional propagation are taken as 0.10–0.30, and the normalization coefficients are taken as 0.80–1.

00. In step 4, the number of intermediate convolutional layers in the three scales of the U-Net network is expanded from the original 3 layers to 5 layers. In the conditional random field post-processing stage, the edge weight is 0.70 when the difference in leaf edge curvature is ≤ 0.10, and 0.30 otherwise.

10. A method for controlling camphor tree chlorosis based on machine vision according to claim 8, characterized in that, In step 5, the fusion processing module sets the initial gating threshold of REP, GPRI, and WBI to 0.

75. Based on the fact that the deviation between the real-time SPAD value and the target SPAD value is ≥2%, the threshold is adjusted linearly within the range [0.70, 0.80].

Images