Organizations: Department of Biological Systems Engineering, University of Nebraska–Lincoln, Lincoln, NE, USA · Panhandle Research and Extension Center, University of Nebraska–Lincoln, Scottsbluff, NE, USA · Department of Entomology, University of Nebraska–Lincoln, Lincoln, NE, USA · Department of Biological Systems Engineering, Florida A&M University, Tallahassee, FL, USA · USDA-ARS, National Soil Erosion Research laboratory, West Lafayette, IN, USA
Insect-induced defoliation threatens grapevine health and yield, yet severity assessment relies on subjective manual scouting. We developed DefoliationCAM (D-CAM), a solar-powered edge AI system that estimates canopy defoliation and transmits results via LoRa to a web dashboard. The pipeline uses YOLOX-Nano to localize grapevine cordons in 4K images and a MobileNet-based classifier with lightweight channel attention to classify 320 x 320 pixel patches. The 29,674 field-collected patches were labeled as background or one of four damage levels: no to negligible, mild, moderate, and severe. Merging moderate and severe classes reduced the class imbalance ratio from approximately 40:1 to 8.8:1. Synthetic augmentation increased macro-F1 from 0.768 to 0.838. The classifier outperformed ResNet-50 in macro-F1 with fewer parameters (1.08 M versus 23.52 M). Patch predictions were aggregated into a canopy-level Non-linear Global Defoliation Severity (NGDS) score and validated against an entomologist's ratings of 86 held-out canopy images. Quadratic discriminant analysis identified boundaries of 9% and 29% separating negligible damage, mild damage requiring monitoring, and action-level damage warranting on-site scouting and insecticide treatment consideration. The 29% boundary aligned with the 25%-30% action threshold range for Japanese beetle management in grapevine. Using TensorRT FP16 on an NVIDIA Jetson Orin Nano, the pipeline processed each 4K image in approximately 1.5 s, with a complete operating cycle of 57-60 s. A five-week deployment of two field units demonstrated the feasibility of autonomous solar-powered operation, LoRa communication, and remote defoliation monitoring. D-CAM supports frequent, objective canopy damage assessment for targeted scouting and timely pest-management decisions.
Figures & tables
Figure 1: Three-phase framework for the development and deployment of D-CAM. (1) scheduled acquisition of 4K grape-canopy images using a solar-powered field unit; (2) dataset preparation, class consolidation, minority-class augmentation, model training, and TensorRT FP16 conversion; and (3) on-device canopy localization, patch-level severity assessment, and LoRa transmission of results to a web dashboard.
Figure 2: System architecture and data flow of D-CAM.
Figure 4: Custom analog battery voltage-sensing circuit. A DC Y-splitter taps the TalentCell 12.8 V battery: one branch powers the Jetson Orin Nano, while the other feeds a resistive divider ( R1=10 k Ω , R2=2.2 k Ω ) whose midpoint ( Vout ) is read on the Maduino (ATmega328) analog pin A1. The divider keeps the source impedance ( ∼ 1.8 k Ω ) below the 10 k Ω ATmega328 ADC limit, and the raw reading is converted to battery voltage via Vbat=ADCraw×0.01461 .
Figure 5: Battery voltage-to-state-of-charge (SOC) calibration curve for the 12.8 V, 6000 mAh LiFePO 4 battery. SOC was estimated by linear interpolation between the calibration points, with voltage readings of 12 V or greater assigned 100% SOC.
Code
Meaning
00
All stages completed successfully
01
Image capture failed (camera/GStreamer error)
02
Model inference failed (TensorRT engine error)
03
Battery voltage read failed (ADC/serial error)
04
LoRa data transmission failed
05
Low-battery power-save
Table 1: Error code system embedded in each 30-byte LoRa payload.
Figure 7: Row (A): real no to negligible damage canopy patches. Row (B): AI-generated damaged images with realistic insect-chewing damage (irregular holes, skeletonized patches, organic edges) via prompt-driven synthesis with randomized hole count (5–9). The synthetic images preserve natural leaf textures and lighting.
Class ID
Label
Count
Proportion (%)
0
Background
10,750
36.2
1
No to negligible
4,736
16.0
2
Mild
12,734
42.9
3
Moderate
1,134
3.9
4
Severe
320
1.1
Total
29,674
100.0
Table 2: Distribution of the 29,674 labeled patches among the five original classes. The moderate damage and severe damage classes were subsequently combined into a single moderate-to-severe damage class, reducing the largest-to-smallest class ratio from approximately ∼ 40:1 to 8.8:1.
Table 3: Test-set performance of the Stage 2 patch-classifier configurations. H-Acc denotes binary healthy-versus-unhealthy accuracy, and D-Acc denotes moderate-to-severe damage-class accuracy.
Backbone
Params (M)
Acc (%)
F1 (macro)
F1 (weighted)
Reference
ResNet-50
23.52
79.15
0.7462
0.7992
[ 45 ]
ResNet-18
11.18
80.24
0.7469
0.8095
[ 45 ]
MobileNetV2
2.23
82.46
0.7745
0.8291
[ 46 ]
MobileNetV3-Small
1.52
82.96
0.7705
0.8337
[ 30 ]
ShuffleNetV2
1.26
81.78
0.7673
0.8231
[ 47 ]
LCA-enhanced (ours)
1.08
83.60
0.8383
0.8341
-
Table 4: Stage 2 patch-classification performance of six backbones trained under the same four-class protocol. The training set was resampled to a target of 2,000 samples per class, with AI-generated patches included in the moderate-to-severe damage class. All models were evaluated on the same held-out, field-collected test set.
Cat 1: No to negligible damage
Cat 2: Mild
Cat 3: Needs action
Avg GT = 7.4%
Avg GT = 17.0%
Avg GT = 47.0%
Panel
Configuration
Pred
∣Δ∣
RMSE
Pred
∣Δ∣
RMSE
Pred
∣Δ∣
RMSE
A
ground truth
7.4
—
—
17.0
—
—
47.0
—
—
B
4-cls, 0% overlap (our deployed model)
6.6
0.8
2.4
16.9
0.1
3.5
46.5
0.5
4.4
C
5-cls, 0% overlap
12.0
4.6
8.0
30.0
13.0
14.0
58.0
11.0
15.0
D
4-cls, 15% overlap
5.9
1.5
2.8
14.0
3.0
4.3
43.0
4.0
8.1
Table 5: Per-category agreement between the predicted NGDS and the expert-derived reference NGDS (GT). Configuration C is the five-class model; the raw input and the four-class configurations (A, B, D–F) are also shown as panels A–E in Fig. 8 . Category definitions: Cat 1 = No to negligible damage ( GT<9% , N=23 ); Cat 2 = Mild ( 9%≤GT≤29% , N=51 ); Cat 3 = Needs action ( GT>29% , N=12 ). Pred = mean predicted NGDS (%); ∣Δ∣ = ∣Pred−GT∣ in percentage points; RMSE per Eq. ( 5 ), in percentage points. Best per category in bold.
Figure 8: Effect of tiling overlap on patch-level severity predictions for a representative canopy image with an expert-derived ground-truth NGDS of 23%. (A) Raw image. Raw input. (B)–(E) Severity maps generated by the LCA-enhanced four-class model using overlap ratios of 0%, 15%, 25%, and 50%, respectively. The corresponding model-derived NGDS values were 25%, 28%, 29%, and 29%.
Figure 9: Distribution of model-derived NGDS values for 86 field images grouped by the entomologist-assigned management category. The QDA-derived decision boundaries define (1) No to negligible damage (NGDS < 9%), (2) Mild damage (9–29%), and (3) Action-level damage ( > 29%).
Figure 10: Visualization across three NGDS categories: no to negligible damage (A), mild (B), and action level damage (C).
Figure 11: Time series of model-derived NGDS values recorded by Unit B in the mature grapevine block during the 2026 field evaluation. Each point represents the NGDS generated during one scheduled operating cycle and transmitted to the web dashboard. The trace shows short-term variability and an increasing frequency of values near 50% later in the evaluation period. Because NGDS is capped at 50% by its formulation, readings at this value represent the upper limit of the metric rather than estimates exceeding 50%.
Figure 12: Battery state-of-charge record for Unit B during the 2026 field evaluation. The orange line shows measurements recorded at 30-min intervals during the daily operating window from 07:00 to 19:00. The blue line represents the daily mean, and the shaded band spans the daily minimum and maximum. Gold dashed lines mark the beginning of each daily operating window at 07:00.
Figure 13: Representative Unit B inference outputs during the 2026 field deployment: (A) August 17, NGDS = 16.9%; (B) September 1, NGDS = 29.5%; and (C) September 11, NGDS = 34.4%. Each example shows the original image (left) and patch predictions (right), with green, yellow, and red representing no to negligible damage, mild, and moderate-to-severe damage, respectively. The increasing NGDS values were consistent with the dashboard trend and observed Japanese beetle feeding damage.
Study
Model
Dataset
Assessment Task
Platform
Camera Vision
Performance
DefoliationCAM
YOLOX-Nano+LCA-enhanced MobileNet classifier (1.08 M)
Field-collected dataset of 29,674 field-collected grape canopy patches with AI-generated damaged-class augmentation
4-class defoliation severity (background, no to negligible damage, mild, moderate-to-severe) aggregated into NGDS score
NVIDIA Jetson Orin Nano (TensorRT FP16)
Field-scale (4K canopy)
83.6% Acc, 0.838 macro-F1; NGDS mapped to three QDA-derived management categories (boundaries at 9% and 29%); ∼ 1.5 s per 4K image
Table 6: Comparison of DefoliationCAM with existing deep-learning approaches for plant defoliation, disease, and pest severity assessment in terms of model, dataset, task, platform, camera vision, and performance.
Most in-field crop-monitoring devices sense a single modality -- either a camera looking at the canopy or a probe reading the soil -- and so cannot detect stress that begins below ground before it becomes visible above it. We present AgriDefend, a self-contained consumer-grade edge device that fuses both. A MobileNetV2 convolutional neural network (CNN) analyses canopy imagery, and a Random Forest classifier integrates trend features from a single soil probe: moisture, the rate of change in pH and in N, P, K, and belowground CO2 relative to ambient (a proxy for soil respiration). A logistic-regression meta-learner fuses the two into a five-class crop-health prediction, all on a Raspberry Pi 5 with no cloud or cellular link. On a controlled greenhouse drought trial with wild radish, the fused model reached 87% accuracy (R^2=0.88, MAE=0.05) at 1.4 s INT8 TensorFlow Lite inference, and its soil signals preceded visible wilting by about four days. The complete bill of materials is $205, roughly a fifth of comparable commercial nodes, positioning multimodal, sustainable crop sensing as a practical consumer-electronics product for connectivity-limited settings.
George Xu, Nikhil Vincent, Elizabeth S. Forbes
Independent Researcher Bothell, WA, USA · Univ. of Michigan, School for Environment and Sustainability Ann Arbor, MI, USA
Cocoa (Theobroma cacao) is a critical cash crop for millions of smallholder farmers in West Africa, where Cocoa Swollen Shoot Virus Disease (CSSVD) and anthracnose cause devastating yield losses. Automated disease detection from leaf images is essential for early intervention, yet deploying such systems in resource-constrained settings demands models that are small, fast, and require no internet connectivity. Existing edge-deployable plant disease systems rely on end-to-end deep learning without uncertainty quantification, while Bayesian methods for edge devices focus on hardware-level inference architectures rather than agricultural applications. We bridge this gap with TinyBayes, the first framework to combine a closed-form Bayesian classifier with a mobile-grade computer vision pipeline for crop disease detection. Our pipeline uses YOLOv8-Nano (5.9 MB) for lesion localisation, MobileNetV3-Small (3.5 MB) for feature extraction, and the Jacobi prior; a Bayesian method that provides a closed form non-iterative estimators via projection, for the classification. The Jacobi-DMR (Distributed Multinomial Regression) classifier adds only 13.5 KB to the pipeline, bringing the total model size within 9.5 MB, while achieving 78.7% accuracy on the Amini Cocoa Contamination Challenge dataset and enabling end-to-end CPU inference under 150 ms per image. We benchmark against seven classifiers including Random Forest, SVM, Ridge, Lasso, Elastic Net, XGBoost, and Jacobi-GP, and demonstrate that the Jacobi-DMR offers the best trade-off between accuracy, model size, and inference speed for edge deployment. We have proved the asymptotic equivalence and consistency, asymptotic normality and the bias correction of Jacobi-DMR. All data and codes are available here: https://github.com/shouvik-sardar/TinyBayes
Cluster closure, defined as the progressive filling of gaps between the berries in a grape bunch, is a key trait in vineyard management, impacting disease risk. However, traditional visual scoring methods are labor-intensive, subjective, and lack temporal resolution. Existing datasets rarely support fine-grained berry-level analysis, limiting the development of robust deep learning models. In this work, we present ViViD-5k, a large-scale in-field Vineyard Vision Dataset containing 5,000 images with dense annotations, including over 648,000 berry centroids and cluster segmentation masks spanning 13 grape varieties. Building on this dataset, we introduce GrapeSAM, a two-stage visual pipeline that combines point-based berry localization with prompt-based segmentation using Segment Anything, followed by transformer-based cluster segmentation. The pipeline enables automated, in-field estimation of cluster closure with minimal supervision. Quantitative results demonstrate strong segmentation and counting accuracy across diverse conditions, while visualizations confirm robustness on both in-domain and out-of-domain samples. This work provides a scalable and objective alternative to manual compactness scoring and supports high-throughput grape phenotyping with enhanced spatial detail.
Xiangzhi Tong, Chengrui Zhang, Mac Flaherty +5
School of Electrical and Computer Engineering, Cornell University, NY, USA. · Department of Electrical Engineering and Electronics, University of Liverpool, Liverpool, UK. · Horticulture Section, School of Integrative Plant Science, Cornell University, Ithaca, NY, USA. +1