Low-Cost RGB Imaging for Plant-Stress Detection in California: A Benchmark-to-Field Evaluation
The Oxford Journal of Student Scholarship · 2026-10
DOI: 10.65161/rec0pfsow9dmfdllk ↗Kurz gefasst
Fachartikel in der Zeitschrift The Oxford Journal of Student Scholarship, 2026-10. Verfasst von Adit Shah.
Abstract (Originalsprache)
Insect pests and other stressors cost California agriculture billions of dollars each year, and early detection is one of the most effective defenses.AI-based detection systems that pair deep learning with multispectral or hyperspectral imagery report roughly 90-95% accuracy, but they require sensors costing $10,000 or more, large labeled datasets, and technical expertise that most small to medium-sized farms lack.This study asks whether a consumer RGB drone (about $400) combined with a training-free, color-based computer-vision algorithm can detect visible plant stress well enough to serve as a first step toward AI-based monitoring.The algorithm segments foliage in HSV color space, computes the fraction of foliage that is discolored, and labels a plant as stressed when that fraction reaches 0.20.On 2,012 plants annotated in drone video from four types of Southern California sites, it reached 70.1% accuracy but only 39.0% stress recall, missing 347 of 569 stressed plants.On all 54,305 images of the PlantVillage leaf benchmark, the same rule ranked diseased leaves above healthy ones well (AUC 0.814) but flagged only 17.6% of them, because most leaf lesions are dark rather than yellow; a classifier trained on RGB color-index features reached 96.7% accuracy on held-out PlantVillage images.The same algorithm's ability to separate stressed plants from healthy plants fell from AUC 0.814 on curated leaves to 0.645 in the field, showing that curated benchmarks overstate field readiness and that visible color alone misses subtle stress.The algorithm measures visible plant stress only; it does not detect insects or determine whether stress was caused by insect damage.We therefore utilized low-cost RGB imaging as a Phase-1 baseline that progresses towards CNN architecture (Phase 2) and that ultimately culminates into multispectral sensing with a proposed ResCNN-HDBSCAN model (Phase 3).
Zitieren
APA
Shah, A. (2026). Low-Cost RGB Imaging for Plant-Stress Detection in California: A Benchmark-to-Field Evaluation. The Oxford Journal of Student Scholarship. https://doi.org/10.65161/rec0pfsow9dmfdllk
Vancouver
Shah A. Low-Cost RGB Imaging for Plant-Stress Detection in California: A Benchmark-to-Field Evaluation. The Oxford Journal of Student Scholarship. 2026. doi:10.65161/rec0pfsow9dmfdllk
BibTeX
@article{Shah2026,
author = {Shah, Adit},
title = {Low-Cost RGB Imaging for Plant-Stress Detection in California: A Benchmark-to-Field Evaluation},
journal = {The Oxford Journal of Student Scholarship},
year = {2026},
doi = {10.65161/rec0pfsow9dmfdllk}
}
EUVIMED