Edge-AI Multispectral Imaging System for Detecting Plant Diseases and Stresses

Overview

Lack of early detection for plant diseases and abiotic stresses has led California to face $3 billion in yearly agricultural losses. Existing approaches rely on laborious manual scouting or delayed laboratory testing, identifying stress after visible damage occurs. Hyperspectral systems are costly and impractical for field deployment, while commercial multispectral imaging (MSI) systems typically use fixed bands, focusing on single stress types. This study presents a novel, low-cost MSI system for unified stress detection via on-device edge-AI processing. The system integrates eight monochrome cameras plus an RGB camera in a compact, UAV-ready form factor. A swappable spectral filter architecture allows cross-crop adaptability without hardware redesign. An on-device SoC runs the full pipeline including AI-based masking, spectral index generation, and stress heatmap outputs. After image capture, YOLOv8-seg performs plant segmentation on the RGB images to generate plant masks. These masks are mapped onto the geometrically aligned and cropped multispectral channel images to define regions of interest, ensuring that spectral indices are computed exclusively on valid pixels for further analysis. Abiotic stress detection was validated using a controlled basil drought experiment. A reflectance-based Water Monitor Index and canopy-scale heatmaps separated irrigated and drought-stressed plants. Biotic stress detection was demonstrated on tomato blight samples collected from Terra Amico Farms in San Martin, CA, using a multi-band disease index. Leaf-level predictions were consistent with ground truth verified by the farmer. These results demonstrate a low-cost, edge-AI multispectral system capable of unified plant stress detection, supporting the feasibility of scalable precision agriculture deployment.