AI-driven stress monitoring in melon crops via graph neural networks

Abstract

Melon cultivation is highly vulnerable to abiotic and biotic stress, and early detection remains difficult when monitoring relies on a single sensing modality. This study investigated a multimodal stress-classification framework that combined root-zone measurements and canopy reflectance descriptors for melon monitoring under greenhouse conditions. Soil pH, nitrogen, phosphorus, potassium, and temperature were acquired using an RS485 multi-parameter sensor, while canopy images were captured using a Raspberry Pi NoIR camera and converted into Normalized Difference Vegetation Index features. Each synchronized observation was represented as a graph with fixed variable nodes and correlation-based edges, enabling relation-aware learning through a Graph Convolutional Network. The proposed model was evaluated using cross-validation and compared against conventional machine learning and non-graph deep learning baselines. The graph-based model achieved the best overall classification performance, indicating that explicit modeling of soil-canopy dependencies improved discrimination between healthy and stressed plants. The results suggest that graph-structured multimodal fusion is a promising strategy for AI-assisted crop stress monitoring.