Database-Driven Crop Disease Surveillance from Remote Sensing and Field Reports
Main article
Abstract
Timely and accurate detection of crop diseases is critical for food security in rain-fed agricultural systems where a single outbreak can eliminate 40–70 percent of seasonal yield. Existing surveillance approaches rely on either manual field scouting—which is spatially sparse and labour-intensive—or on remote sensing pipelines that lack structured integration with ground-truth observations. This paper introduces CropDiseaseDB, a relational–spatial database system that unifies multi-spectral satellite imagery from Sentinel-2 and Landsat-8 with georeferenced field disease reports across five major crops and twelve disease classes in a semi-arid agricultural region of India. The database is designed around a formally specified schema with spatial indexing, field dictionaries, version-controlled data pipelines, and an open REST API. We demonstrate its utility by training and evaluating a Spatio-Temporal Graph Neural Network (ST-GNN) that jointly exploits spectral features and inter-farm adjacency relationships to predict disease outbreaks. Evaluated over five growing seasons (2020–2022), the ST-GNN achieves an F1 score of 0.847 and a mean early-warning lead time of 3.6 days, outperforming SVM, Random Forest, CNN, and LSTM baselines. Ablation experiments confirm that both remote sensing and graph connectivity components are necessary for cross-season generalisation. CropDiseaseDB is openly available and supports reproducible experimentation, automated analytical pipelines, and evidence-based plant health management.
