Spatiotemporal Deep Learning for Forecasting Vector-Borne Crop Disease Risk across Indian States

Abstract

Vector-borne crop diseases pose a severe threat to agriculture in India, reducing crop yields and threatening food security. Effective forecasting of disease outbreaks can enable timely interventions, but traditional epidemiological models often struggle with complex spatiotemporal dynamics and climate influences. This paper explores the application of spatiotemporal deep learning models to forecast the risk of major vector-borne and spatially spreading crop diseases (such as rice blast, wheat yellow rust, and leaf curl virus diseases) across different Indian states. We review recent advances in deep learning – including convolutional Long Short-Term Memory (ConvLSTM) networks and spatiotemporal graph convolutional networks (ST-GCN) – which can capture both spatial spread and temporal progression of diseases. Multi-modal data, including weather variables, remote sensing indices, and historical disease incidence, are integrated to train these models. In experimental evaluations, deep learning models significantly outperformed baseline statistical models, achieving high predictive accuracy (e.g., area under ROC > 0.90 in some cases) for outbreak risk forecasting. A case study is presented illustrating how a ConvLSTM model accurately predicts seasonal disease incidence patterns in a sample state. The results demonstrate that spatiotemporal deep learning can provide early warning of crop disease outbreaks, potentially guiding preemptive management strategies. However, challenges remain in terms of data availability, model interpretability, and generalization across diverse crops and regions. This study highlights the promise of deep learning-based forecasting as a component of next-generation crop protection for India, enabling data-driven decisions to mitigate vector-borne disease impacts under current and future climate conditions.