Deep learning enabled detection of low incidence plant disease with integrated proximal and remote sensing

Abstract:

Agriculture comprises 50% of Earth’s habitable land and is critical to human health and well-being. Plant disease causes 15-30% global crop loss annually (a loss upwards of ~$220 billion USD), making disease one of the greatest threats to the environmental and financial sustainability of agriculture worldwide. Challenging our ability to use remote sensing to solve this problem is the heterogenous distribution of early stage disease at both the micro-, within canopy, and meso-, within block, scale. New, higher spectral and spatial resolution tools in remote sensing offer the potential to revolutionize disease surveillance and management in crops with low cost and high accuracy decision support. Here, we find that we can increase the utility of high resolution (sub-1m) satellite imagery to detect low severity and incidence, together known as “intensity,” foliar fungal disease in vineyards with deep learning and proximal sensing derived training data. We paired each satellite pixel with high resolution (sub-cm), color imagery of the side canopy collected by a semi-autonomous rover acquired at the same spatial interval, allowing us to accurately determine the precise level of foliar disease within each pixel. Using recently developed deep learning models to quantify disease intensity from the side canopy imagery, we developed disease maps that facilitated more accurate satellite disease detection than training with human scouting data alone. This work shows promise for the use of new spaceborne spectroscopic instruments, such as WorldView3, Planet SkySat, and NASA Surface, Biology, and Geology (SBG) as passive disease surveillance systems to support sustainable crop management.