An Edge-to-Cloud System enabling Geospatial Machine Learning for Soil-Plant Moisture Classification

In Smart Farming, geospatial machine learning has the potential to enable data-driven crop management across large areas for optimization of resource consumption. In this work, we describe an edge-to-cloud system designed for soil-plant moisture detection and classification in a realistic scenario, where ground truth data are scarce and inaccurate and yet district-level irrigation water provisioning must rely on these data. Global data sources are scouted and integrated with purposely-designed edge and fog components to support soil and crop data collection, data filtering and validation, and moisture assessment. Exploiting suitable combinations of large scale data (retrieved in public satellite data hubs) and specific but inaccurate manual data, sufficiently accurate moisture assessments are obtained by machine learning algorithms, thereby avoiding expensive manual labeling of large data collections as well as human inspection of individual plots. The resulting accuracy (in the order of 80%) is comparable with that obtained by human-laboured approaches.