Early Within-Season Yield Prediction and Disease Detection Using Sentinel Satellite Imageries and Machine Learning Technologies in Biomass Sorghum

摘要

Sorghum is grown for several purposes including biomass for producing energy and fodder, and grain for producing health-promoting foods. Sorghum is a drought resistant cereal with low input requirements, making it one of the most promising crops under the world's tropics and higher latitudes. Crop monitoring, one of the leading activities in smart fanning, can help cut production costs and more so under climate change. In this study, Sentinel 2A and 2B-derived fAPAR and NDVI data were used to monitor sorghum phenology, foliar diseases, and to predict aboveground biomass yields months before harvest, using machine learning approaches including Bayesian methods and region-convolutional neural network. The results obtained in this work were encouraging. We were able to predict biomass yields up to 6 months before harvest with mean absolute percentage error (MAPE) < 0.2, while diseases were detected with accuracy up to 90%. The best machine learning algorithm was Bayesian additive regression trees (bartMachine method), while the best biomass yields prediction regressors were the days of year 150 and 165. These results were achieved at a Pilot level and the technologies showed industrial scale implementation potentials with tremendous benefits for the fanner, extension services, policy makers, and other parties at interest