IFPRI and UMD are working as partners to understand how satellite data can help us better understand crop production and loss in Tanzania, using drones for ground truthing.
Ground-truthed data are needed to verify satellite agricultural data. This study investigates a method to validate NDVIs from satellite data using derived NDVIs from Unmanned Aerial Vehicles (UAV) and to validate yields using self-reported surveys of farmers by mobile technology. A partnership was formed between IFPRI, University of Maryland and local private sector organizations Agrinfo Social Enterprise, WeRobotics and Tanzania Flying Labs to help achieve this goal.
The initial analysis on yield estimation from NDVIs measured with satellite imagery (Sentinel and Landsat) was completed by UMD using survey data from IFPRI. Satellite-based NDVI data were used to build field-scale yield models, 5 popular machine learning algorithms were applied to forecast crop yields. Overall, models perform better for 2018 compared to 2017, but r-square values are low and error (RMSE) is high. CatBoost, based on gradient boosting, seems to be performing the best of the 5 models. The team is looking into reasons for strength or weakness of predictive power, which will be useful for others in their use of satellite data to track yields.