Machine learning for yield prediction in Fergana valley, Central Asia

Accurate yield prediction is essential for growers, researchers, governments, the farming industry, and policymakers for social peace, food safety, security, and sustainable development. The results of earlier techniques of data collecting and analysis for yield forecasts were typically delayed, expensive, time-consuming, site-specific, and riddled with errors and uncertainties. This study is a novel approach to using high-resolution satellite data in conjunction with environmental and topographic data to predict wheat yield variability at the farm scale using machine learning. In this research, winter wheat yield prediction was based on 36 indicators in machine learning using correlation and different regression models. Winter wheat yield was predicted using linear regression (LR), decision tree (DT), and random forest (RF) regression models with scikit-learn in machine learning. More than 10,000 data points from 45 farms were trained and validated in Fergana valley, Central Asia. Results indicate that at 10 m resolution using Sentinel-2 and other secondary data such as topographic, soil, environmental, and filed data can generate an accurate wheat yield prediction map. The accuracy of all regressions is lowest for LR (R2:95, RMSE: 2.31), highest for RF (R2:98, RMSE: 1.40), and intermediate for DT regression (R2:97, RMSE: 1.85). Results also indicate that prediction in the early stage of the crop is less accurate in comparison to harvesting time as LR (R2:85, RMSE: 2.66), DT (R2:95, RMSE: 2.06), RF (R2:97, RMSE: 1.54) have different R2 and RMSE values. Appling the RF model, the winter wheat prediction is 3.29 to 4.30 t/ha therefore the total wheat production is approximately 100 t in the study area. Thus this study will demonstrate the capability of high-resolution satellite imagery and secondary data for highly accurate real-time crop yield prediction at the field scale, which can be used to assist precision agriculture and will provide a point of reference for crop area extraction, mapping, monitoring, and sustainable development with food security.