Abstract
Crop yield forecasting activities are essential to support decision making of farmers, private companies and public entities. While standard systems use georeferenced agro-climatic data as input to process-based simulation models, new trends entail the application of machine learning for yield prediction. In this paper we present HADES (HAzelnut yielD forEcaSt), a hazelnut yield prediction system, in which process-based modeling and machine learning techniques are hybridized and applied in Turkey. Official yields in the top hazelnut producing municipalities in 2004-2019 are used as reference data, whereas ground observations of phenology and weather data represent the main HADES inputs. A statistical analysis allows inferring the occurrence and magnitude of biennial bearing in official yields and is used to aid the calibration of a process-based hazelnut simulation model. Then, a Random Forest algorithm is deployed in regression mode using the outputs of the process-based model as predictors, together with information on hazelnut varieties, the presence of alternate bearing in the yield series, and agro-meteorological indicators. HADES predictive ability in calibration and validation was balanced, with relative root mean square error below 20%, and R(2) and Nash-Sutcliffe modeling efficiency above 0.7 considering all municipalities together. HADES paves the way for a next-generation yield prediction system, to deliver timely and robust information and enhance the sustainability of the hazelnut sector across the globe.