Yield forecasting with machine learning and small data: What gains for grains?

• We present a system to forecast crop yield with machine learning at regional level. This fully automated selects the best features and model. Leaving one year out is required rigorously assess model performance. Yield forecasts based on are accurate even in low-yielding years. Extensive calibration deliver significant gains predictive power. Forecasting yields important for food security, particular predict where production likely drop. Climate records remotely-sensed data have become instrumental sources of forecasting systems. Similarly, methods increasingly used process big Earth observation data. However, access necessary train such algorithms often limited food-insecure countries. Here, we evaluate performance small monthly basis between start end growing season. To do so, developed robust machine-learning pipeline which prediction. Taking Algeria as case study, predicted national barley, soft wheat durum an accuracy 0.16–0.2 t/ha (13-14% mean yield) within The models always outperformed simple benchmark models. was confirmed years, particularly relevant early warning. Nonetheless, differences were not practical significance. Besides, up 60% that tested, stresses importance proper selection. For forecasting, like many application domains, has delivered improvement superiority over benchmarks achieved after extensive calibration, especially when dealing