Making optimal use of limited field-scale data for crop yield forecasting using transfer learning and Sentinel-1 and 2 data

Climate change increasingly threatens global agriculture, necessitating optimised resource management to ensure food security for an increasing population. Field-scale crop yield forecasts, using machine learning and Earth observation data, have great potential for adaptive farm management, but the development of such models is limited by the scarcity of field-scale training data. We evaluated a transfer learning (TL) approach, which entailed training and testing the model on different spatial domains, using artificial neural networks based on Sentinel-1 and 2 data to forecast crop yields in Austria and Czechia. We compared four model setups: training and testing both at the field scale (i), both at the regional scale (ii), TL with training on a regional scale and testing on a field scale with (iii) and without fine-tuning (iv). We calculated forecasts at four lead times (1-4 months) before harvest. The fine-tuned model demonstrated superior performance, achieving median R 2 of 0.52-0.69 at a one-month lead time for all crops while outperforming the model trained and applied at field scale by 0.05-0.12. TL required significantly less field-level data to achieve a performance comparable to the model trained only at the field level: 50% of the data for spring barley and maize, and only 25% for winter wheat. The model showed limitations in the leave-1-year-out cross-validation due to large differences between the years in the predictors and yields and the low number of years (6) available for training. Still, TL improved the efficiency of crop yield data utilisation and the performance of field-level forecasts. • We forecast maize, wheat & barley yields at field scale using Sentinel-1/2 data & ANN. • Transfer learning outperforms purely field-trained models for crop yield forecasting. • Transfer learning needs <50% field data to achieve skill of field-trained model.