Transfer learning for improved crop yield predictions in a cross-scale pathway: a case study for Brazilian national soybean

Abstract

Crop yield prediction at a fine spatial scale is crucial for improving agricultural management and resource allocations. Many countries and regions lack fine-scale yield data for fine-scale modeling and thus have to use a “cross-scale pathway”, where coarse-scale (e.g., state-level) data are used to train a model for fine-scale (e.g., county-level or field-level) yield predictions. However, the cross-scale pathway has limited effectiveness in predicting yield due to issues with data availability and model scalability. In this study, we quantify the benefits of transfer learning in the cross-scale pathway. We applied transfer learning by fine-tuning a previously trained and validated AI-based machine learning model, originally developed for field-level soybean yield predictions in the United States, using Brazilian state-level data to predict Brazilian municipal-level soybean yield. Despite differences in environmental conditions, crop phenology, and yield responses between the U.S. and Brazil, we show that transfer learning improves the municipal-level predictions from the cross-scale pathway by increasing the R2 from 0.29 (without transfer learning) to 0.44. Notably, this is achieved without using any municipal-level data and relying only on scarce state-level observations. When the municipal-level data were used, the transfer learning achieved an R2 of 0.57, the most stable high performance compared with previous studies. The effectiveness of the cross-scale pathway, thus, increases from 50% to 78% with transfer learning. These findings demonstrate the benefits of transfer learning in the cross-scale pathway under data-limited conditions, and underscore the potential for global crop yield predictions across scales.

Noemi Vergopolan

Assistant Professor of Earth, Environmental and Planetary Sciences

Computational hydrologist working on high-resolution water prediction for climate, agriculture and decision-making.