A weakly supervised framework for high-resolution crop yield forecasts

Abstract

Predictor inputs and labels (e.g. yield data) for crop yield forecasting are not always available at the same spatial resolution. Common statistical and machine learning methods require inputs and labels at the same resolution. Therefore, they cannot produce high resolution (HR) yield forecasts in the absence of HR yield data. We propose a weakly supervised (WS) deep learning framework that uses HR inputs and low resolution (LR) labels (crop areas and yields) to produce HR forecasts. The forecasting model was calibrated by aggregating HR forecasts and comparing with LR crop area and yield statistics. The framework was evaluated by disaggregating yields from parent statistical regions to sub-regions for five countries and two crops in Europe. Similarly, corn yields were disaggregated from counties to 10 km grids in the US. The performance of WS models was compared with naive disaggregation (ND) models, which assigned LR forecasts for a region or county to all HR sub-units, and strongly supervised models trained with HR yield labels. In Europe, all models (ND, WS and strongly supervised) were statistically similar, mainly due to the effect of yield trend. In the US, the WS models performed even better than the strongly supervised models. Based on Kendall’s rank correlation coefficient, the WS model forecasts captured significant amounts of HR yield variability. Combining information from WS with Trend model (using LR yield trend) and WS No Trend model (not using yield trend) provided good estimates of yields as well as spatial variability among sub-regions or grids. High resolution crop yield forecasts are useful to policymakers and other stakeholders for local analysis and monitoring. Our weakly supervised framework produces such forecasts even in the absence of high resolution yield data.