Aligning satellite-based phenology in a deep learning model for improved crop yield estimates over large regions

Accurate estimate of crop yield is crucial for agricultural planning and resource allocation. Crop yield estimation over large regions is especially challenging because of large variations of crop phenology and environmental conditions. Environmental stresses, such as drought and heat, during different phenological stages impose different impacts on plant growth and thus on final grain yield. We hypothesize that incorporating the phenology information, specifically, aligning environmental conditions with plant phenological timing, can better simulate the impact of environmental conditions on crop yield and thus improve crop yield estimates. To test our hypothesis, we build a deep learning model to predict crop yield using growing-season weather and satellite time series that are either aligned with field-level phenology or not. Based on rainfed corn fields in the 12 states in the US Midwest between 2011 and 2020, we show that applying the aligned phenology in the deep learning model decreases the error of corn yield estimates from 34.1 bu/ac (or 21.8 %) to 29.2 bu/ac (or 18.7 %) and increase the explained yield variability from 61 % to 71 % for leave-one-year-out predictions. The improvements are most significant in fields with yield loss that are associated with early phenological timing. Specifically, in the extreme drought year 2012, the phenology was ∼20 days earlier than normal years, aligning the input sequences with field-level phenological timing captures 78 % of the yield loss, reduces the error by 15 %, and increases the explained variability over 50 %. Our results demonstrate the importance of aligned phenological timing for capturing environmental stress in crop yield estimates and show great potential of phenology alignment in accurate crop yield estimates over large regions.