A novel deep learning framework for within-season field-level crop phenology characterization

Abstract

Crop phenology is an important indicator of how agricultural systems respond to climate and environmental changes. Timely and accurate crop phenology characterization of farm fields is crucial for assessing agricultural management and productivity under climate change and weather extremes. Previous efforts have mainly been devoted to retroactively monitoring crop phenology at regional scales with coarse resolution satellite imagery. Yet our ability remains limited in timely within-season crop phenology characterization at field levels, due to the challenges in high resolution and near-real-time crop monitoring. In this study, we propose a novel deep learning framework for within-season phenology retrieval of corn fields by integrating multi-scale satellite imagery, including MODIS, Harmonized Landsat Sentinel-2 (HLS), and Sentinel-1. MODIS and HLS time series are first fused to reconstruct temporally dense time series of optical remote sensing observations. Built upon transformer, a novel deep learning model is trained to identify phenological transition dates using a combination of time series of optical, SAR, and meteorological observations. A variety of corn phenological stages (i.e., emerged, silking, dough, dented, mature, and harvested) are retrieved and validated with field observations in Illinois, US from 2020 to 2022. Results for corn in Illinois suggest that the proposed deep learning framework can accurately predict the transition dates for various corn phenological stages at the field level. Trained with within-season time series, the framework also holds great potential for early-season crop phenology characterization and forecasting, opening unique opportunities to support optimized management strategies for more sustainable agricultural production.