Near real-time crop phenology characterization of the US Corn Belt

Keywords: 
Abstract Type: Paper Abstract

Abstract

Crop phenology suggests the impacts of climate and environmental changes on agricultural systems. Shifts in crop phenology have profound influence on agricultural productivity and food security. Yet our ability remains limited in leveraging high-resolution images for timely and accurate crop phenology characterization at field scale. In this study, we propose a weakly supervised deep learning approach for near real-time (NRT) crop phenology characterization, which incorporates multi-source satellite imagery, including MODIS, Harmonized Landsat Sentinel-2 (HLS), and Sentinel-1, as well as environmental variables, such as temperature and soil moisture. A masked autoencoder is first trained to reconstruct temporally satellite dense time series from sparse satellite observations. The model is then fine-tuned to identify corn phenological transition dates of various corn phenological stages (i.e., emerged, silking, dough, dented, mature, and harvested) with pseudo labels estimated from the time-series observations. Validated with Crop Progress Reports (CPR) from 10 states in the U.S. Corn Belt, our proposed model can well predict the transition dates with mean absolute errors ranging from 4 to 5.5 days for different stages. Results are further validated with field observations in Illinois, U.S. from 2021 to 2022, and the mean absolution errors for most phenological stages are within 7 days for field-level phenology characterization. Trained with within-season time series, the deep learning model is capable of NRT crop phenology characterization, opening unique opportunities to understand crop growth progress and support optimized management strategies for more sustainable agricultural production.