Phenological normalization can improve in-season classification of maize and soybean: A case study in the central US Corn Belt

Abstract

Within-season estimates of crop-specific planted area, conditions, and expected yields are critical for decision- and policy-making related to agriculture and food security. However, these estimates require in-season crop type maps which are not currently widely available. Machine learning and remote sensing data can be used to create crop type maps that provide crop-specific land cover classifications to enable timely analysis at field scales throughout the growing season to complement traditional reporting efforts. However, existing methods are often limited by lower performance on test data from seasons with different patterns not seen during training and inability to provide classifications during the growing season when most operationally relevant. We present a new approach to in-season crop type mapping that addresses inter-annual domain shift by normalizing satellite observations by land surface phenology stage. These phenology-normalized observations are input to a neural network that extracts temporal and spatial features using both recurrent and convolutional layers respectively. Using Harmonized Landsat and Sentinel-2 (HLS) and Sentinel-1 SAR observations with test data from the U.S. states of Iowa in 2019 (late planting year) and Illinois in 2020 (standard year), we showed that this method enabled good performance in both scenarios throughout the growing season (71% and 72% in early-season, 80% and 84% in mid-season, 81% and 85% in late-season for Iowa 2019 and Illinois 2020 respectively).