A Novel Phenology Guided Deep Learning Model for Within-Season Field-Level Crop Mapping

Zijun Yang,  University of Illinois at Urbana-Champaign,  zijuny2@illinois.edu (Presenter)
Chunyuan Diao,  University of Illinois,  chunyuan@illinois.edu
Feng Gao,  USDA-ARS HRSL,  feng.gao@ars.usda.gov

Timely and accurate crop type mapping during the growing season is critical for a variety of agricultural applications, including near-real-time monitoring of the crop growing conditions and crop yield forecasting. The Incorporation of historical time-series remote sensing data with crop type reference for model building has provided a feasible means for within-season crop mapping. Yet the spatial and temporal variations in crop phenology may hinder the model scalability and transferability, as such models are trained with the historical data but applied to the current season. In this study, we propose a novel phenology guided modeling strategy which incorporates within-season crop phenology information identified from temporally dense satellite fusion images for scalable within-season field-level crop mapping. Daily 30-meter satellite images are first generated by a hybrid deep learning spatiotemporal fusion model through blending harmonized Landsat and Sentinel-2 and MODIS data. The within-season emergence (WISE) algorithm is then employed to extract within-season crop phenology information, which is utilized for normalizing the time-series remote sensing data. The WISE-phenology normalized data are subsequently fed into machine learning/deep learning models for crop type mapping in different years and regions. Results from the US Midwest suggest that the phenology guided model achieves high accuracy at the end of the growing season, ranging from 91% to 93.5%. An overall accuracy of 90% can be reached in late July. Compared to conventional approaches, the phenology guided model exhibits advantages throughout the growing season. Through the integration of spatiotemporal image fusion, within-season crop phenology, and advanced deep learning techniques, The proposed phenology guided modeling strategy holds great potential in providing accurate and timely field-level crop type maps over extended geographical regions.

Poster Location ID: 3-45

Presentation Type: Poster

Session: Poster Session 3

Session Date: Thu (May 11) 3:00-5:00 PM

CCE Program: Other