CropSync: towards a large-scale operational framework for within-season crop type mapping using Google Street View and Harmonized Landsat and Sentinel-2 imagery

Abstract

Within-season crop type mapping plays a critical role in guiding agricultural policy decisions by providing timely crop type data in given regions. Existing studies on within-season crop type mapping typically leverage historical remote sensing and crop type reference data. However, the adaptability and transferability of these models to the current season are limited due to annual and regional variations in crop growth. Our study proposes a novel CropSync framework for scalable, within-season crop type mapping, combining Google Street View (GSV) and Harmonized Landsat and Sentinel-2 (HLS) satellite images. CropSync incorporates three innovative components: a GSV-based crop type ground truth collection method, the emergence-based thermal phenological normalization (EMET) method, and a self-attention based LSTM (SAtLSTM) model. The GSV-based method is developed to retrieve high-quality, current-year crop type labels from GSV images. The EMET method synergizes the within-season crop emergence model with thermal time accumulation to normalize crop phenological stages in HLS time series data. The EMET normalized HLS time series and the collected crop type labels are then utilized to train the SAtLSTM model for within-season crop type mapping. CropSync significantly addresses the challenge caused by spatiotemporal variations in crop phenological dynamics, offering a substantial improvement over conventional calendar-based and historical data-driven methods. Results in the southern United States from 2018 to 2023 demonstrate CropSync's superiority, achieving over 92% accuracy in classifying key crops including corn, soybean, rice, and cotton by season's end. The robustness and adaptability of CropSync highlight its potential for extensive operational application in within-season crop type mapping.