You, NS; Dong, JW (2020). Examining earliest identifiable timing of crops using all available Sentinel 1/2 imagery and Google Earth Engine. ISPRS JOURNAL OF PHOTOGRAMMETRY AND REMOTE SENSING, 161, 109-123.

Timely and accurate information on crop planting areas is critical for estimating crop production, and earlier crop mapping can benefit decision-making related to crop insurance, land rental, supply-chain logistics, and food market. Previous efforts generally produce crop planting area maps after harvest and early season cropping information is rarely available. New opportunities emerge with rapid increase in satellite data acquisition and cloud computing platform such as Google Earth Engine (GEE) which can access and process a vast volume of multi-sensor images. Here we aimed to examine earliest identifiable timing (EIT) of major crops (rice, soybean, and corn) and generate early season crop maps independent of within-year field surveys in the Heilongjiang province, one most important province of grain production in China. The Random Forest classifiers were trained based on early season images and field samples in 2017, then were transferred (applied) to corresponding images in 2018 to obtain resultant maps. Six scenarios with different temporal intervals (10d, 15d, 20d, and 30d) and data integration (Sentinel-2 and Sentinel-1, a total of 16, 450 images) were compared to get the optimal crop maps. The results showed that the Sentinel-2 time series and 10-day composite outperformed in obtaining EITs and crop maps. We found various EITs for the three grain staples. Specifically, rice could be identified in the late transplanting stage (four months before harvest) with F1 score of 0.93, following by corn recognizable in the early heading stage (two months before harvest, with F1 score of 0.92) and soybean in the early pod setting stage (50 days before harvest, with F1 score of 0.91). The crop maps in the EITs based on the classifier transfer approach have comparable accuracies (overall accuracy = 0.91) comparing to the traditional post-season mapping approach based on current year's all available images and samples (overall accuracy = 0.95). This study suggests the potential of growing fine resolution observations for timely monitoring of crop planting area within season, which provides valuable and timely information for different stakeholders and decision makers.