LAUSR

ABSTRACT Due to differences in environmental factors, the phenology of the same crop is different every year, causing divergent performances of the classifier built by spectral or time-series features Here, we proposed a random forest classifier (RFC) based on an asymmetric double S curve model fitted by accumulated temperature (AT) and Vegetation Index (VI), which can be applied in different years without ground samples. We built AT and VI time series from Moderate Resolution Imaging Spectroradiometer 8-day composites of land surface temperatures and Sentinel-2 and Landsat-8, respectively. The RFC was trained by characteristics from the asymmetric double S curve. We prepared RFC by ground samples of 2018 and 2019 and then mapped crops of the same region in 2017. Results indicated that, compared with diverse VI-AT series, the overall accuracy based on universal normalized vegetation index (UNVI) was the best of all (2017: F1 = 0.91, 2018: F1 = 0.92, 2019: F1 = 0.91) and better than that based on the UNVI-TIME series (2017: F1 = 0.84, 2018: F1 = 0.81, 2019: F1 = 0.88). It proved that the classification features from the VI-AT series have smaller intra-class differences in 2017, 2018, and 2019.