Developing an Automatic Phenology-Based Algorithm for Rice Detection Using Sentinel-2 Time-Series Data

Abstract

Phenology-based classification methods have been developed with the goal of removing the need for ground sampling. In contrast, machine-learning-based classification methods are costly since they need extensive ground truth data to be collected, a process which is time-consuming and must be repeated on a seasonal or annual basis. Hence, in this research, we present a new automatic rule-based method, based on crop phenology, to detect rice using the time series of Sentinel-2 imagery. To do the research, the 10-m spatial resolution Sentinel-2 data acquired during the rice growing season in the red and near-infrared spectral bands for three regions in Iran were used. To develop the rules, the near-infrared band reflectance at the rice cultivation time, the red band reflectance close to the rice harvest time, and the temporal Normalized Difference Vegetation Index data were used to detect rice and discriminate it from other crops. Furthermore, the dates of data associated with the phenological stages used to develop the rice classification rules were extracted from the rice crop calendar. Although the rice fields had extensive intra-class temporal phenology variability, the algorithm performed with excellence in detecting them. The kappa coefficients obtained were 0.73, 0.94, and 0.70 for Marvdasht, Dargaz, and Qazvin, respectively.