A novel method for automatic potato mapping using time series of Sentinel-2 images

Abstract

Crop maps produced by remote sensing data play an important role in agricultural crop studies. Most of the crop mapping methods rely on field samples to train a model, which is a costly and time-consuming process. On the other hand, automatic crop mapping methods which are based on unique spectral characteristics are independent of ground truth data. Since most crops have specific spectral and temporal features during the growing season, this research aims at developing a new automatic method to discriminate potato from other crops using Sentinel-2 time series imageries. In this research, Crop type data of three study sites in Iran consisting of 2019 fields of potato and other crops, which were sampled by a GPS receiver were used. Moreover, an additional site in the United States comprised of 880 fields from Cropland Data Layer (CDL) in raster format, was also utilized. We employed 50% (292) of the Hamedan fields to train the model and the remaining 2607 fields from other sites and 50% from Hamedan were used to validate the model. Then, the temporal reflectance spectra of various crops and potato were extracted and considered. Results show that potato has four unique spectral characteristics which can be utilized to distinguish potato fields. These include the near-infrared reflectance values at the cultivation and harvest dates, variations of the near-infrared reflectance at the greenness peak time and the ratio of the near-infrared reflectance values to the red reflectance values at the greenness peak. Therefore, a novel feature was proposed based on a combination of the above spectral characteristics for discrimination of potato fields with a kappa coefficient of higher than 0.8 and an overall accuracy of better than 90%, in the four study sites.