A comparative analysis of different phenological information retrieved from Sentinel-2 time series images to improve crop classification: a machine learning approach

Abstract

In this study, the potential of phenological indicators derived from Sentinel-2A (S2) time series were evaluated to explore the key variables that allow identifying both cropland and crop types. Based on the derived S2 phenological metrics and fitted vegetation indices (VI), 10 feature sets were developed and assessed to discriminate different crop types via Random Forest (RF) classifier. The comparison between VI data-based classifications has shown that NDVI and EVI2 phenological sets could delineate and identify crop types more accurately compared to RENDVI data. Overall, the combined use of fitted VI and phenological features rather than being used separately achieved the best performances. Further, the result of using optimum features was the most accurate among 10 feature sets, with an overall accuracy of 88% and kappa of 0.84. This study constitutes a substantial improvement in crop type identification, which gives a valuable tool to monitor agricultural areas.

Keywords:

• Phenological metrics
• TIMESAT
• random forest
• crop mapping
• Sentinel-2

Disclosure statement

No potential conflict of interest was reported by the authors.