Crop mapping using Sentinel-1 and Sentinel-2 images and random forest algorithm

Abstract

Crop mapping can provide valuable information for agricultural land management and crop estimation. This study investigated the spectral bands of Sentinel-2 time series, their vegetation indices, Sentinel-1 VV and VH time series of the radar backscatter coefficient, and the VV/VH ratio. This study explored the importance of the red-edge wavelengths of Sentinel-2 imagery for crop mapping using the random forest (RF) method. Therefore, the 2019 time series of Sentinel-1 and 2 images for the growing season in northwest Ardabil, Iran, were retrieved from the Google Earth Engine. After pre-processing, these images were segmented using the multi-scale method, and then the spectral features of optical imagery and the radar backscatter coefficient were extracted for each segment. To examine the importance and role of red-edge wavelengths, in addition to the three red-edge bands, visible and infrared wavelengths, and plant indices derived from these bands, red-edge indices were also factored in as input features. Overall, nine scenarios were simulated using different inputs and combinations. In each scenario, key features were identified using RF feature selection and introduced as inputs for the RF algorithm for an object-oriented classification. The research results showed that the addition of red-edge bands and the derived indices increased the accuracy of crop type mapping. The best result was obtained for a combination of optical and radar images with an overall accuracy of 87.59% and a kappa coefficient of 85.40%.