Random forest classification analysis of Sentinel-2 and Landsat-8 images over semi-arid environment in the Eastern Mediterranean

Abstract

Sentinel-2 land monitoring constellation mission aims to generate products similar with the Landsat-8 images, the world’s longest continuously acquired collection of space-based land earth observation data. Though both sensors share similar spectral characteristics, their Relative Spectral Response Filters (RSRFs) are not identical. It is consequently important to assess whether and to what extent endproducts, such as land use maps, may vary between these two sensors. For this purpose, the random forest classifier was applied over a semi-arid environment in the Eastern Mediterranean (Cyprus). Initially the Sentinel-2 image was sampled to the Landsat-8 spatial resolution. Then, two different classification strategies have been followed: the first one using an equal (balance) training sample between the 11 land use classes, while the second classification was based on a random training sample. In addition, land use maps were also generated based on maximum likelihood, mahalanobis distance and minimum distance pixel-based supervised classification algorithms. The overall results were evaluated based on kappa, overall, producer’s and user’s accuracies. Random forest classification has provided the best results with a kappa accuracy of 90% for both datasets while maximum likelihood algorithm has provided a kappa coefficient between 79.06% and 81.27% for Sentinel-2 and Landsat-8 sensors respectively. These results were much more improved compared to the mahalanobis and minimum distance classifiers, with an approximately kappa coefficient 69% and 66% respectively. In addition, the results obtained from the random forest have demonstrated that only a very small variance between the two datasets (Sentinel-2 and Landsat-8) exists (<3 % kappa coefficient), which can be due to the non-identical RSR filters of the sensor.