Assessment of the effectiveness of supervised and unsupervised methods: maximizing land-cover classification accuracy with spectral indices data

Abstract

This study is aimed at evaluating the effectiveness of different supervised and unsupervised methods with information derived from Landsat satellite images and fieldwork in order to maximize the land cover classification accuracy in an area with geomorphologic differences and heterogeneous edaphic characteristics located in the southwest of the Pampas (Argentina). We test two datasets: bands-based and indices-based and also we analyze the spectral behavior of each land cover identified by field trips and surveys with farmers to improve the spatial samples employed in the digital processing. Complementarily, we study the spatial and temporal information about the land cover changes during 2000 to 2016. The classification based on indices widely outperforms the analyses based on bands. The best methods to classify the land cover are the Mahalanobis distance and the maximum likelihood. The values of kappa coefficient and overall accuracy obtain from these two methods allow us to realize a multitemporal study. This study provides essential information for semiarid regions with rain-fed agriculture and livestock activities worldwide. The knowledge obtained quickly and accurately about the land covers and their changes provides essential information about the past and current situations and can be used to predict likely future trends.