Combining time series features and data mining to detect land cover patterns: A case study in northern Mato grosso state, Brazil

Abstract

One product of the MODIS sensor (Moderate Resolution Imaging Spectroradiometer) is the EVI2 (Enhanced Vegetation Index). It generates images of around 23 observations each year, that combined can be interpreted as time series. This work presents the results of using two types of features obtained from EVI2 time series: basic and polar features. Such features were employed in automatic classification for land cover mapping, and we compared the influence of using single pixel versus object-based observations. The features were used to generate classification models using the Random Forest algorithm. Classes of interest included Agricultural Area, Pasture and Forest. Results achieved accuracies up to 91,70% for the northern region of Mato Grosso state, Brazil.