A Semi-Supervised Hybrid Approach for Multitemporal Multi-Region Multisensor Landsat Data Classification

Abstract

The classification of land covers is one of the most relevant tasks carried on to understand the state of a certain region. Additional studies about the biodiversity, hydrology, human impact, modeling dynamics, and phenology in the study area, can be carried on. In these cases, a wide temporal series of images need to be considered in order to get the tendencies throughout the years. In some regions, such as the South-West part of Colombia (Andean region), studies over large areas are needed in order to obtain unified and coherent statistics that can be representative of the region. This means that different images, acquired by the same satellite and over different areas, or acquired by different sensors, or at different times, need to be classified. Standard classification methods do not work properly to perform this task, due to the heterogeneity in both land cover and orography. This paper presents a hybrid approach for the classification of multitemporal, multiregion, and multisensor images. Classification and regression trees (CART) decision tree and an SVM-based clustering were used in cascade in order to get the final classification maps. Experimental results carried over three Landsat Path/Rows, three sensors, and six different years, confirm the effectiveness of the proposed approach, where the overall accuracy was of 93% with a kappa factor of 0.92.