A Comparison Between Classical Object Based Methods and Conditional Random Fields

Author: Jhonatan Contreras Duarte

Original title: A Comparison Between Classical Object Based Methods and Conditional Random Fields

Language: English

Electronic version: Portuguese|English

Abstract

This dissertation investigates semantic segmentation techniques for the analysis of Earth observation data. This study has two main task. The first one is to assess the potential of semantic segmentation techniques as an option to traditional image segmentation methods that typically ignore the semantic information. The second objective is to compare the semantic segmentation with the typical object-based approach (OBIA). The study is based on an implementation of semantic segmentation based on Conditional Random Fields. The object-based approach is represented in this study by the segmentation algorithm known as Multiresolution. The Random Forests classifier is used to generate the association potentials for the conditional random fields and to perform the classification task in a representative implementation of the typical object-based approach. Experiments carried out on two high spatial resolution images (8 cm) indicated a clear superiority of semantic segmentation, both in terms of spatial accuracy and thematic accuracy. Although a more extensive analysis is required for the generalization of the aforementioned conclusions, the results of this study provide enough evidence to encourage a future research on the use of semantic segmentation to compose sophisticated image classification models, in particular being part of models inspired in the OBIA approach.