Object-based image analysis for urban land cover classification in the city of Campinas - SP, Brazil

Abstract

Classifiers that make use of pixel-by-pixel approaches are limited in the high spatial and radiometric resolution of urban areas, that happens mostly because of the similarity between the target's spectral response like ceramic roofs and bare soil. Because of that, the literature favors approaches that make use of object-oriented analysis for image interpretation, those approaches make a better use of the high spatial resolution and do not use only the target spectral response. Assuming that the object-oriented analysis is a favorable approach to be employed for intra-urban image classification, this paper will assess the results of such approach through an implementation of it in an urbanized area from the city of Campinas (Brazil), which has a size close to twelve square kilometers. Making use of the fusion of high spatial resolution image from Worldview-2 sensor and it's panchromatic band, the experiments were performed with the use of eCognition Developer 8 as the segmentation platform, and the classification being based on a decision tree generated by J48 (C4.5) algorithm on the software WEKA. This work also assess which approach best suits the experiment needs, being an optimal attribute selection achieved through a Wrapper filter, with a final kappa statistic of 0.9425.

Avaliação da performance do algoritmo J48 para construção de modelos baseados em árvores de decisão

“…Diversos trabalhos têm divulgado a superioridade do método Wrapper em relação a outros métodos para seleção de atributos. A exemplo, França et al (2015) mostra em seu trabalho um incremento no poder preditivo do modelo em árvore de decisão com a seleção de atributos por meio do método Wrapper reduzindo em cerca de 9% o erro em relação às predições sem seleção de atributos e, em cerca de 2,5% em relação à seleção realizada pelo método CFS. A Tab.…”