Grupo de investigación RNM-177
Universidad de Sevilla
PAIDI. Junta de Andalucía

Rodríguez Galiano VF., Chica Olmo Mario. Analysis of the parameterization needs of different land cover classifiers: the case study of Granada Province (Spain). Vol. XXXVI. En: Mathematics of Planet Earth. Springer. 2013. 123-126.

Introduction

The application of classification methods for the mapping and monitoring of landcovers is one of the most relevant remote sensing applications. The accuracy of thegenerated mapping can be assessed quantitatively using the kappa index (K) or theoverall accuracy (OA) [1]. This accuracy is not only dependent on the classificationscene or the data themselves, but it is also strongly bound to the applied classifica-tion method. The accuracy of models induced by means of ML classifiers stronglydepends of the combination of parameters used. Therefore, a detailed comparisoncannot be carried out without previously establishing the optimal parameterizationof each model.Most studies focus on land cover mapping accuracy only, avoiding details aboutthe parameter setting used in the training of classifiers. Additionally, it is also frequentto apply the default settings recommended by the commercial software used, oronly a very limited number of experiments are carried out to determine the optimalparameters [5]. However, few are the studies based on remote sensing data whichanalyse the effect of parameter selection in ML algorithms [3]. Hence there is a needto study the impact of the parameterization of these algorithms for the classificationof land covers and land uses in depth. This chapter discusses the crucial issuesrelated to the parametrization of different up-to-date ML classifiers: classificationtrees (CT), artificial neural networks (ANN), support vector machines (SVM) andRandom Forest (RF) [2].

Top