Chapter 33 - Object-Oriented Random Forest for High Resolution Land Cover Mapping Using Quickbird-2 Imagery

Abstract

This chapter presents an experimental research study on the use of random forest (RF) ensemble learning in conjunction with the object-based image analysis (OBIA) for classification of land use and land cover types using a Quickbird-2 image. It starts by presenting research objectives together with the explanations for the need of advanced classifiers and image analysis methods for satellite image classification that is a complex process affected by some uncertainties and decisions made by the analysts. Limitations of conventional parametric classifiers are discussed, and the relevancy of random forest ensemble method, which considers portions of the samples for training trees in the forest to eliminate the negative effect of mixed and atypical pixels, is established. An application is presented for the object-based random forest method with optimal parameter setting of the RF and OBIA approaches. Results are thoroughly analyzed using common accuracy metrics, and McNemar's statistical test.