Improving Image Classification

Abstract

Multidate imagery and satellite image derivatives such as vegetation and texture indices have been reported to improve image classification. However, the increase in additional predictor variables has also resulted in high data dimensionality and redundancy. Feature selection and extraction can be used to reduce high data dimensionality and minimize redundancy. The purpose of this chapter is to test whether feature selection can improve image classification. In this chapter, image classification will be performed using two different approaches. First, image classification is performed using the random forests (RF) classifier and multiple data sets (that consist of multidate Landsat 5 TM imagery, and vegetation and texture indices). Second, image classification is performed using the RF classifier with feature selection and multiple data sets. While the tutorial exercises indicate that feature selection did not improve image classification accuracy, it reduced the number of predictor variables.