Dimensionality reduction for hyperspectral images to improve object-based image classification using feature selection and principal components analysis

In parallel to the increasing accessibility of high resolution imagery, object-based image analysis (OBIA) has recently become a hot topic in remote sensing. Segmented objects significantly reduce the high-dimensionality and low- training size problems for classification process. On the other hand, Estimation of Scale Parameter (ESP-2) tool, which is commonly used to estimate optimal scale value, is limited to 30 spectral bands. This limits its use in hyperspectral image analysis and, thus, compulsory reduction in number of spectral bands is required. In this study, a 103-band Pavia University hyperspectral dataset was utilized to conduct the objectives of the study. In this context, a feature extraction method (Principal Components Analysis-PCA) and a feature selection method (random forest-RF) were utilized for reducing the number of spectral bands to be used in ESP-2 tool for searching optimal scale parameter. While multi-resolution segmentation approach was employed with optimal parameter setting using ESP-2 tool, two robust machine learning methods, namely RF and rotation forest (RotFor) were applied for classification of the constructed image objects. The results showed that the classification accuracy obtained using RotFor was much higher than the random forest classifier (up to 6%) for the dataset selected by the random forest algorithm (10 bands). However, the difference in classifiers' performances was about 1.5% for the PCA dataset (first seven components representing 99% of the total variance). The performance of RF and RotFor classifiers was statistically analyzed using McNemar's test and found that difference in classifier performances was statisticallydifferent for the PCA and RF-selected datasets.