Evaluating the relative importance of spectral, topographic, and texture features in the ecosystem classification

In the past decade, the ensemble of classifiers such as decision trees has been proposed as a new strategy for the improvement of the performance of individual tree classifiers. However, it posed a new challenge of feature selection for high-dimensional data classification. The traditional feature ranking methods for individual tree was not suitable for the AdaBoost trees algorithm. In this study, we proposed an improved method to evaluate the relative feature importance of multi-source input data in the AdaBoost tree algorithm. The feature selection algorithm has been applied to an ecosystem classification in the Eastern Mojave Desert through multi-season LANDSAT TM/ETM+ images, QuickBird images and terrain-related GIS data layers. A total of 60 spectral layers derived from multi-season TM/ETM+ images, 76 texture layers from high-spatial resolution QuickBird images, and 6 terrain-related GIS layers were pooled in the AdaBoost trees classifier. We analyzed and discussed the feature ranking in the AdaBoost trees, selected the top ranking features for the AdaBoost trees, and acquire the similar classification accuracy. The results showed that the topographic features had major influence for ecosystem classification, followed by spectral layers; and the texture layers derived from QuickBird images could further significantly increase the classification accuracy.