Extreme value theory-based calibration for the fusion of multiple features in high-resolution satellite scene classification

Abstract

This article presents a hierarchical classification method for high-resolution satellite imagery incorporating extreme value theory (EVT)-based normalization to calibrate multiple-feature scores. First, a simple linear iterative clustering algorithm is used to over-segment an image to build a superpixel representation of the scene. Then, each superpixel is characterized by three different visual descriptors. Finally, a two-phase classification model is proposed for achieving classification of the scene: (1) in the first phase, a support vector machine (SVM) with histogram intersection kernel is applied to each feature channel to obtain raw soft probability; and (2) in the second phase, the derived soft outputs are multiplied to build a product space for score-level fusion. The fused scores are subsequently further calibrated using the EVT and fed to an L1-regularized L2-loss SVM to obtain the final result. Experimental analysis on high-resolution satellite scenes shows that the proposed method achieves promising classification results and outperforms other competitive methods.

Funding

The research was supported in part by the National Key Basic Research and Development Programme of China under Contract 2013CB733404; the Chinese National Natural Sciences Foundation (NSFC) [grant number 61271401, 61331016].