Synergetic Classification of Long-Wave Infrared Hyperspectral and Visible Images

Abstract

A decision-level-based synergetic classification method has been proposed to conduct land-cover classification for long-wave infrared (LWIR) and high-resolution visible (VIS) images in this paper. The problem of synergic classification is challenging, since we aim for classification map at the spatial resolution of the VIS image under different imaging modes. The proposed method consists of two stages, i.e., stage of classifications for LWIR hyperspectral (HS) and VIS images separately, and stage of decision-level fusion for both classification results. For LWIR HS image, we have proposed a new semisupervised feature extraction method named as semisupervised local discriminant analysis (SLDA) for SVM classification. In parallel, spatial features which have been extracted from the high-resolution VIS image are used to combine with the spectral features for classification. In the second stage, several common decision-level fusion rules have been employed to integrate both classification results. We also present a context-based opinion pools (CBP) strategy to enhance the classification accuracy. Experiments conducted on the dataset of 2014 IEEE GRSS Data Fusion Contest show the advantage of our proposed SLDA method for HS image, and the effect of spatial–spectral features for high-resolution VIS image. Especially, the presented synergic classification strategy for HS and VIS images has higher overall accuracy and better visual effect than those only using single source image and those compared fusion methods in the experiments.