Authors: Xiaoai Dai*, Chengdu University of Technology, University of Kansas, Shouheng Guo, Chengdu University of Technology, Xingong Li, University of Kansas, Shenglan Xu, Chengdu University of Technology
Topics: Remote Sensing, Qualitative Methods, Land Use
Keywords: hyperspectral image classification, t-SNE dimension reduction, AdaBoost algorithm, feature extraction, ensemble learning
Session Type: Poster
Start / End Time: 1:20 PM / 3:00 PM
Room: Napoleon Foyer/Common St. Corridor, Sheraton, 3rd Floor
Presentation File: Download
Compared with panchromatic or multi-spectral images, hyperspectral images have significant advantage on spectral resolution which enables precise identification and classification of land surface components. However, traditional image classification methods do not fit for hyperspectral images because of its complicated data structure. In this study, we propose a novel hyperspectral image classification method which combines the t-SNE (t-distributed stochastic neighbor embedding) and AdaBoost algorithms and fully considers spatial context in the classification process. T-SNE algorithm was first used to map high dimensional data to low dimensional subspace with high inter-class separability. After that, the pixel features within the neighborhood of a pixel is also added. The AdaBoost algorithm is used to construct a highly accurate classifier ensemble which integrates different decision tree classifiers randomly. High accuracy can be obtained using our method without intensive parameter calibration and the convergence rate is faster than other conventional methods, such as support vector machines，random forest and iterative integration of decision tree algorithm. We tested our method on Hyspex imaging spectrometer data, and our results indicated that the optimal scenario is when the characteristic number of t-SNE algorithm is 4 and the number of decision tree classifier of AdaBoost algorithm is 150. The best classification accuracy from our method is 93.06%. We further improved our method by adding spatial correlation information to pixel feature vector, which increased classification accuracy to 94.03%.