关键词: 有偏SVM, 支持向量数据描述, 高斯数据描述, 单类支持向量机, 遥感图像分类, 多类SVM

Abstract: Traditional classifier is made up of both positive and negative data. It is a common situation in remote sensing image classification：users are only interested in one specific land-cover type. However, labeling land-cover is a time consuming and labor intensive process, and unlabeled data are usually obtained easily and contain useful information. For this reason, a remote sensing image classification method based on Positive and Unlabeled data（PUL） is proposed. Firstly, according to the inherent characteristics of positive data and combined with support vector data description confident positive and negative samples can be extracted from unlabeled data, and those examples are eliminated from unlabeled set. Then it uses above extracted samples to train a SVM classifier and extract relative confident positive and negative sample from unlabeled set again. The extraction rule is based on the performance of unlabeled set in the SVM classifier. The last step is weighted SVM process. The weight of initial positive and negative samples is 1. The weight of samples extracted by SVM classifier is between 0 and 1. To verify the effectiveness of PUL method, it does classification experiment in remote sensing image and is compared with One-Class SVM（OC-SVM）, Gauss Data Description（GDD）, Support Vector Data Description（SVDD）, Biased SVM and Multi-class SVM. The results show that PUL is helpful to the improvement of classification and better than above OC-SVM methods and Multi-class SVM.

Key words: Biased Support Vector Machine（SVM）, Support Vector Data Description（SVDD）, Gauss Data Description（GDD）, One-Class SVM（OC-SVM）, remote sensing image classification, Multi-class SVM