Remote Sensing Image Classification Based On Semi-supervised Learning

Learning methods classifier used is one of the most important parts in Pattern Recognition. In traditional learning methods, Supervised-Learning is mainly used. Mostly used in the field where labeled samples are easy to get, it needs a large amount of labeled samples to learn from.In the field of remote sensing, it's difficult to get samples labeled, and incorrect information are easily brought in because of visual interpretation. In order to identify unlabeled samples as accurately as possible and to improve classification accuracy, we need to label a large number of samples and improve visual interpretation accuracy, which is very time-consuming and labor-intensive.In response to above problems, this paper introduces Semi-Supervised Learning, It only needs a small number of labeled samples , together with a large number of unlabeled samples, which greatly reduces time and labor cost. Based on Self-Training , Co-Training and Low Density Separation, first builds initial classifiers using labeled samples, then improve themselves using unlabeled samples recursively.A lot of experiments have been done using the same remote sensing images based on different kind of methods including supervision and semi-supervision, self-training and co-training, and Low Density Separation in this paper. Experimental results show that when the proportion of labeled sample to unlabeled sample is between 1:2 and 1:9, classification based on semi-supervised learning can get higher accuracy than that based on full-supervised learning using much less labeled sample, which has certain significance both on research and practice.