Automated Remote Sensing Image Interpretation with Limited Labeled Training Data
Automated remote sensing image interpretation has been investigated for more than a decade. In early years, most work was based on the assumption that there are sufficient labeled samples to be used for training. However, ground-truth collection is a very tedious and time-consuming task and sometimes very expensive, especially in the field of remote sensing that usually relies on field surveys to collect ground truth. In recent years, as the development of advanced machine learning techniques, remote sensing image interpretation with limited ground-truth has caught the attention of researchers in the fields of both remote sensing and computer science. Three approaches that focus on different aspects of the interpretation process, i.e., feature extraction, classification, and segmentation, are proposed to deal with the limited ground truth problem. First, feature extraction techniques, which usually serve as a pre-processing step for remote sensing image classification are explored. Instead of only focusing on feature extraction, a joint feature extraction and classification framework is proposed based on ensemble local manifold learning. Second, classifiers in the case of limited labeled training data are investigated, and an enhanced ensemble learning method that outperforms state-of-the-art classification methods is proposed. Third, image segmentation techniques are investigated, with the aid of unlabeled samples and spatial information. A semi-supervised self-training method is proposed, which is capable of expanding the number of training samples by its own and hence improving classification performance iteratively. Experiments show that the proposed approaches outperform state-of-the-art techniques in terms of classification accuracy on benchmark remote sensing datasets.