Active Learning Based Methods For Remote Sensing Image Classification

Supervised models have been popularly used in remote sensing data classification. The generalization capacity of these models always depend on the quality and quantity of training samples. With the advance of data collection and storage techniques, it is easy to accumulate a large amount of unlabeled samples in practical application. However, labeling such a large amount of samples can be laborious, time consuming and prone to human error. So how to exploit the abundant unlabeled samples and train a strong enough classifier with a small amount of available labeled samples has become a hot topic in the past few years. Active learning is such an approach to address this problem.Current active learning methods are essentially based on pixel-wise classification using the spectral feature of every pixel without spatial information, reducing to a large amount of isolate points in the classification map. To solve this problem, in this paper, we propose two active learning methods which combine the patch information with pixel information. The first one is the patch based active learning method. Firstly, the original remote sensing images are partitioned into overlapping patches. Then, for each patch, the spectral and spatial information are extracted. Lastly, three active learning methods, including margin sampling, entropy query-by-bagging and multi-class level uncertainty, are used to classify the patches. Experimental results show better performance of proposed patch based active learning model on the classification accuracy and the computational time on three different hyper-spectral data sets as compared to pixel based active learning model. The second one is a new committee based active learning approach. Firstly, the original remote sensing images are partitioned into overlapping patches and the feature of every patch is extracted. After that, a small number of patches are randomly selected to train the patch based support vector machine (SVM). Meanwhile, pixels contained in these patches are used to train the pixel based SVM. A set of unlabeled pixels from the pixel candidate set whose prediction labels disagree by the two classifiers are added to the contention pool. Lastly, a margin sampling based active learning method is employed to select the most informative pixels from contention pool. These pixels are labeled by human annotator and added to the training set to retrain the two SVM models. This process will repeat until a predefined convergence condition is satisfied. Experimental results show excellent performance on two hyper-spectral remote sensing data sets as compared to the state-of-the-art margin sampling and entropy query-by-bagging based active learning models.