Automatic land-cover classification using semi-supervised multilayer perceptron for analyzing remotely sensed images

In this article, a land-cover classification technique is proposed using semi-supervised Multilayer Perceptron (MLP) where only a few labeled samples are available to train the classifier. Here, the MLP initially trained using a few labeled samples is used to predict the class labels of the unlabeled samples to obtain soft class labels. The soft class labels of each unlabelled samples are boosted with respect to that of their k nearest neighbors. This is done to exploit the spatial relationship among the patterns. Finally re-training for MLP is carried out iteratively by using the available labeled samples along with the selected samples from the most confidently predicted ones. The retraining continues until the prediction values are significantly stable across iterations. Here a differential weighted mean of k neighborhood is used for boosting of soft class label and a semi-automatic threshold selection algorithm is investigated to identify the most confidently predicted samples. To assess the effectiveness of the proposed methodology, experiments on land-cover classification are conducted on patterns from Landsat8 satellite images for two different geographical regions. The obtained results outperform the corresponding supervised strategy under the scenario of labeled data scarcity.