Modified Semi-Supervised Adversarial Deep Network and Classifier Combination for Segmentation of Satellite Images

Content extraction from satellite images continues to evolve with the application of learning aided approaches. Recently, with the addition of deep learning (DL) based methods, content extraction from satellite images has become more reliable and efficient, yet challenges continue to exist as these methods require a large number of training and annotated images to enable effective learning by these networks. For high-resolution satellite images, limited training data is a familiar problem. Therefore, amongst the DL-based methods, semi-supervised adversarial approaches represent an emerging area of application in content extraction from satellite images. Semi-supervised adversarial methods adopt a combination of unsupervised training and labeled data to process applied inputs to generate reliable classification. In this paper, a semi-supervised adversarial learning method, which includes architectural expansion and several other additions, is reported that is used for content extraction from satellite images. The objective of the work is to design a modified structure of a semi-adversarial network working in concert with a classifier layer to extract the region of interests (ROIs) of the satellite image with limited training and annotated data. The proposed method has been tested with four different input feeding mechanisms which enhance the quality of processed data by generating a higher correlation. Two learning-based networks constitute the core of the semi-supervised adversarial learning method reported in this work. The first is a segmentation network that combines unlabeled data and supervised learning for processing. Next is a discriminator block which is a variant of the popular Convolutional Neural Network (CNN) trained sufficiently to improve the segmentation accuracy. For all the experiments performed, both labeled and unlabeled cases are considered. Adversarial and semi-supervised losses are the cost functions to train the system with images of the DeepGlobe Land Cover Classification Challenge dataset. The outputs of the segmentation network which are the semantic labels of the satellite images are used as inputs to the classifier to extract the ROIs of the satellite image. A comparison of the performance of the classifiers used is also included for ascertaining the most suitable one for the specific combination for the discriminator-segmentation blocks. Experimental data reveal that the proposed work is reliable compared to earlier reported approaches.