Scene classification for aerial images based on CNN using sparse coding technique

• Main Authors: Qayyum, A., Malik, A.S., Saad, N.M., Iqbal, M., Faris Abdullah, M., Rasheed, W., Rashid Abdullah, T.A., Bin Jafaar, M.Y.
• Format: Article
• Published: Taylor and Francis Ltd. 2017
• Online Access: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85014553885&doi=10.1080%2f01431161.2017.1296206&partnerID=40&md5=c41a65077a8f7c4fbc7f59ab0a7955ef; http://eprints.utp.edu.my/19498/

Aerial scene classification purposes to automatically label aerial images with specific semantic categories. However, cataloguing presents a fundamental problem for high-resolution remote-sensing imagery (HRRS). Recent developments include several approaches and numerous algorithms address the task. This article proposes a convolutional neural network (CNN) approach that utilizes sparse coding for scene classification applicable for HRRS unmanned aerial vehicle (UAV) and satellite imagery. The article has two major sections: the first describes the extraction of dense multiscale features (multiple scales) from the last convolutional layer of a pre-trained CNN models; the second describes the encoding of extracted features into global image features via sparse coding to achieve scene classification. The authors compared experimental outcomes with existing techniques such as Scale-Invariant Feature Transform and demonstrated that features from pre-trained CNNs generalized well with HRRS datasets and were more expressive than low- and mid-level features, exhibiting an overall 90.3 accuracy rate for scene classification compared to 85.4 achieved by SIFT with sparse coding. Thus, the proposed CNN-based sparse coding approach obtained a robust performance that holds promising potential for future applications in satellite and UAV imaging. © 2017 Informa UK Limited, trading as Taylor & Francis Group.