Feature extraction and classification of VHR images with attribute profiles and convolutional neural networks

Effective feature extraction plays an important role in the classification of very high resolution (VHR) remote sensing (RS) images. Current researches mainly focus on individual shallow or deep feature extraction methods, remarkable representatives of which include Morphological Attribute Profiles (APs) and Convolutional Neural Networks (CNNs). Actually, to combine low-level and high-level features may take advantages of each approach and fully exploit the description capability. In this paper, APs and CNNs are integrated to characterize VHR RS images in order to improve the pixel description. Moreover, during the training of CNNs, regularization, dropout and fine-tuning strategies are all utilized to mitigate over-fitting problems due to insufficient samples in RS applications. Evaluations using QuickBird datasets demonstrate that our proposed method leads to a higher classification accuracy compared to individual method for VHR images.