Enhancing the Classification of Remote Sensing Data Using Multiband Compact Texture Unit Descriptor and Deep Convolutional Neural Network

This abstract proposes a method to enhance the classification of high spatial remotely sensed imagery by using Multiband Compact Texture Unit (MBCTU) descriptor and pre-trained convolutional neural networks (CNN) feature extractor. MBCTU was used in order to take into account intra- and interband spatial interactions by characterizing texture using relative pixel values in a multispectral neighborhood instead of a monoband neighborhood. The derived new encoded images are fed to a deep feature extractor (fine-tuned CNN). Finally, the obtained deep features are classified using k nearest neighbor'S (KNN) algorithm to produce the classification map. The proposed approach is applied to the classification of hyperspectral and multispectral datasets. Results indicate that the proposed approach can achieve accurate classification results compared to other approaches using the full spectral dataset.