ENHANCED PIXEL BASED URBAN AREA CLASSIFICATION OF SATELLITE IMAGES USING CONVOLUTIONAL NEURAL NETWORK

Abstract

Recent years have witnessed a great development in the use of deep learning in the applied fields in general, including the improvement of remote sensing. Satellite imagery classification has played a prominent role in various development processes. This paper presents a new improvement in automatic urban classification using One Dimension Convolutional Neural Network (1DCNN) architecture. The suggested approach has three enhancement processes. First, select training boxes for different classes and create many pixels with variable class signatures. This makes the training process dependent on the broadband of signature for the classes. Second, modified 1D convolution was used to re-encode pixel values to increase distinguish power. Third, adding a new median filter layer at the end of network architecture to remove pixels like noise to make the resulting map smoother. An image of Greater Cairo is used and the different urban classes are defined within it. The proposed method was compared to other methods based on pixels. The proposed method proved to be numerically and visually superior.