An object-based supervised classification framework for very-high-resolution remote sensing images using convolutional neural networks

Object-based image classification (OBIC) is presented to overcome the drawbacks of pixel-based image classification (PBIC) when very-high-resolution (VHR) imagery is classified. However, most of classification methods in OBIC are dealing with 1D hand-crafted features extracted from segmented image objects (superpixels). To extract 2D deep features of superpixels, a new deep OBIC framework is introduced in this letter by using convolutional neural networks (CNNs). We first analyze the different mask policies of superpixels and design two architectures of networks. Then, we determine the specific details of our framework before experiments. The results of comparison experiments show that our DiCNN-4 (Double-input CNN) model achieves higher overall accuracy, κ coefficient and F-measure than conventional OBIC methods on our image dataset.