Object-Based Convolutional Neural Network for High-Resolution Imagery Classification

Abstract

Timely and accurate classification and interpretation of high-resolution images are very important for urban planning and disaster rescue. However, as spatial resolution gets finer, it is increasingly difficult to recognize complex patterns in high-resolution remote sensing images. Deep learning offers an efficient strategy to fill the gap between complex image patterns and their semantic labels. However, due to the hierarchical abstract nature of deep learning methods, it is difficult to capture the precise outline of different objects at the pixel level. To further reduce this problem, we propose an object-based deep learning method to accurately classify the high-resolution imagery without intensive human involvement. In this study, high-resolution images were used to accurately classify three different urban scenes: Beijing (China), Pavia (Italy), and Vaihingen (Germany). The proposed method is built on a combination of a deep feature learning strategy and an object-based classification for the interpretation of high-resolution images. Specifically, high-level feature representations extracted through the convolutional neural networks framework have been systematically investigated over five different layer configurations. Furthermore, to improve the classification accuracy, an object-based classification method also has been integrated with the deep learning strategy for more efficient image classification. Experimental results indicate that with the combination of deep learning and object-based classification, it is possible to discriminate different building types in Beijing Scene, such as commercial buildings and residential buildings with classification accuracies above 90%.