High spatial-resolution classification of urban surfaces using a deep learning method

Abstract Urban surface composition is key information for global carbon emission estimation, mesoscale numerical simulations, and outdoor environment studies at both city and neighborhood scales. High spatial-resolution satellite imagery can provide accurate urban surface data. Due to the large volume of data, it is time-consuming to extract manually information from satellite imagery or use other conventional classification methods. Deep leaning-based classification methods have great potential in satellite imaginary recognition due to their high efficiency and accuracy. In this study, we established a novel workflow for deep learning-based satellite imagery classification. At the stage of training dataset preparation, an object-based image analysis (OBIA) classification with the assistance of open source ancillary data from OpenStreetMap (OSM), i.e., OBIA-OSM method, was proposed to accelerate the procedure and improve the accuracy. An area of 33.5 km2 training dataset with 1 m spatial resolution was built based on the Gaofen2 satellite imagery of Hangzhou, China (GF2-HZ dataset). A Res-UNet + inception model was developed for the deep learning process. The results using the newly built model were compared with those from previous FCN and UNet models. The overall accuracy of our proposed model on the GF2-HZ dataset reached 83.1% which outperforms previous models. The influences of raw data-related factors, such as spectral bands composition and spatial resolution, were carefully tested. Data normalization and ‘transfer learning’ techniques were also analyzed and applied to improve the generalization ability of deep learning models.