A Classification Method of Local Climate Zones With the Skip-Connection Variable-Scale Network and a Case Study in the Main Urban Area of Xi'an, China

Local climate zones (LCZ) provide a refined tool and research framework for the study of urban heat island effects. Remote sensing images, with high coverage and easy accessibility, are widely used in LCZ mapping. However, since the acquisition of traditional LCZ samples depends on expert knowledge and incurs a high cost of annotation, urban LCZ mapping are facing difficulties in sample generation and accuracy increasing. In this work, a skip-connection variablescale network (SVS-Net) was proposed for LCZ classification with remote sensing images, which was used in the main urban areas of Xi'an. Experiments demonstrated the proposed method could achieve the LCZ map of Xi'an efficiently and accurately. In using the SVS-Net based LCZ classification method, a training sample number of more than 500 was suggested, which would running with an overall accuracy more than 0.8928. The experiments also suggested a structure with the best performance that combined with two skip convolutional layers and one fully connected layer.