Urban Local Climate Zone Classification with a Residual Convolutional Neural Network and Multi-Seasonal Sentinel-2 images

Abstract

This study presents a classification framework for the urban Local Climate Zones (LCZs) based on a Residual Convolutional Neural Network (ResNet) architecture. In order to make full use of the temporal and spectral information contained in modern Earth observation data, multi-seasonal Sentinel-2 images are exploited. After training the ResNet, independent predictions are made from the multi-seasonal images. Subsequently, the seasonal predictions are fused in a decision fusion step based on majority voting. A systematical experiment is carried out in a large-scale study area located in the center of Europe. A significant accuracy improvement can be achieved by applying majority voting on multi-seasonal predictions. Based on the results, the main challenges and possible solutions of urban LCZ classification are further discussed, providing guidance for large-scale urban LCZ mapping.