Identifying degrees of built arrangement in Indian cities through mid-resolution satellite imagery and Convolutional Neural Networks

Abstract

The performance of Convolutional Neural Networks (CNN) in satellite image classification tasks has been found superior to that of traditional algorithms. However, comparatively fewer studies have experimented with CNN-based classifiers to classify intra-urban characteristics with open mid and low-resolution earth observation (EO) datasets. The current pace of urbanization necessitates understanding and mapping of inherent urban forms, which would further assist in devising policies pertaining to sustainable urban development. While several remote sensing methods have been utilized to understand the urban structure, the replicability and generalizability of such approaches have been some of the key limitations. This study creates the CNN model to identify the degrees of built arrangement in mid resolution Sentinel 2B imagery of ten largest Indian cities. Training and testing datasets for seven land cover classes such as compact, open, sparse built, paved, unpaved, greens, and water are manually created with the help of Google Earth Pro platform. The definitions of the classes are acquired from the LCZ classification scheme. The CNN model trained with the prepared dataset provides the overall accuracy of 90% and kappa value of 0.88. The classification results are plotted for each city and compared with each other. The study has potential in the large-scale assessment of built forms of cities for quick assessment and policy formulation.