Urban-analysis-through-Satellite-Imagery

Abstract:

World population is increasing day by day. New areas are built in cities and already built areas are getting densely populated. Categorization of urban communities into planned or unplanned areas is very important for allocation of resources and planning of future projects. Using satellite images with the help of Deep Learning has been very beneficial in urban analysis. A lot of work has been done related to classification of buildings, segmentation and counting of different structures. In this work we want classify urban areas whether they are planned or unplanned, from satellite images. There are certain challenges for solving this problem, like extraction of a dataset, selecting useful features for classification and annotating data with these features. After dataset is collected, it has its own challenges and limitation; complex topologies of cities, shadows and other variations. After collecting dataset, we use U-net to detect images with more than 50% built-up areas. These images are then labeled with 7 different categories and classified through a multi-task learning (MTL) classifier. Resultant vector of MTL classifier is combined with original images to improve classification results and train the building count regressor. This way results are improved.