Counting Buildings from Unmanned Aerial Vehicle Images using a Deep Learning Based Approach

Abstract

Effective urban planning requires accurate and up-to-date spatial information. The advent of remote sensing has contributed immensely to the efficiency with which this information can be collated. With remotely sensed high spatial resolution imagery important matrixes such as number of buildings in an area can be manually counted or building densities quantified and mapped. On the other hand, traditional methods are tedious and subjective. The aim of this study therefore was to develop deep learning-based algorithms to automate the counting of buildings from high spatial resolution aerial imagery. Unmanned Aerial Vehicle (UAV) sensed imagery was acquired, preprocessed, annotated and augmented using python scripts. A deep learning algorithm based on Convoluted Neural Networks was then developed in the You Only Look Once (YOLO) V2 environment to identify and then count the buildings in the area of interest. The algorithm was trained and tested on the google co-laboratory platform. The model achieved a high accuracy with a recall rate of 0.89, F1 score of 0.89, and average precision of 91.12% on the validation data. When applied to a new set of testing data, the algorithm successfully identified and counted the number of buildings with an overall accuracy of 71%. The developed methodology yields a lot of promise to accurately extract building matrixes for change detection, biomass modelling and urban planning.