Automatic Detection and Counting of Urban Housing and Settlement in Depok City, Indonesia: An Object-Based Deep Learning Model on Optical Satellite Imageries and Points of Interests

Abstract

Detecting urban housing and settlements has a substantial position in decision-
making problems such as monitoring housing and development, not to mention the widely
required urban mapping application. One of the most important goals in the United Nations
Sustainable Development Goals (SDGs) is to improve urban living conditions globally by
2030. We propose an automatic detection of urban housing and settlements on remote sensing
satellite imagery data using object detection-based deep learning using semantic segmentation
and the potential availability of remote sensing datasets at high spatial resolutions, Open Street
Map (OSM) geolocation point of interest dataset, and Sentinel-2 optical satellite imagery data.
The detection model using Mask Region-based Convolutional Neural Networks (Mask R-
CNN) is implemented in Depok City, Indonesia. These regions were chosen because it is the
second most populous suburb in Indonesia and the tenth most populous globally and, making it
challenging to extract building features from satellite imagery. This model categorizes dense,
moderate, and sparse conditions and has a promising result of an average precision of 100%
and an F1-score of 67% with evaluation performance metrics only considering points
associated with buildings, not building boundaries or the intersection over union (IoU). The
model performance has been compared to ground check results of field surveys, and it
performs best in sparse conditions. Our findings offer the potential implementation of the
model for fast and accurate monitoring of housing, settlement, and regional planning in urban
areas.