Assessing Human Settlement Sprawl in Mexico via Remote Sensing and Deep Learning

Abstract

Understanding human settlements' geographic location and extent can support decision-making in resource distribution, urban growth policies, and natural resource protection. This research presents an approach to assess human settlement sprawl using labeled multispectral satellite image patches and Convolutional Neural Networks (CNN). By training deep learning classifiers with a dataset of 5,359,442 records consisting of satellite images and census data from 2010, we evaluate sprawl for settlements across the country. The study focuses on major cities in Mexico, comparing ground truth results for 2015 and 2020. EfficientNet-B7 achieved the best performance among various CNN architectures evaluated with an ROC AUC of 0.970 and a PR AUC of 0.972. To assess human settlement sprawl, we introduce an information-based metric that offers advantages over entropy-based alternatives. We demonstrate its application to major cities in Mexico as examples.