Assessing the performance of machine learning models at predicting demographics from satellite imagery

Abstract

Deep learning, a subset of machine learning, has seen enormous growth in recent years, enabling deep learning models to comprehend abstract details and understand context with high accuracy. Consequently, deep learning holds significant promise for a wide range of applications, particularly in GIS applications where population estimates demand accuracy and a wealth of abstract decision making on a huge scale. However, the sheer size of incoming data causes deep learning models to escalate computational cost and complexity which results in many prerequisites and a lengthy processing time, potentially making shallow machine learning models a more viable choice. With that in mind, the aim of this study is to compare a classical random forest machine learning model with a model that leverages convolutional neural networks (CNNs) for population estimation. We apply this approach by trying to predict the on-the-ground demographics of local areas using satellite imagery. For example, it is expensive in time and money to undertake detailed population surveys, therefore open-source datasets from reputable outlets such as WorldPop, commonly use such a predictive approach. Our research explores the potential for traditional shallow machine learning models to still hold precedence over newer deep learning applications given their superior logistical prerequisites while having comparable accuracy on curtain applications.