A deep learning approach to water point detection and mapping using street-level imagery

Abstract

ABSTRACT Households in developing countries often rely on alternative shared water sources that exist outside of the datasets of public service providers. This poses a significant challenge to accurately measuring the number of households outside the public service system that use a safe and accessible water source. The article proposed a novel deep learning approach that utilizes a convolutional neural network to detect and geo-reference communal water points using Google Street View imagery. Using a case study of the Agege local government area in Lagos, Nigeria, the model processed 39 kilometres of street network in 26 minutes, successfully detecting 36 previously unregistered water points with 94.7% precision and US$0 out-of-pocket expenses. In doing so, it presents a highly precise, low-cost, and scalable solution to closing geospatial data gaps on WASH access in developing countries.