Introduction to a cloud-based tool for on-demand urban expansion mapping in Africa: the DIY-BU Mapping tool

Abstract

The rapid urbanization of Africa, driven by natural population growth and rural-to-urban migration, significantly impacts the environment and presents challenges for urban management. Monitoring and understanding these dynamics requires data that are accurate, up-to-date, and analysis-ready. Currently, this information is only provided by global products that are not tailored to the African context with documented high uncertainties in classification accuracies. The DIY-BU-Mapping tool is an open, cloud-based mapping tool designed to generate local analysis-ready data on urban expansion in African cities. It is intended for use by decision-makers, stakeholders, and the scientific community. The tool addresses the growing need for objective, accurate, frequent, and timely data in developing urban environments. It utilizes Sentinel-1 and Sentinel-2 imagery along with various data sources. The tool is designed to assist users with varying skill levels in the classification process, from data collection and preparation to final map production. It consists of two parts: The first one creates a sampling dataset using reference data from 2021. For built-up areas, 'open buildings' from Google or 'building footprints' from Microsoft are used as reference datasets, while the WorldCover v200 from ESA is used for other land covers. The second part performs the classification, starting with dividing the sample data into training and validation (70/30). For each year starting from 2016, Sentinel images are collected, clouds are masked, and several spectral, textural, and statistical indices are calculated. These, along with a slope map, are used as features in a random forest classifier that is trained for 2021 and applied to each year. To ensure that built-up pixels are consistent over time, a pixel change trajectory analysis is applied. The accuracy of the method was compared with reference data as well as with globally available datasets, such as Dynamic World, WorldCover, ESRI land cover, WSF2019, and GHSL2023. The findings indicate that the locally fine-tuned maps produced by our tool outperform existing global multi-temporal products. Furthermore, we discuss both the limitations and strengths of the tool and the resulting maps. We release both the tool and its code freely and openly to encourage the scientific community to use the code, fostering the advancement of new Earth observation applications.