A novel ensemble learning approach to extract urban impervious surface based on machine learning algorithms using SAR and optical data

• Published in: International journal of applied earth observation and geoinformation Vol. 132; p. 104013
• Main Authors: Ahmad, Muhammad Nasar, Shao, Zhenfeng, Xiao, Xiongwu, Fu, Peng, Javed, Akib, Ara, Iffat
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.08.2024; Elsevier
• Subjects: data collection; Ensemble Learning; Impervious surface; land cover; Machine Learning; radar; Remote Sensing; SAR; spatial data; Ensemble Learning; Remote Sensing; Impervious surface; SAR; Machine Learning
• ISSN: 1569-8432; 1872-826X
• DOI: 10.1016/j.jag.2024.104013

•Proposed a novel ensemble learning framework to improve urban impervious surface accuracy.•Integrated open source Optical and SAR datasets and four ensemble algorithms.•UISEM outperformed three existing global data products with 92% accuracy.•XGB algorithm is optimal for mapping UIS for diverse global cities. Accurate urban impervious surface (UIS) extraction from open-source remote sensing data remains challenging, especially for cities with heterogeneous climatic backgrounds. Contemporary, state-of-the-art techniques achieve promising results at a global scale, but accuracy is compromised at the city level. Therefore, a ensemble machine learning approach using open-source Optical-SAR remote sensing datasets was implemented to enhance the accuracy of UIS mapping. Initially, we integrated optical and radar datasets with modified urban indices to generate input features. Then, we applied four ensemble machine learning algorithms, including AdaBoost, Gradient Boost (GB), Extreme Gradient Boosting (XGBoost), and Random Forest (RF), and fine-tuned them via a soft voting ensemble approach. The optimized UISEM approach showed a model accuracy of 98%. The UISEM method achieved a classification accuracy of 92% and consistently performed across 32 cities globally with heterogeneous climatic zones. Regarding accuracy and predictive power, the XGB ensemble classifier outperformed other ML classifiers in mapping UIS. Furthermore, a comparative analysis against three well-known datasets (ESA World Cover, ESRI Land Cover, and Dynamic World) was also performed. The proposed UISEM model outperformed renowned global datasets with a 92% classification accuracy, followed by DW with 83%, ESA with 86%, and ESRI with 82%. In the future, developing a spatial–temporal version of UISEM can support diverse urban applications globally. The datasets and (GEE and Python) codes are available at https://github.com/mnasarahmad/UISEM.