Analysis of asbestos-cement roof classification in urban areas: Supervised and unsupervised methods with multispectral and hyperspectral remote sensing

Abstract

Asbestos-cement roofs, commonly found in urban areas, pose environmental and health risks as they deteriorate, releasing asbestos fibres into the atmosphere. Accurate identification and classification of these roofs are essential for assessing potential hazards and implementing appropriate remediation measures. This study presents a comprehensive analysis of supervised and unsupervised classification methods for the identification of asbestos-cement roofs in an urban area using both multispectral and hyperspectral remote sensing data. Six well-established supervised classification methods and two unsupervised classification methods were employed to analyse multispectral (WorldView 3 satellite) and hyperspectral data (overflight), offering ground pixel resolutions of 3.7 m and 1.2 m for both images. ENVI® was utilized for classification purposes. The supervised methods included in the study were Parallelepiped (PP), Minimum Distance (MiD), Mahalanobis Distance (MhD), Spectral Angle Mapper (SAM), Support Vector Machine (SVM) and Spectral Information Divergence (SID). In contrast, unsupervised methods were K-Means and ISO-Data. The classification performance of each method was assessed based on several metrics. The novelty of this study lies in the first-ever comparison of six supervised and two unsupervised methods applied to hyperspectral imagery captured via aerial survey and satellite imagery over the same urban area. Results indicate that hyperspectral data outperformed multispectral data in terms of asbestos-cement roof classification, demonstrating the potential of hyperspectral imagery for more precise identification. Additionally, the supervised classifiers consistently outperformed the unsupervised methods, highlighting the importance of a priori knowledge for accurate classification. In contrast, the cost-benefit analysis reveals that multispectral imagery is significantly more cost-efficient, being up to 6.5 times less expensive and requiring approximately 32 times fewer computational resources than hyperspectral imagery. This study provides important insights for urban planning, environmental assessment, and public health management by enabling accurate and efficient identification of asbestos-cement roofs in urban areas. The findings highlight the critical role of selecting appropriate remote sensing data and classification techniques for such applications. The methodology and results offer valuable guidance to local authorities, researchers, and policymakers in addressing asbestos-related risks, particularly in developing countries confronting these challenges.