Mapping Asbestos-Cement Corrugated Roofing Tiles with Imagery Cube via Machine Learning in Taiwan

Abstract

Locating and calculating the number of asbestos-cement corrugated roofing tiles is the first step in the demolition process. In this work, archived image cubes of Taiwan served as the fundamental data source used via machine learning approach to identify the existence of asbestos-cement corrugated roofing tiles with more than 85% accuracy. An adequate quantity of ground-truth data covering all the types of roofs via aerial hyperspectral scan was the key to success for this study. Twenty randomly picked samples from the ground-truth group were examined by X-ray refraction detection to ensure correct identification of asbestos-cement corrugated roofing tiles with remote sensing. To improve the classifying accuracy ratio, two different machine learning algorithms were applied to gather the target layers individually using the same universal training model established from 400 ground-truth samples. The agreement portions within the overlapping layers of these two approaches were labeled as the potential targets, and the pixel growth technique was performed to detect the roofing boundary and create the polygon layer with size information. Exacting images from aerial photos within the chosen polygon were compared to up-to-date Sentinel-1 images to find the temporal disagreements and remove the mismatched buildings, identified as non-asbestos roofs, from the database to reflect the actual condition of present data. This automatic matching could be easily performed by machine learning to resolve the information lag while using archived data, which is an essential issue when detecting targets with non-simultaneous acquired images over a large area. To meet the 85% kappa accuracy requirement, the recurring processes were applied to find the optimal parameters of the machine learning model. Meanwhile, this study found that the support vector machine method was easier to handle, and the convolution neuro network method offered better accuracy in automatic classification with a universal training model for vast areas. This work demonstrated a feasible approach using low-cost and low-resolution archived images to automatically detect the existence of asbestos-cement corrugated roofing tiles over large regions. The entire work was completed within 16 months for an area of 36,000 km2, and the detected number of asbestos-cement corrugated roofing tiles was more than three times the initial estimation by statistics method from two small-area field surveys.