Asbestos roofing recognition by use of convolutional neural networks and high-resolution aerial imagery. Testing different scenarios

Abstract

Asbestos has been widely used in the economy due to the mineral's unique physical and chemical properties. At the beginning of the 21st century, research confirmed that asbestos is a carcinogen. In 1997, in Poland it was forbidden to produce and use products containing asbestos. Statistics on the import of asbestos to Poland were collected, but unfortunately, there are no statistics on the quantity of asbestos products in use. Over 90% of still used products containing asbestos are asbestos-cement tiles. Therefore, various methods of estimating the number of these roofs are being sought to eliminate them safely from the use by the end of 2032. Our previous study has used CNN to test asbestos cement roofs identification in one commune. The purpose of the study is to present the possibilities of using the new artificial neural network architecture enhanced by inception-net which was developed to identify asbestos cement roofs based on high-resolution aerial images. In addition to our previous study, another commune was added to the research area. This study was conducted in two communes in Poland: Chęciny and Baranów. The study used orthophotos with a spatial resolution of 25 cm. Information on asbestos cement roofs was obtained during fieldwork. A new architecture of convolutional neural network (CNN) with a feature extraction block based on inception-net was used, which was not tested in our previous study. The classification was performed with the use of aerial imagery in the RGB composition since the results obtained before were promising. In this research, four different classification scenarios were tested: (1) the new network architecture with an inception-net-based network was trained and validated on signatures obtained for the same commune as in our previously published results (Krówczyńska et al., 2020) [18]; (2) the new network was trained and validated on signatures obtained for another commune in Poland; (3) the new network architecture with an inception-net-based network was trained on a dataset that combined both image signatures from two communes, from which training and validation signatures were selected randomly; (4) the new network architecture testing image signatures derived from one commune in another which was not seen by CNN. Moreover, the high-resolution imagery in those two communes was taken under different conditions. The choice of final model selection was done on the basis of the validation dataset loss function value (lowest is the best). The overall accuracy of the classification of different scenarios tested ranges from 88.0% to 93.0%. The presented research results indicate that there is a possibility to map asbestos cement roofs with high accuracy using the high-resolution aerial imagery taken under different conditions and CNN with inception-net using image signatures from different areas. This method may be used for decision-making communities to estimate the amount of asbestos cement roofing still in use to issue policies that will enable to safely eliminate those carcinogenic building materials from the environment.