Identification of asbestos-cement in NIR images via spectral similarity and artificial neural networks

This study employs airborne multispectral imagery obtained from the repository of the Environmental Public Agency of Cartagena (EPA), which provides a spatial resolution of 0.17 m/pixel. The dataset consists of four spectral bands (R, G, B, and NIR), from which 300 spectral signatures of asbestos–cement roofing and 300 signatures of other materials were extracted. The labeling was conducted through visual interpretation supported by field visits. Two configurations of the data cube were evaluated: one containing the four multispectral bands and another with five bands, generated by incorporating the first principal component obtained through PCA. Three spectral similarity measures (Cosine-Based Similarity (CBS), Spectral Distance Similarity (SDS), and Euclidean Distance Similarity (EDS)) were compared along with an artificial neural network (ANN) model. The performance was assessed using a 70–30 split based on precision, recall, F1-score, and execution time. The results indicate that EDS and CBS achieved the highest detection accuracy, with EDS being computationally more efficient (CBS was 2.22 times slower). These findings demonstrate that EDS and CBS are promising methods for the efficient detection of asbestos–cement roofing from multispectral imagery and can be extrapolated to the identification of other materials under similar spatial-resolution and data-quality conditions.