IJASEIT

International Journal on Advanced Science, Engineering and Information Technology, Vol. 11 (2021) No. 5, pages: 1897-1902, DOI:10.18517/ijaseit.11.5.13102

Utilization of Band Combination for Feature Selection in Machine Learning-Based Roof Material Types Identification

Ayom Widipaminto, Yohanes Fridolin Hestrio, Donna Monica, Yuvita Dian Safitri, Dedi Irawadi, - Rokhmatuloh, Djoko Triyono, Erna Sri Adiningsih

Abstract

Land monitoring requires remote sensing data, which varies in its spectral and spatial resolution. Remote sensing data with the high spatial resolution is especially needed for urban monitoring. However, high spatial resolution data is usually expensive with limited coverage and complex analysis. This paper aims to find the most efficient way to do urban monitoring, specifically surface material identification. In material identification, the distinctive feature that can be used to differentiate one material surface from one another is its reflectance responses. This leads to a question of which absorption features are significant to different surface materials, especially roofing materials, and which absorption features are not discriminant enough to be used at classification. This paper proposed a machine learning-based identification of roof material types using band combinations as classification features. The experiment was done on Pleiades data, multispectral satellite imagery with very high spatial resolution. We first calculated the image’s reflectance values for each band and then grouped them based on their spectral range, yielding 11 possible combinations as the classification features. The experiment found that reflectance responses for band Red and NIR are the most distinctive trait of a material type and thus sufficient for material identification. We minimized the number of spectral responses used in material identification down to two bands, which can help the data collection and processing of material identification easier, cheaper, and less time-cost. Our experiment yields overall accuracy of 0.9959, with a computational time of 19.72 seconds.