A hyperspectral evaluation approach for quantifying salt-induced weathering of sandstone.

Salt-induced weathering is a common phenomenon in stone relics, and its traditional artificial evaluation of severity is greatly affected by subjective consciousness and lacks systematic standards. Here, we propose a hyperspectral evaluation method for quantifying salt-induced weathering on sandstone surfaces in laboratory tests. Our novel approach consists of two parts: data acquisition of microscopic observations of sandstone in salt-induced weathering environments, and machine learning technology for a predictive model. We first obtain the microscopic morphology of sandstone surfaces by near-infrared hyperspectral imaging technique. Then, a salt-induced weathering reflectivity index is proposed according to analyses of spectral reflectance variation. Next, a principal components analysis-Kmeans (PCA-Kmeans) algorithm is applied to bridge the gaps between the salt-induced weathering degree and the associated hyperspectral images. Furthermore, machine learning technologies, such as Random Forest (RF), Support Vector Machine (SVM), Artificial Neural Network (ANN), and K-Nearest Neighbors (KNN), are trained for better evaluating the salt-induced weathering degree of sandstone. Tests demonstrate that the RF algorithm is feasible and active in weathering classification based on spectral data. The proposed evaluation approach is finally applied to the analysis of salt-induced weathering degree on Dazu Rock Carvings.