Integrating the Mahalanobis Distance Metric with Spectral Clustering: A Hybrid Aerosol Classification Algorithm

Aerosols play a critical role in the Earth’s radiation budget by scattering and absorbing solar radiation; however, their classification remains a major source of uncertainty due to overlapping fine and coarse modes, complex mixing states, and strong spatio-temporal variability, particularly over mountainous terrain. This study presents a hybrid aerosol classification framework applied to long-term (2008–2025) sky–sun radiometer (SKYNET) observations from three high-altitude sites in the Ladakh region, together with global AERONET observations spanning 171 sites across six continents from 1993 to 2025. The algorithm is tested under both climatically sensitive high-altitude environments and diverse global conditions to evaluate its robustness and credibility. The approach integrates unsupervised spectral clustering with the statistical Mahalanobis distance (MD) metric to improve aerosol regime separation in high-dimensional feature space. The spectral clustering technique, an unsupervised data-driven approach, involves three main steps: constructing a similarity graph, projecting the data into a low-dimensional space, and forming clusters. Although spectral clustering partitions the entire dataset, real aerosol regimes typically exhibit a dense core of representative observations, with transitional or mixed cases occurring at the periphery. To reduce this overlap, the MD metric is introduced to retain only the core inliers.  Internal validation of the algorithm is performed using the Silhouette coefficient, Calinski–Harabasz index, and Davies–Bouldin index. A traditional threshold-based classification method is employed for external validation of the proposed framework. Using the hybrid algorithm, aerosols are classified into four types: Dust, Mixed, Absorbing, and Non-absorbing. Among the 171 sites analysed, 83 sites are dominated by Absorbing aerosols, 19 by Dust, 1 by Mixed, and 68 by Non-absorbing aerosol types. Africa is primarily dominated by dust aerosols, accounting for 50% of the sites. Absorbing aerosols dominate in Asia (67.3%), Australia (55.6%), and South America (77.3%). In contrast, Europe and North America are largely characterised by Non-absorbing aerosol types, representing 75.8% and 73.5% of the sites, respectively. A strong and statistically significant positive correlation (Pearson’s r = 0.89, p = 0.0166) is observed between the continent-wise dominant aerosol fractions derived from the threshold-based and hybrid classification methods. Individual continental comparisons reveal small deviations for Africa, Europe, and North America (<3%), identical results for Australia, and comparatively larger differences for Asia (+8.4%) and South America (−9.1%), suggesting an enhanced sensitivity of the hybrid approach in regions characterised by complex aerosol regimes. At high-altitude sites, low aerosol concentrations make the development of robust aerosol classification schemes particularly challenging. Nevertheless, major aerosol types—such as absorbing, non-absorbing, and mixed aerosols—can be effectively distinguished using spectral clustering algorithms, thereby enhancing the effectiveness of the proposed hybrid method.