Hyperspectral Band Selection for Zinc Concentrate Composition Prediction via Bayesian Nonparametric Segmentation and Multiobjective Optimization

Hyperspectral imaging (HSI) provides a rapid and non-destructive technique for compositional analysis of zinc concentrate; however, the intrinsic high dimensionality of spectral data often results in significant redundancy, complicating subsequent modeling and analysis. Clustering-based band selection (BS) methods have shown promise in reducing dimensionality while retaining informative bands, yet these approaches fail to maintain spectral contiguity, adequately capture inter-band similarity, and optimize the quality of selected subsets comprehensively. To overcome these challenges, we propose a Bayesian nonparametric segmentation framework integrated with evolutionary multi-objective optimization (BNS-MOP) for robust and efficient BS. Initially, kernel density estimation is adopted to characterize the reflectance distribution of each spectral band. Bayesian inference then quantifies inter-band similarity and groups adjacent bands within a maximum-posterior framework, thus preserving the inherent wavelength contiguity and physical interpretability. To ensure computational efficiency, the band-grouping problem is transformed into a dynamic-programming formulation and resolved by employing the Pruned Exact Linear Time (PELT) algorithm, whose overall computational complexity isO(n) with respect to the number of spectral bands. Subsequently, representative band selection is modeled as a multi-objective optimization problem, balancing redundancy minimization and information maximization, integrating prior knowledge of element-sensitive wavelengths. Moreover, an innovative integer encoding strategy is introduced to significantly reduce computational overhead relative to traditional binary encoding methods. Finally, the proposed BNS-MOP algorithm is validated against ten state-of-the-art band-selection techniques on hyperspectral data from zinc-concentrate samples. Across four target elements (Zn, Fe, Pb, and Cu), BNS-MOP consistently achieved superior performance, confirming the practical efficacy of the redundancy-aware, information-maximizing approach.