An ant interaction scheme based wrapper strategy for hyperspectral band selection

Hyperspectral imaging acquires information in an extensive range of narrow and contiguous spectral bands. However, information redundancy in neighboring bands directly affects classification performance. Substantial work has been proposed to select pertinent but fewer spectral bands by employing metaheuristic-based band selection strategies. Despite encouraging results, slow convergence and susceptibility to local optima remain significant challenges for metaheuristic techniques. Additionally, the selection of optimal hyperparameters for the underlying classifier significantly influences the algorithm’s efficacy. The present work suggests a novel wrapper technique to address the band selection problem. Two strategies are employed to select optimal bands and hyperparameter simultaneously. Firstly, a novel power distribution-based ant interaction and pheromone modeling technique is utilized for information sharing and efficiently balancing localized and globalized exploration of the search space. Secondly, to maintain diversity in ant tours, a random initialization strategy is applied to those ant tours that fail to improve over successive iterations. The proposed approach also has the advantage of requiring fewer parameters compared to existing continuous ant colony optimization algorithms, thereby significantly reducing the computational effort needed for fine-tuning. Experimental results demonstrate that the proposed strategy selects fewer spectral bands without compromising classification accuracy and outperforms other metaheuristic techniques.