Abstract：Aiming at the issues of strong randomization in the initial population, imbalance between global convergence and local diversity, and low local search efficiency of the Non-dominated Sorting Genetic Algorithm Ⅲ (NSGA-Ⅲ) for hyperspectral band selection, an improved algorithm—INSGA-Ⅲ (Improved NSGA-Ⅲ driven by feature classification)—is proposed. Firstly, Latin Hypercube Sampling and a reference-point guidance mechanism were integrated to generate a high-quality initial population that ensures both comprehensive search space coverage and targeted focus in the objective space. Secondly, a classification accuracy-driven term based on the Adaptive Rotating Forest algorithm and a correlation penalty term based on the Pearson correlation coefficient were designed to construct a multi-objective fitness function that balances global exploration and local exploitation. Finally, the search mechanism of Particle Swarm Optimization was introduced to enhance regional search efficiency. Experiments are conducted on four types of hyperspectral datasets: Indian Pines (agricultural scenes), Pavia University (urban features), Salinas (vegetation monitoring), and Botswana (mineral identification). Widely used algorithms, including Sequential Forward Selection (SFS), Competitive Adaptive Reweighted Sampling (CARS), Multi-Objective Particle Swarm Optimization (MOPSO), Multi-Objective Evolutionary Algorithm Based on Decomposition (MOEA/D), and the original NSGA-Ⅲ, are selected as benchmarks to verify the universal advantages of INSGA-Ⅲ. Experimental results show that, in terms of band selection performance, INSGA-Ⅲ improves information entropy by 8.5% and reduces the band correlation metric by 9.7%, compared to the mean values of all benchmark algorithms (p<0.01). In the SVM classification task, INSGA-Ⅲ outperforms the benchmark mean values by 10.3% in Overall Accuracy (OA) and 11.6% in Kappa coefficient (p<0.01). Regarding algorithmic efficiency, INSGA-Ⅲ requires 32% fewer iterations to reach 90% Pareto front approximation than NSGA-Ⅲ, and shows significantly lower accuracy fluctuation (standard deviation ±1.23%) than the benchmark mean (±4.2%) under 25% Gaussian noise (averaged over 10 runs). The proposed algorithm provides an efficient and robust band selection scheme for applications such as agricultural crop monitoring, urban feature classification, and mineral identification, effectively balancing information content, redundancy, and classification accuracy, while significantly reducing the dimensionality and processing cost of hyperspectral data.
Key words：Non-dominated sorting genetic algorithm Ⅲ; Multi-objective hyperspectral band selection; Latin hypercube sampling; Adaptive rotating forest algorithm; Particle swarm optimization
[1] Weiss M, Jacob F, Duveiller G. Remote Sensing of Environment, 2020, 236: 111402.
[2] Zhao J, Wang J, Ruan C, et al. IEEE Transactions on Geoscience and Remote Sensing, 2024, 62: 1.
[3] Sun L, Zhao G, Zheng Y, et al. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 1.
[4] Maseno E M, Wang Z. Journal of Big Data，2024，11(1): 24.
[5] Wu K, Zhu T, Wang Z, et al. European Food Research and Technology，2024，250(1): 191.
[6] Vallese F D, Paoloni S G, Springer V, et al. Journal of Food Composition and Analysis, 2024, 126: 105925.
[7] Xu X F, Wang K, Ma W H, et al. Renewable Energy, 2024, 223: 120086.
[8] He M, Wang Z, Chen H, et al. Expert Systems, 2025, 42(2): e13802.
[9] Wang Y, Zhu Q, Ma H, et al. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 1.
[10] Zhang Y, Lin Q, Li L, et al. Swarm and Evolutionary Computation, 2024, 89: 101614.
[11] Wang Q, Liu Y, Xu K, et al. Swarm and Evolutionary Computation, 2024, 90: 101665.
[12] Wei Y, Hu H, Xu H, et al. Sensors, 2023, 23(4): 2129.
[13] Phaneendra Kumar B L N, Vaddi R, Manoharan P, et al. Scientific Reports, 2024, 14(1): 31836.
[14] Gu Q, Xu Q, Li X. Expert Systems with Applications, 2022, 207: 117738.
[15] Sawant S S, Manoharan P. International Journal of Remote Sensing, 2019, 40(20): 7852.
[16] Naik P, Chakraborty R, Thiele S, et al. Mult-Objective Optimization Based Hyperspectral Feature Engineering for Spectral Abundance Mapping, 2024 14th Workshop on Hyperspectral Imaging and Signal Processing: Evolution in Remote Sensing (WHISPERS)，IEEE, 2024.
[17] Iordanis I, Koukouvinos C, Silou I. Applied Numerical Mathematics, 2025, 208: 256.
[18] Feng W, Quan Y, Dauphin G, et al. Information Sciences, 2021, 575: 611.
[19] Shami T M, El-Saleh A A, Alswaitti M, et al. IEEE Access, 2022, 10: 10031.
