MOBS-TD: Multiobjective Band Selection With Ideal Solution Optimization Strategy for Hyperspectral Target Detection
Xudong Sun,Peng Lin,Xiaodi Shang,Huijuan Pang,Xianping Fu
DOI: https://doi.org/10.1109/jstars.2024.3402381
IF: 4.715
2024-06-15
IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Abstract:Band selection (BS) is a crucial concept within the realm of remote sensing, involving the selection of the most suitable bands to accurately capture features of landforms and surfaces. Despite the promising results achieved by many existing methods, certain limitations remain. First, most methods rely on a single criterion for band evaluation, leading to an incomplete assessment and limited generalizability of bands. Second, there is a lack of emphasis on target detection; thus, some BS techniques commonly used for classification are less effective for detection. Therefore, this article proposes MOBS-TD, a multiobjective optimization (MO) based BS method specifically designed for target detection, which aims to select bands with better target separation and stronger robustness across various application scenes. Initially, we develop an MO model with three objectives and introduce a novel metric to quantify the target–background separability of bands. Subsequently, a weighted similarity to ideal solution strategy is developed to clearly describe the dominance relations and strike a balance among multiple objectives in evolution. In addition, we devise an evaluation mechanism based on the ratio of maximum to submaximum, which is devised for selecting the optimal solution from the Pareto front, which has been empirically validated to be effective in reducing false alarms. Extensive experiments on real-world datasets demonstrate the competitiveness of MOBS-TD in remote sensing applications.
engineering, electrical & electronic,imaging science & photographic technology,remote sensing,geography, physical