Non-Overlapping Sampling with Extreme Data Utilization in Hyperspectral Imagery

Data-driven hyperspectral classification has gained significant attention across various applications, leading to the development of numerous novel neural networks for effective classification. However, existing spatial-spectral works often overlook the critical aspect of how training and testing samples are split, resulting in data leaks between these sets and suboptimal performance in real-world scenarios. To address this issue, we propose a non-overlapping sampling approach based on the extremal animals theory to achieve extreme data utilization, i.e., maximizing the number of testing samples. Specifically, it begins by determining the number of training samples n in each connectivity area according to the label map and pre-defined sampling ratio. Subsequently, it searches for all possible polyominos with the minimum perimeter p(n) and at least n pixels intersecting with the connectivity area. Finally, we remove unnecessary pixels according to the priority of their class in the connectivity area and degree in polyominos. Experimental results on Indian Pines validate the superior effectiveness of our proposed non-overlap sampling strategy.