Hyperspectral reconnaissance has practical significance for target recognition due to its rich image and spectral information. However, abundant but complicated information makes hyperspectral camouflage segmentation improperly solved yet. To tackle this problem, we propose a self-adaptive two-stage framework derived from the morphology watershed, which is integrated with a pre-processing module and a fine-tuning module. Firstly, in the pre-processing stage, we propose a method that employs adaptive thresholding determined by 2D-OTSU (Two-dimensional OTSU) to mitigate unwanted gray-scale changes caused by variations in the lighting environment. Then, in the fine-tuning stage, for suppression of over-segmentation, we optimize marker points of the pre-processing stage via clustering centers. The empirical evaluation shows that our framework outperforms other methods in terms of segmentation accuracy and completeness. It also demonstrates that our framework owns good segmentation performance on hyperspectral images.