AU-Super: Superpixel Scale Optimization and Training Data Augmentation Strategy for Hyperspectral Image Classification

Yiran Wang¹Lingling Li^2*Mingming Xu¹Shanwei Liu¹Muhammad Yasir^1*Manuel Á. Aguilar³Fernando J. Aguilar³

• ¹College of Oceanography and Space Informatics, China university of petroleum (East China), Qingdao , china, Qingdao, China
• ²Shandong Academy for Environmental Planning, Key Laboratory of Land and Sea Ecological Governance and Systematic Regulation, Ministry of Ecology and Environment, Jinan, China, jinan, China
• ³Universidad de Almeria, Almería, Spain

The spectral information of each pixel in hyperspectral images contains valuable information about object properties, although accurate labeling is required in supervised classification to guide the model in distinguishing different land cover types. However, labeling data for hyperspectral images is difficult to obtain, especially in complex or remote areas. This results in a shortage of labeled samples, which prevents the model from fully learning the features of different classes. To overcome this challenge, this work proposes a hyperspectral image classification method, called AU-Super, that combines adaptive superpixel scale selection, superpixel label expansion, and data augmentation. First, an adaptive method is developed to determine an appropriate superpixel segmentation scale based on feature values, thereby ensuring that superpixel segmentation effectively cap-tures the spatiospectral information of the image. Second, feature extraction is performed at the previously estimated superpixel scale. Third, pixel labels are converted to Running Title superpixel labels to reduce the effects of labeling noise during the training process. Furthermore, superpixel-level label-based data augmentation techniques are introduced to mitigate the problem of under-labeled patterns. The comparative results against various state-of-the-art algorithms demonstrate that AU-Super-RF consistently achieves superior performance across multiple accuracy metrics. Under few-shot training scenarios (with only 1–10 samples per class) on the Indian Pines, Salinas, and Pavia University datasets, it improves overall accuracy (OA) by 3%–7%, average accuracy (AA) by 2%–6%, and the Kappa coefficient by 3%–8%, highlighting the robustness and practical utility of the method.