Abstract:Changbaishan Mountain, one of the most well- preserved polygenetic composite active volcanoes, boasts a diverse range of surface cover types, making it an ideal location for studying volcanic eruption landforms. The clear interpretation labels in remote sensing imagery make this region a valuable resource for large- scale volcanic landform classification using deep learning methods. This paper introduces the volcanic eruption landform scene dataset (VELSD1. 0), derived from Sentinel- 2 remote sensing imagery of Changbaishan Mountain, China. The dataset was developed in conjunction with geological data and field investigations. In the dataset, VELSD1. 0 comprises ten surface cover types: high vegetation coverage, middle vegetation coverage, low vegetation coverage, exposed soil, basaltic rocks, trachytic rocks, mixed rocks of trachyte and pumice, water, shadow, and artificial landscape. The dataset includes three constituent elements: remote sensing images, labeled data, and an interpreted label and explanatory file. Covering an area of 2500 km2, the dataset contains approximately 40,000 sample images of volcanic eruption landform. Each sample image measures 25 pixels×25 pixels with the spatial resolution of 10 m. To validate and analyze VELSD1. 0, we employed both classical convolutional neural networks (e. g. , GoogLeNet, ResNet) and transformer- based architectures (e. g. , Vision Transformer, Swin Transformer). Experimental results demonstrate the dataset' s strong applicability and transferability for deep learning models, achieving an overall classification accuracy (OA) of 82. 93%, a Kappa coefficient (KC) of 75. 64%, and a mean average precision (mAP) of 84. 22%. VELSD1. 0 effectively supports deep learning classification of other volcanic eruption landforms from remote sensing imagery, contributing to the informatization and intellectualization of remote sensing surveys for volcanic landforms.