International Journal of Applied Earth Observation and Geoinformation ( IF 7.5 ) Pub Date : 2022-09-18 , DOI: 10.1016/j.jag.2022.103022 Jiayi Li, Xin Huang, Lilin Tu
Hyperspectral image (HSI) classification is one of the most important remote sensing techniques. Currently, the performances of most of the HSI classification networks on the public HSI datasets are overoptimistic (i.e., the overall accuracy exceeds 98 %). This deficiency is partly due to the very limited scale of these existing datasets, which impedes the network generalization and poses limitations for the future development. The existing hyperspectral datasets urgently need to be scaled up in size. Therefore, in this study, we built a dataset named the WHU-OHS dataset, which consists of about 90 million manually labeled samples of 7795 Orbita hyperspectral satellite (OHS) image patches (sized 512 × 512) from 40 Chinese locations. This dataset ranges from the visible to near-infrared range, with an average spectral resolution of 15 nm. The extensive geographical distribution, large spatial coverage, and widely used classification system make the WHU-OHS dataset a challenging benchmark. This dataset was validated by comprehensive experiments using several representative deep HSI classification networks. Furthermore, the transferability of the HSI classification networks under the conditions of the same/different HSI sensors was tested. In particular, when classifying the existing public HSI datasets, using initial parameters obtained by pre-training on the WHU-OHS dataset can further improve the inference accuracy as well as the training efficiency. The WHU-OHS dataset and a PyTorch toolbox for large-scale HSI classification are available at https://irsip.whu.edu.cn/resources/resources_v2.php and https://github.com/zjjerica/WHU-OHS-Pytorch, respectively.