A large-scale remote sensing scene dataset construction for semantic segmentation

Abstract

ABSTRACT As fuelled by the advancement of deep learning for computer vision tasks, its application in other fields has been boosted. This technology has been increasingly applied to the interpretation of remote sensing image, showing high potential economic and societal significance, such as automatically mapping land cover. However, the model requires a considerable number of samples for training, and it is now adversely affected by the lack of a large-scale dataset. Moreover, labelling samples is a time-consuming and laborious task, and a complete land classification system suitable for deep learning has not been established. This limitation hinders the development and application of deep learning. To meet the data needs of deep learning in the field of remote sensing, this study develops JSsampleP, a large-scale dataset for segmentation, generating 110,170 data samples that cover various categories of scenes within Jiangsu Province, China. The existing Geographical Condition Dataset (GCD) and Basic Surveying and Mapping Dataset (BSMD) in Jiangsu were fully utilised, significantly reducing the cost of labelling samples. Furthermore, the samples were subject to a rigorous cleaning process to ensure data quality. Finally, the accuracy of the dataset is verified using the U-Net model, and the future version will be optimised continuously.