Remote Sensing Image Scene Classification: Benchmark and State of the Art

Abstract

Remote sensing image scene classification plays an important role in a wide\nrange of applications and hence has been receiving remarkable attention. During\nthe past years, significant efforts have been made to develop various datasets\nor present a variety of approaches for scene classification from remote sensing\nimages. However, a systematic review of the literature concerning datasets and\nmethods for scene classification is still lacking. In addition, almost all\nexisting datasets have a number of limitations, including the small scale of\nscene classes and the image numbers, the lack of image variations and\ndiversity, and the saturation of accuracy. These limitations severely limit the\ndevelopment of new approaches especially deep learning-based methods. This\npaper first provides a comprehensive review of the recent progress. Then, we\npropose a large-scale dataset, termed "NWPU-RESISC45", which is a publicly\navailable benchmark for REmote Sensing Image Scene Classification (RESISC),\ncreated by Northwestern Polytechnical University (NWPU). This dataset contains\n31,500 images, covering 45 scene classes with 700 images in each class. The\nproposed NWPU-RESISC45 (i) is large-scale on the scene classes and the total\nimage number, (ii) holds big variations in translation, spatial resolution,\nviewpoint, object pose, illumination, background, and occlusion, and (iii) has\nhigh within-class diversity and between-class similarity. The creation of this\ndataset will enable the community to develop and evaluate various data-driven\nalgorithms. Finally, several representative methods are evaluated using the\nproposed dataset and the results are reported as a useful baseline for future\nresearch.\n