Remote sensing scene classification under scarcity of labelled samples—A survey of the state-of-the-arts

Semantic labelling of remote sensing images, technically termed as remote sensing scene classification, plays significant role in understanding huge volume of complex remote sensing images. Eventually, this aids in a number of real-world decision making tasks, including land-cover management, urban planning, environmental monitoring, and so on. Over the past few years, different strategies have been adopted for proper labelling of remote sensing image scenes, and these are already studied by the existing surveys or review papers. Unlike these exploratory works, in this article, we exclusively focus on the recently proposed models that deal with remote sensing image scene classification under limited labelled sample scenario. We classify the existing works into three broad categories: Data-level, Model-level, and Algorithm-level. For each category, we formally present the working principles to tackle scarcity of labelled data, and also, provide insights into advantages/disadvantages of the adopted schemes. We further systematically analyse several interesting and open challenges, such as the incremental class problem (in streaming environment), presence of noisy training samples, multi-modality of the data, privacy issue, and so on, which give rise to additional complications during scene-level classification under limited labelled sample scenario. Finally, we provide a future vision of opportunities and discuss new perspectives pertaining to the development of this field.