GrapHiSM: a graph-based hierarchical semantics-driven model for aerial scene classification under scarcity of labelled samples

Article Type

Research Article

Publication Title

Applied Intelligence

Abstract

Aerial scene classification can be treated as the problem of acquiring high-level semantic interpretations of the earth-surface images, remotely captured from the space or from the aerial vehicles. Although the topic is already extensively explored thus far, due to the high complexities and diversities in geometrical and spatial texture of aerial scenes, there still remain several open challenges. This paper primarily focuses on a comparatively new challenge of aerial scene classification under labelled sample scarcity, which restricts the promising deep network models, such as convolutional neural networks, to attain the desired accuracy. Even the graph convolutional networks, which have added strength of capturing spatial relationships, fail to perform well when trained with limited training samples. We address this issue by generating hierarchical semantics-driven multiple graph representations for each image, and subsequently, employing graph representation learning over these multitude of graphs which act as augmented training samples. Our graph-based hierarchical semantics-driven model (GrapHiSM) is evaluated using benchmark UC-Merced and AID datasets. Experimental results exhibit efficacy of GrapHiSM, in handling labelled sample scarcity at the time of aerial scene classification.

First Page

25919

Last Page

25930

DOI

https://10.1007/s10489-023-04919-4

Publication Date

11-1-2023

This document is currently not available here.

Share

COinS