Efficient Semantic Segmentation with Hyperspectral Images

Abstract

Many robotic application domains, such as agriculture, environmental management, or industrial automation, could benefit from hyperspectral images and their ample spectrum information. However, the huge volume of information makes automated semantic segmentation of hyperspectral images both a daunting task and a barrier to its adoption. This paper explores alternatives to obtain accurate and more efficient results for semantic segmentation in hyperspectral imagery. The proposed pipeline combines standard methods to reduce the dimensionality of the input, followed by an efficient deep neural network to segment the information. Our results show that this pipeline performs better than available related work solutions in terms of accuracy and performance. Besides, this work makes a detailed study of the trade-offs between the amount of information received by the segmentation network and the accuracy, execution time, and energy efficiency of the solution.