Buosi, Samuele and Sonarghare, Shubham and McDonald, John and Mc Carthy, Timothy (2021) More efficient Geospatial ML modelling techniques for identifying man-made features in Aerial Ortho-imagery. Proceedings of the 23rd Irish Machine Vision and Image Processing conference IMVIP 2021. pp. 25-32. ISSN 978-0-9934207-6-4

Preview

Download (1MB) | Preview

Abstract

Deep learning techniques are used to achieve state-of-art accuracy in semantic segmentation on aerial ortho-imagery datasets. These algorithms are known to be efficient in terms of accuracy but at the expense of computational power required for training and subsequent inference operations. In this paper we strive to achieve a comparable performance but with lower floating point operations per second (FLOPS) and less training time. With this in mind, we chose to evaluate the EfficientNet-B0 network configured with 5.3 millions parameters and 0.39 billion FLOPS as a feature extractor operating inside a U-net architecture, achieving accuracy levels (mean F1 score of 0.869) comparable to a state-of-the-art deep learning architecture (U-net with Resnet50 as backbone) configured with 25.6 million parameters and 4.1 billion FLOPS which achieved a mean F1 score of 0.87. These promising results demonstrate that employing EfficientNet as the feature extractor in semantic segmentation on aerial ortho-imagery can be an effective strategy, in achieving higher performance results in terms of computational power, especially when running these networks on the edge.

Item Type:	Article
Keywords:	Deep Learning; Supervised Image Segmentation; semantic segmentation; ortho-imagery; Deep convolutional neural network;
Academic Unit:	Faculty of Science and Engineering > Computer Science
Item ID:	15154
Depositing User:	Tim McCarthy
Date Deposited:	04 Jan 2022 11:33
Journal or Publication Title:	Proceedings of the 23rd Irish Machine Vision and Image Processing conference IMVIP 2021
Publisher:	Dublin City University
Refereed:	Yes
Funders:	Science Foundation Ireland
URI:	Publisher
Use Licence:	This item is available under a Creative Commons Attribution Non Commercial Share Alike Licence (CC BY-NC-SA). Details of this licence are available here

Repository Staff Only(login required)

Item control page

Downloads

Origin of downloads