A Comparison of Convolutional Neural Networks for Remotely Sensed Imagery
Laurel Hopkins, Oregon State University, hopkilau@oregonstate.edu (Presenter)
Nate Butler, Oregon State University, butlenat@oregonstate.edu
Weng-Keen Wong, Oregon State University, wongwe@eecs.oregonstate.edu
Rebecca Hutchinson, Oregon State University, rah@engr.orst.edu
When applying machine learning to satellite imagery, it has become standard practice to apply off-the-shelf models (e.g., ResNet) to satellite images. While convolutional neural networks (CNNs) have been shown to outperform baseline methods in satellite imagery prediction tasks, there are substantial enough differences between natural images (i.e., images that comprise common datasets like ImageNet and CIFAR-10) and satellite images that off-the-shelf CNNs should not automatically be assumed to be the best suited for satellite imagery. We present a comparison of off-the-shelf CNNs to much simpler CNN architectures over a range of satellite imagery prediction tasks. We show that CNNs with simpler architectures and significantly fewer parameters perform on par with standard CNN architectures.
Associated Project(s):
Poster Location ID: 3-24
Presentation Type: Poster
Session: Poster Session 3
Session Date: Thu (May 11) 3:00-5:00 PM
CCE Program: BDEC