Convolutional neural networks and satellite imagery: How deep is necessary?

NeurIPS 2023 Workshop CompSust Submission21 Authors

Keywords: satellite imagery, CNNs, remote sensing

TL;DR: ReNet-type CNNs may be overkill for many satellite imagery tasks; we show a simple 3-layer CNN is comparable to ResNet in many satellite imagery tasks.

Abstract: Applying off-the-shelf models (e.g., ResNet) to satellite imagery has become standard practice. While convolutional neural networks (CNNs) have been shown to outperform baseline methods in remote sensing prediction tasks, differences in satellite and natural images (i.e., images that comprise common datasets like ImageNet and CIFAR-10) may make ResNet-type models overkill for many satellite imagery tasks. In this paper, we present a comparison of off-the-shelf CNNs to a much smaller CNN over a range of satellite imagery tasks and show that a CNN with significantly fewer parameters performs on par with standard CNN architectures for five out of six tasks. Our findings are especially pertinent to those working with satellite imagery who face computational constraints.

Other Workshops: No

Submission Number: 21