Impact of architecture on robustness and interpretability of multispectral deep neural networks

TL;DR

This paper evaluates how different multispectral fusion strategies in deep neural networks affect their performance, interpretability, and robustness to image corruptions, providing insights for optimal spectral information integration.

Contribution

It systematically compares various fusion methods in multispectral deep learning models, highlighting their impact on robustness and interpretability.

Findings

Early fusion models are more sensitive to spectral corruptions.

Late fusion approaches improve robustness to channel-specific noise.

Fusion strategy significantly influences model interpretability.

Abstract

Including information from additional spectral bands (e.g., near-infrared) can improve deep learning model performance for many vision-oriented tasks. There are many possible ways to incorporate this additional information into a deep learning model, but the optimal fusion strategy has not yet been determined and can vary between applications. At one extreme, known as "early fusion," additional bands are stacked as extra channels to obtain an input image with more than three channels. At the other extreme, known as "late fusion," RGB and non-RGB bands are passed through separate branches of a deep learning model and merged immediately before a final classification or segmentation layer. In this work, we characterize the performance of a suite of multispectral deep learning models with different fusion approaches, quantify their relative reliance on different input bands and evaluate their robustness to naturalistic image corruptions affecting one or more input channels.