Differential Analysis of Multispectral Images for Terrain Identification
For autonomous robot navigation, DRIFT addresses the problem of unreliable RGB-based terrain perception under challenging conditions with a practical, edge-deployable solution.
DRIFT, a lightweight multispectral framework combining raw spectral bands and band-ratio representations via dual-stream residual architecture and differential fusion, improves terrain identification robustness under low illumination and material ambiguities, achieving consistent improvements over strong baselines on a new oil-on-soil dataset and a controlled water-on-grass study.
Reliable terrain understanding is a prerequisite for autonomous robot navigation. Yet, the widespread RGB-based perception can fail under low illumination, shadows, and material ambiguities. In this work we propose DRIFT, a lightweight multispectral framework that combines raw spectral bands and illumination-tolerant band-ratio representations through a dual-stream residual architecture and a differential fusion branch. Band ratios attenuate multiplicative acquisition effects (illumination/sensor gains), while the differential fusion explicitly highlights discrepancies between absolute-band and ratio-derived cues, which improves the robustness to noisy or partially unreliable spectral measurements. In the paper (i) we evaluate DRIFT on a new oil-on-soil multispectral dataset acquired using a MicaSense RedEdge-P camera mounted on an Unmanned Aerial Vehicle, and (ii) we provide an additional controlled study on water-on-grass under varying illumination and thermal perturbations (hot/cold water) to analyze NIR-sensitive effects. DRIFT consistently improves over strong baselines, while remaining compatible with edge deployment.