Homogeneous multimodal wide-area vehicle type recognition using multispectral satellite imagery

To address the challenge of vehicle-type recognition in complex traffic scenes within wide-area satellite imagery, where vehicles are easily confused with the surrounding background, this paper exploits the distinct spectral-response differences among various ground objects in multispectral data to guide the model to focus on vehicle targets. Based on satellite imagery comprising seven visible (VIS) bands and one near-infrared (NIR) band, this research is structured into a two-stage framework. In the first stage, a dedicated Preprocessing Framework for Vehicle Type Recognition (PFVTR) is developed for the satellite imagery, aiming to identify the sensitive bands within the visible-light spectrum for this specific task. Consequently, the Visible and Near-Infrared Multispectral Dataset (VNMD) is established based on the sensitive VIS bands and NIR imagery. In the second stage, to exploit the environmental robustness of the near-infrared band, the Multispectral Information Fusion-based Multimodal Vehicle Type Recognition model (MIF-MVTR) is proposed. By adaptively fusing the sensitive visible bands and the near-infrared spectrum, the model counters environmental interference to bolster recognition precision. Experimental results on the VNMD and public VEDAI datasets, as well as under complex illumination conditions, demonstrate that MIF-MVTR effectively exploits complementary spectral information to improve vehicle-type recognition performance in complex scenarios. These results further validate the robustness and generalization capability of the proposed model. Overall, spectral information plays an important role in alleviating vehicle-background confusion in wide-area traffic scenes.