Skip to Content
Remote SensingRemote Sensing
  • Article
  • Open Access

19 November 2026

39 Pages

LightMaskMamba: A Near-Ground Remote Sensing Object Detection Network with Visible–Infrared Fusion

,
,
,
,
,
and
1
School of Software and Artificial Intelligence, Guangzhou University of Software, Guangzhou 510990, China
2
National Key Laboratory of Agricultural Equipment Technology, Guangzhou 510642, China
3
School of Mechanical and Electrical Engineering, Dalian Minzu University, Dalian 116650, China
4
School of Mechatronical Engineering, Beijing Institute of Technology, Beijing 100081, China
This article belongs to the Section AI Remote Sensing

Abstract

Visible–infrared fusion is a main route to all-weather object extraction in near-ground remote sensing, but single-band RGB remains unstable under occlusion, clutter, and low light and illumination variation. This study presents LightMaskMamba, a lightweight visible–infrared object detection network for near-ground remote sensing. Public benchmarks test fusion quality across scenes, and the self-built near-ground wampee field dataset is for deployment-oriented field validation. The core of the network is a lightweight mask-guided Mamba fusion architecture. Its key module, the Lightweight Mask Mamba Fusion Block (LMMFB), uses an object-related mask to drive shallow-to-deep VIS–NIR feature fusion, focuses computation on object regions, and reduces scan redundancy. Dual-band detection therefore stays light enough to run at the edge. On the self-built wampee dataset, LightMaskMamba-v11s reaches 97.4% mAP50 and 52.0% mAP50:95. Relative to single-band YOLOv11s, mAP50 is 2.3 and 4.6 percentage points higher on IR and RGB. Parameters rise by 3.7M, from 8.8M to 12.5M. The same model is likewise effective in detection accuracy, and lighter, on public multimodal object-detection datasets. On a Jetson Orin NX Super, TensorRT INT8 mixed-precision inference reaches about 17.3 FPS, or 58 ms per image.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.