Unveiling the hidden: Early detection of invasive vegetation in crops with UAV multispectral imaging

This work presents a practical framework for early detection of camouflaged weeds in agricultural production environments using UAV-based near-infrared (NIR) imagery. The proposed approach integrates RGB-NIR image fusion with advanced camouflaged object detection (COD) models to support precision agriculture applications at farm scale. Fourteen state-of-the-art (SOTA) fusion techniques are evaluated, and the three best-performing methods are selected for multispectral integration, while nine SOTA COD models are trained and tested on RGB and RGB-NIR fused data. Experimental validation on two UAV-based case studies–a commercial banana plantation (Weeds-Banana) and a maize field (WeedsGalore)–demonstrates that RGB-NIR fusion consistently improves weed segmentation accuracy compared with RGB-only inputs. Overall, the proposed framework provides a transferable and application-oriented solution for weed detection, enabling targeted interventions and contributing to reduced herbicide use and more sustainable crop management. The dataset is publicly available at GitHub: https://cod-espol.github.io/COD-Weeds/ .