Robust UAV Image Mosaicking Using SIFT and LightGlue

Abstract. Unmanned Aerial Vehicle (UAV) imagery is playing an important role in various remote sensing applications, including precision agriculture, environmental monitoring, and urban planning. To construct a seamless and geometrically accurate mosaic from multiple overlapping UAV images, it is essential to interpret the camera geometry accurately and extract tiepoints between images reliably. In this study, we focus on evaluating the robustness and effectiveness of the combination of SIFT with LightGlue, a hybrid matching approach that integrates the rotationally invariant properties of SIFT with the contextual matching capabilities of LightGlue. For comparison, we tested two traditional methods, the SIFT with Brute-Force matcher and the SIFT with FLANN matcher, and one AI-based method, the SuperPoint with LightGlue matcher. These matching algorithms were applied to UAV datasets covering both high-texture regions, such as urban environments, and low-texture areas, such as agricultural fields. The performance evaluation was conducted based on several criteria, including the number and spatial distribution of tiepoints, epipolar error, rotational robustness, bundle adjustment stability, and mosaic completeness. Among all methods, the SIFT with LightGlue matcher consistently demonstrated the most reliable performance. This approach not only achieved robust and accurate matching in low-texture and high-rotation scenarios but also led to superior spatial consistency in bundle adjustment and final mosaics, confirming its suitability for practical UAV image mosaicking tasks.