Objective With the widespread use of unmanned aerial vehicles (UAVs) in modern warfare, the demand for autonomous and safe landing in complex environments has become increasingly significant. For fixed-wing UAVs, the landing stage is characterized by relatively high speed, limited maneuvering time, strict trajectory constraints, and short decision windows. In practical missions, the aircraft often faces incomplete ground information, unknown obstacle distribution, and adverse weather disturbances such as fog or reduced visibility. Under such conditions, conventional landing approaches relying on prior maps, manual guidance, or static route planning are often insufficient to guarantee safety and efficiency. Therefore, rapidly obtaining reliable three-dimensional structural information of the landing area through onboard vision systems has become an important technical approach for improving autonomous landing capability. However, most existing three-dimensional reconstruction methods primarily pursue global reconstruction accuracy and scene completeness, while lacking task-oriented strategies that prioritize key landing-related regions. In addition, under degraded environments, image quality reduction, ineffective observations, and slower model convergence further limit practical deployment. To address these issues, this study proposes a three-dimensional reconstruction strategy for UAV autonomous landing in complex environments, aiming to improve perception efficiency and reconstruction effectiveness within limited observation opportunities.

Methods  The proposed method employs 3D Gaussian splatting (3DGS) as the baseline reconstruction framework because of its high rendering efficiency, fast optimization speed, and suitability for multi-view aerial image modeling. Compared with conventional implicit neural field methods, 3DGS is more appropriate for incremental scene updating and near-real-time reconstruction tasks required during landing operations. On this basis, an uncertainty-guided active perception mechanism was introduced. Specifically, reconstruction uncertainty was estimated according to the discrepancy between rendered images generated by the current model and newly acquired real observations. Regions with larger residual errors were considered insufficiently reconstructed or weakly observed areas, indicating greater potential information gain for subsequent sensing. Based on the uncertainty distribution, candidate viewpoints were evaluated through expected information gain, and the next observation direction was adaptively selected to maximize uncertainty reduction in critical regions. In this way, observation resources were preferentially allocated to high-value areas such as occluded structures, distant zones, and potential landing-related regions, rather than uniformly covering the entire scene. After new images were collected, the reconstructed model was incrementally updated, and the procedure of observation, reconstruction, evaluation, and replanning was iteratively repeated until the target quality requirement was reached. To verify the effectiveness of the proposed strategy, comparative experiments were conducted under different weather conditions, observation policies, and sensor configurations. Random viewpoint selection and sequential viewpoint sampling were used as baseline strategies. Sensor settings included different image resolutions and noise levels to simulate practical onboard camera conditions. Quantitative evaluation adopted the number of views required to achieve target reconstruction quality as the primary metric, while peak signal-to-noise ratio (PSNR) was used as an auxiliary indicator to measure reconstruction fidelity.

Results and Discussions  Experimental results demonstrated that the proposed method consistently required fewer observation views to reach predefined reconstruction quality than passive baseline strategies. Compared with random viewpoint selection and sequential observation, the required number of views was reduced by up to 43.5% and 67.5%, respectively, indicating that active viewpoint planning can significantly improve information acquisition efficiency during the reconstruction process. This reduction is particularly meaningful for fixed-wing UAV landing tasks, where available sensing time and repeated observation opportunities are inherently constrained. Under a fixed budget of 20 views, the proposed strategy also achieved higher reconstruction quality, with PSNR improvements of up to 2.11 dB and 4.24 dB compared with the random and sequential strategies, respectively. These results confirm that the method not only accelerates convergence but also enhances model quality under the same observation cost. Under different sensor resolutions and noise levels, the proposed strategy maintained stable relative advantages, demonstrating good adaptability to varying device configurations. Higher-resolution sensors generally provided richer structural details and clearer textures, which benefited reconstruction quality. Lower-noise sensors also improved feature consistency and reduced uncertainty accumulation. Nevertheless, regardless of hardware settings, the proposed active strategy still outperformed passive observation methods, showing that viewpoint scheduling remains an important factor beyond sensor capability alone. In foggy environments, the performance of all methods degraded to some extent because of visibility reduction, contrast attenuation, and texture loss caused by atmospheric scattering. However, the proposed method still preserved faster convergence and higher observation efficiency than baseline strategies. By continuously directing sensing actions toward uncertain regions, the method reduced redundant observations in already well-reconstructed areas and compensated for perception loss caused by environmental degradation. These findings indicate that uncertainty-guided observation planning is effective for improving robustness under adverse weather conditions. Overall, the experimental results verify that optimizing reconstruction strategy is as important as improving reconstruction models themselves for task-oriented UAV landing applications.

Conclusions This study proposed a three-dimensional reconstruction strategy for UAV autonomous landing in complex environments. By integrating 3D Gaussian splatting with uncertainty-driven active viewpoint selection, the proposed framework enables adaptive observation planning and incremental scene reconstruction during landing-related perception tasks. Experimental results verified that the method can achieve target reconstruction quality with substantially fewer views, obtain higher PSNR under fixed observation budgets, and maintain stable advantages under foggy and noisy conditions. The study demonstrates the practical value of combining reconstruction models with task-oriented sensing strategies, and provides useful reference for 3D perception planning, onboard vision system design, and autonomous landing technology development for UAVs operating in complex environments.