Global geological conditions are complex and variable, characterized by frequent plate movements, earthquakes, and volcanic eruptions. Coupled with significant climate differences, various factors interact to trigger frequent landslide disasters, resulting in substantial losses of life and property. Therefore, landslide monitoring is crucial. Traditional monitoring technologies face limitations when dealing with complex terrains and meeting the demands for high timeliness, while unmanned aerial vehicles (UAVs), with their maneuverability, high resolution, and ability to operate in hazardous environments, have been widely applied in landslide monitoring. This paper provides a comprehensive review of UAV-based 3D reconstruction for landslides, detailing the characteristics and application cases of UAVs, explaining the functions and limitations of sensors such as optical sensors and light detection and ranging (LiDAR), and exploring 3D reconstruction methods based on UAV imagery, LiDAR, and hybrid approaches. It analyzes the applications of UAV 3D reconstruction in landslide emergency investigation, monitoring, and disaster assessment. The paper identifies the technical challenges faced in these applications and proposes corresponding solutions. In addition, UAV-based 3D reconstruction technology—with its centimeter-level spatial resolution—enables the precise delineation of landslide extent and hazard potential, thereby enhancing monitoring accuracy and improving the efficiency of emergency investigations. This technology provides strong technical support for landslide research and prevention, with significant implications for reducing landslide disaster losses.

Recent advances in deep learning have substantially improved remote sensing change detection. However, most existing models still describe bi-temporal differences mainly from the spatial domain, making it difficult to fully capture complementary frequency domain cues in complex scenes. To address this limitation, this paper introduces a Spatial-Frequency Collaborative Learning Network (SFCLNet) for remote sensing change detection. In particular, hierarchical features are extracted from bi-temporal images using a Siamese backbone. A Spatial Domain Feature Fusion (SDFF) module is then designed to enhance local structural variation details by modeling the structural consistency between bi-temporal features. Meanwhile, a Frequency Domain Feature Fusion (FDFF) module is introduced to characterize frequency domain cues by separately modeling phase and amplitude components. Furthermore, a Spatial-Frequency Collaborative Fusion (SFCF) module is developed to obtain more discriminative change feature representations by integrating the fused spatial domain and frequency domain features in a channel-wise competitive way. Finally, the pixel-wise results are predicted using a UNet-based decoder that progressively aggregates the fused multi-level features. Experimental results on Google, LEVIR, and MSRS benchmark datasets show that SFCLNet achieves F1 scores of 88.94%, 91.39%, and 74.97%, respectively, outperforming several recently published methods. These results verify the effectiveness of jointly exploiting the frequency domain and spatial domain for remote sensing change detection.