Optical remote sensing imagery provides critical spectral information for mapping landslide inventories but is frequently hindered by cloud cover and adverse weather. In contrast, Synthetic Aperture Radar (SAR) penetrates clouds and is highly sensitive to surface backscatter changes, offering complementary insights for detecting landslide disturbances. This study presents an integrated approach combining optical and SAR data to enable rapid and reliable landslide detection under challenging conditions. The proposed framework applies object-based image analysis (OBIA) to segment terrain into meaningful units and derives NDVIdiff and NDSI indices from pre- and post-event imagery, together with six GLCM-based texture features, to characterize surface disturbances. Four classification scenarios, including optical-only, SAR-only, cloud-obstructed optical, and fusion-based models, were systematically compared. Results demonstrate that the fusion approach consistently yields more spatially coherent and complete landslide inventories, while SAR-based mapping alone successfully delineates most large-scale landslides under heavy cloud cover. These findings confirm the operational effectiveness of the proposed optical–SAR framework for rapid, event-driven landslide mapping and disaster risk assessment.

This study investigates the effects of Digital Elevation Model (DEM) and Digital Bathymetric Model (DBM) resolutions, as well as coastline positioning accuracy, on gravimetric geoid modeling across the land–sea interface along the Qingshui Cliff coast in eastern Taiwan. Two coastline datasets, namely the Global Self-consistent, Hierarchical, High-resolution Geography (GSHHG) database and manually digitized coastlines, were combined with DEM-DBM datasets of different spatial resolutions. Under the remove–compute–restore framework, gravimetric geoid models were constructed using the Residual Terrain Model and Least-Squares Collocation methods, and their accuracy was evaluated using GNSS/leveling-derived geoid heights. The results show that the GSHHG coastline deviates noticeably from the actual shoreline in cliff-type coastal areas, leading to terrain masking and correction errors near the coast. In contrast, manually digitized coastlines improve boundary consistency and model stability. Resolution comparisons further indicate that the 3″×3″ DEM-DBM provides better overall accuracy than the 9″×9″ and 1″×1″ models. These findings suggest that accurate coastline definition and appropriate terrain resolution are essential for improving geoid modeling in complex land–sea transition zones.

With the increasing use of unmanned aerial vehicles (UAVs) in infrastructure monitoring and environmental inspection, stable image quality has become critical for deep learning and photogrammetry applications. However, UAV images are often degraded by environmental disturbances, while existing quality filtering still relies on manual inspection, making it unsuitable for high-frequency or real-time deployment. This study proposes a realtime, reference-free image quality assessment (IQA) framework based on a Swin-Unet architecture to improve screening efficiency and ensure data quality stability, while simultaneously generating image quality maps (IQMs) for downstream applications. To overcome limitations of traditional SSIM-based methods, including the requirement for reference images and sensitivity to pixel misalignment, an improved metric, termed CLIP-SSIM (CSSIM), is introduced to construct an image scoring model. A probability-weighted Swin-Transformer is first employed to generate high-accuracy IQMs (RMSE = 0.0193); however, its pixel-wise inference is computationally expensive (532 seconds per image). Therefore, the generated IQMs are used as supervisory labels to train a Swin-Unet model, enabling real-time inference (0.3 s per image) with acceptable accuracy (RMSE = 0.04). The proposed approach provides an efficient, accurate, and scalable solution for UAV image screening, effectively replacing manual inspection in high-frequency UAV applications.

Image stitching can expand the field of view and eliminate blind spots, but differences in scene depth often lead to parallax and ghosting. To address this, this study integrates single-image depth estimation and semantic segmentation models to establish a facade image stitching workflow for bridges, reconstructing a complete structural appearance map for damage analysis and management. Through transfer learning, a bridge-side image dataset was constructed, and a pretrained RGB-D semantic segmentation model was fine-tuned. The model achieved an mIoU of 86.44%, mAcc of 91.24%, recall of 92.11%, and F1-score of 91.56%, demonstrating stability and generalization capability, while indirectly validating the accuracy of the depth estimation model. To correct geometric misalignment caused by image tilt, depth maps were used to reconstruct point clouds for correction. A stitching accuracy comparison shows that the average SSIM of images corrected with the segmentation model (0.6807) was higher than that of the traditional method (0.5081), confirming the advantages of the proposed approach in terms of accuracy and visual consistency.