Landslide detection plays a crucial role in rapid Disaster Risk Reduction Management (DRRM), enabling swift decision-making after a disaster strikes. This study investigates the effectiveness of semantic segmentation (SS) and change detection (CD) models for automated landslide mapping using high-resolution optical remote sensing imagery. We employ state-of-the-art deep learning techniques, comparing four models with and without the CLIP foundation model. Our geographically diverse dataset incorporates high-resolution RGB images from the GVLM Benchmark Dataset. Our experiments reveal that CD models generally outperform SS models, particularly the BAN model, which demonstrates superior robustness and granularity, especially for unseen datasets like those from the Philippines. These findings highlight the importance of incorporating diverse datasets and advanced pre-training methods (e.g., CLIP) to enhance the generalizability and reliability of automated landslide detection systems.