<p>Global landslide mapping presents a significant challenge due to the complex morphology of landslides and the difficulty of identifying them across diverse regions. Conventional approaches to mapping rely on manual data collection, which is time-consuming and limited to documented landslides, often missing those in remote areas. In contrast, modern methods, such as deep learning-based approaches, offer faster, more comprehensive mapping using remote sensing images, reducing time and data gaps while improving the visualization of complex landslide morphology. While deep learning-based landslide mapping methods have shown promising results in small-scale areas, automatic identification on a global scale remains difficult due to the diverse geographic distribution and complexity of landslides. Model limitations, such as difficulty generalizing across regions and data scarcity in remote areas, hinder global detection efforts. To address these challenges, we propose a novel global landslide detection model (GLDM) that improves the YOLOX architecture to better handle large-scale data diversity and enhance model generalization. ASFF (Adaptive Spatial Feature Fusion) is incorporated into YOLOX to adaptively fuse features from different scales, ensuring consistency in feature representation across different feature scales. a Tibetan Plateau landslide remote sensing image dataset is curated, consisting of landslides, rainfall-induced debris, and freeze–thaw-induced debris. The primary focus of the research is on landslides, as they are the main target for global landslide detection, while the inclusion of rainfall-induced and freeze–thaw debris helps the model become more robust by training it to distinguish landslides from similar features. Finally, GLDM is trained using this dataset and subsequently deployed for global landslide detection, culminating in the identification of 10,016 landslide areas. These results highlight the effectiveness of ASFF in improving detection performance by enabling the model to better handle diverse landslide features. Deep convolutional neural networks prove effective in overcoming regional dataset limitations, as demonstrated by the model's ability to detect landslides in varied global regions despite the initial focus on the Tibetan Plateau dataset. Notably, the similarity between the landslide features in the Tibetan Plateau and those globally allows the model to generalize better across regions, playing a key role in the effectiveness of this research. These global landslide mapping efforts provide valuable insights for advancing research and mitigation strategies by offering a comprehensive global landslide distribution map, which can aid in identifying high-risk areas, improving early warning systems, and guiding targeted intervention measures.</p>