Uncertainty-aware Spatiotemporal Inference of Landslide Events by Fusing Multimodal Social Media Information with Geographic Features

Social media can provide rapid on-site information that helps to improve situational awareness in disaster response. Nevertheless, social media posts often provide imprecise or ambiguous location information (e.g., toponyms), leaving the exact location within the referenced area highly uncertain. In addition, the actual event time may deviate from the posting time. Existing toponym-based geocoding approaches typically reduce a place name to a single representative point, which is insufficient to capture within-area spatial uncertainty and to integrate heterogeneous evidence.

We propose an uncertainty-aware spatiotemporal inference framework that fuses geographic factors with multimodal social media information to estimate both the most likely event location and occurrence date, using landslides as an event type with topographic and hydro-climatic location and time constraints. The framework is evaluated using landslide-related social media posts monitored by the Global Landslide Detector in the contiguous United States. First, toponyms extracted from posts are geocoded into candidate geometries that constrain the spatial search domain. Second, we build a spatial probability map by combining a landslide susceptibility raster representing topographic constraints with image-derived semantic cues. CLIP is used to detect roads and water bodies from post images, which adaptively weight road/river buffer zones before normalization. Third, within a time window before the post date, we extract PRISM daily precipitation series as a hydro-climatic constraint, and fuse it with the spatial probability to form a joint spatiotemporal score. The framework outputs (i) a spatial probability map and (ii) the most likely occurrence date.

We evaluate the method using posts with manually annotated coordinates and assess map quality using the Percentile Rank (PR) of the ground-truth pixel, among other metrics. Preliminary results indicate that incorporating road–water features with image-driven semantic modulation consistently concentrates the true landslide location into smaller high-probability areas and yields event-time estimates consistent with rainfall-triggering processes. This provides an uncertainty-aware transferable framework for rapid, social-media-driven event localization and verification for event types with geographic constraints.