Ricarido Saturay, Jr.

The presentation will demonstrate the NLP-geocoding pipeline, utilizing news reports from a specific news agency to generate a sample landslide inventory, with the Philippines as the use case. **Landslide inventories** are important inputs for mapping landslide susceptibility, and for establishing rainfall and earthquake thresholds for landslide triggering. Online news reports comprise a body of largely untapped landslide information. A **natural language processing** (NLP) and **geocoding** pipeline, built on [*Trafilatura*](https://github.com/adbar/trafilatura), [*Spacy*](https://github.com/explosion/spaCy), [*Networkx*](https://github.com/networkx), [*Geopy*](https://github.com/geopy/geopy), and [*Nominatim*](https://github.com/osm-search/Nominatim), uses open source language models ([*en-core-web-\[sm*](https://github.com/explosion/spacy-models/releases//tag/en_core_web_sm-2.3.1)[*/md*](https://github.com/explosion/spacy-models/releases//tag/en_core_web_md-2.3.1)[*/lg\]*](https://github.com/explosion/spacy-models/releases//tag/en_core_web_lg-2.3.1)) and [*Openstreetmap*](https://www.openstreetmap.org/copyright/) data to output structured information about landslides reported in a news article, particularly the inferred dates and locations which are relevant to landslide inventories. The pipeline can be linked upstream to modules which provide URLs to news articles , and downstream to modules which allow further validation by manual or AI-driven methods. This application has the potential to provide accurate landslide dates, as news reports typically have an expressed date stamp for the occurrence of a landslide. It complements remote sensing-based methods which provide site-specific landslide locations but with less accurate dates due to the fixed time intervals and coarse temporal resolution of satellite observations. This work contributes to efforts in landslide mapping and early warning, within the frameworks of disaster risk reduction and climate change adaptation.