Semi-automated landslide database development through online news and satelite images

With the increase in the frequency and magnitude of landslides observed in recent years, it is essential to improve risk management tools. To this end, the development of landslides databases must be improved in order to train and refine these tools more efficiently. The GDELT project, a global database that monitors and collects news from around the world, was used to collect news available on the web regarding landslides which occurred between 2015 and 2024 in the city of Petrópolis, the selected study area for the project. The result was compared with the landslide database prepared and provided by the Civil Defense of Petrópolis-RJ. The comparison was made visually, through graphs, and mathematically, through the Pearson correlation coefficient and through Spearman's rank correlation. Moreover, in an attempt to improve the temporal accuracy of the news-based database, keywords referring to periods of the day were identified. The results were compared to the times registered by the Civil Defense, and the news related to the cases in which there was a divergence were studied, in order to assess which result was closer to reality. Finally, seeking to improve spatial accuracy, satellite images were used in order to identify the difference in the vegetation index (in particular, MSAVI2) between before and after the date of a landslide occurrence to ascertain the appearance of slope failures. The news-based database presented a good annual and monthly precision and reasonable weekly precision for identifying landslide events. Moreover, it proved to be useful for identifying the period of the day in which a particular landslide with a significant impact occurred. However, this strategy is less accurate for events involving multiple landslides with a large impact. The Civil Defense database, on the other hand, may be useful in order to consider a larger number of landslides, including those of lesser impact, but it is not prone to highlighting high-impact particular events. Calculating the difference in vegetation index from multispectral images has proven useful for identifying the emergence of landslide scars.