A new climate impact database using generative AI

Storms, heat waves, wildfires, floods, and other extreme weather climate-related disasters pose a significant threat to society and ecosystems, which in many cases is being aggravated by climate change. Understanding and quantifying the impacts of extreme weather climate events is thus a crucial scientific and societal challenge. Disaster databases are extremely useful for establishing the link between climate events and socio-economic impacts. However, publicly available data on impacts is generally scarce. Apart from existing open disaster databases such as EM-DAT, robust data on the impacts of climate extremes can also be found in textual documents, such as newspapers, reports and Wikipedia articles. Here we present a new climate impact database that has been built based on multiple public textual entries using a pipeline of data cleaning, key information extraction and validation. In particular, we constructed the database by using the state-of-the-art generative artificial intelligence language models GPT4, Llama2 and other advanced natural language processing techniques. We note that our dataset contains more records in the early time period of 1900-1960 and in specific areas such as than the benchmark database EM-DAT. Our research highlights the opportunities of natural language processing to collect data on climate impacts, which can complement existing open impact datasets to provide a more robust information on the impacts of weather and climate events.