Beichen Zhang , Kelly Helm Smith, Frank Schilder, Fatima K. Abu Salem, Ashok Samal, Tsegaye Tadesse, Michael J. Hayes

Drought is recognized for its extensive and varied impacts. Based on the drought-related textual datasets from the National Drought Mitigation Center, our research applies advanced artificial intelligence techniques, including deep learning and natural language processing, to enhance the monitoring of multifaceted drought impacts in the United States. This study also delves into predicting drought-related impact labels from informal social media texts through transfer learning from the textual databases. Our findings reveal that deep learning models based on Transformers significantly outperform the baseline model based on the word frequency in labeling various drought impacts from media-sourced data and a collected tweet dataset spanning 2020 to 2022. Comparative analyses of predicted labels of drought-related tweets underscore the added value of social media data, which offers distinct insights into drought impacts beyond what is captured in news media or citizen science contributions. Case studies in California and Nebraska illustrate dynamic characteristics of drought impacts from the predicted labels of the tweet dataset at spatial and temporal scales. The analyses indicate that the monthly quantity of drought-related information in California is linked to drought severity, urban areas, and water supply and quality. In contrast, it is associated with the growing season, irrigated cropland, and agriculture in Nebraska. Consequently, this study suggests applying social media as a valuable supplementary data source, boosted by cutting-edge deep learning models, for monitoring drought impacts with the potential for quantitatively defining socioeconomic droughts from societal and public perspectives.