A Global Multimodal Flood Event Dataset

The increasing human activities have been aggravating global environmental changes, especially the frequent flood disasters which have caused severe threats to human society.  The analysis of flood events on a global scale in terms of their causes, social impacts, duration and scope is essential for disaster relief and prevention.We present Global Multimodal Flood Event Dataset (MFED) which is built upon the foundation of the Global Flood Database, which systematically collects remote-sensing image data and conducts comprehensive data cleansing procedures for textual data. MFED consists of an image modality which is composed of multi-source remote sensing images and a text modality which is composed of flood event information and flood news reports on the Internet. The image modality is derived from various remote sensing sources, capturing flooded areas, while the text modality draws from information records within the global flood database and authentic news reports.MFED collected remote sensing images and news records from 913 flood events in flood-prone areas of the world between 2000 and 2018, covering 29 countries. It contains 1138 scenes of optical images covering flood areas, 281 scenes of radar images, 9000 pieces of news data related to flood events, and 913 pieces of flood event record information.MFED consists of two parts: image and text. The image file is in geotiff format and stored in 'MFED/MFED_data' folder and packaged in MFED.zip file.The statistics in the Global Flood Database are named floodinfo.csv and contain a summary of the collected event data, including the central latitude and longitude coordinates of the flood event, the duration of the flood, the major country in which it flowed, the main cause of the flood, the path of the product file, and the latitude and longitude range of the remote sensing image collected based on the event information.The news report records crawled from Google are numbered by country-event and stored in csv format. Each file contains three parts: title, content and time. All news record files are packaged in the 'Text.zip'.Finally, we provide the 'MFED.jsonl' file, which records the correspondence between the optical image and the SAR image in jsonl format, and the matching text information for each pair.