The need for training and benchmark datasets for convolutional neural networks in flood applications

Khouakhi, AbdouZawadzka, JoannaTruckell, Ian2022-07-192022-07-192022-05-17Khouakhi A, Zawadzka J. Truckell I. (2022) The need for training and benchmark datasets for convolutional neural networks in flood applications, Hydrology Research, Volume 53, Issue 6, pp. 795-8061998-9563https://doi.org/10.2166/nh.2022.093https://dspace.lib.cranfield.ac.uk/handle/1826/18188Flood-related image datasets from social media, smartphones, CCTV cameras, and unmanned aerial vehicles (UAVs) present valuable data for the management of flood risk, and particularly for the application of modern convolutional neural networks (CNNs) to specific flood-related problems such as flood extent detection and flood depth estimation. This review discusses the increasing role of CNNs in flood research with a growing number of published datasets, particularly since 2018. We note the lack of open and labelled flood image datasets and the growing need for an open, benchmark data library for image classification, object detection, and segmentation relevant to flood management. Such a library would provide benchmark datasets to advance CNN flood applications in general and serve as a resource, providing data scientists and the flood research community with the necessary data for model training and validation.enAttribution 4.0 Internationalconvolutional neural networksflood imagesfloodsThe need for training and benchmark datasets for convolutional neural networks in flood applicationsArticle