Predictive Modeling of Flash Floods: Investigating Hydrology and Land Cover Dynamics through Remote Sensing Data

Abstract

Accounting for more than 75% of Federal Disaster Declarations, floods in the United States of America have increased both in intensity and frequency, outpacing the development of reliable predictive models and warning systems. This issue has only been exacerbated by a lack of large-scale public participation and access to predictive models. This paper focuses on classifying risk factors and building a prototype flood-prediction tool to clearly communicate results within a defined Area of Interest (AOI). The investigative strategy involves gathering and analyzing several geospatial indicators, which could result in high-risk flood conditions, such as relevant soil properties, elevation/slope, Curve Number (CN) values, and land cover types within an AOI. To ensure dataset reliability and scientific validity, the considered geological and atmospheric variables were sourced from a range of open-source databases such as Earth Map, GLOBE Citizen Science, FEMA, and Collect Earth Online. The second portion of this study concentrates on developing a classification-based prototype machine learning model that outputs flood risk in a given area into five levels. This model facilitates citizen science efforts through an interactive Colab project and subsequent GitHub repository, which allows for user input. Findings were that Land cover and slope/elevation are key factors in determining runoff potential, while soil properties do not contribute to Curve Number differences in the AOI due to a lack of detailed data. The S.H.I.E.L.D. model can easily be ported into a cross-platform application, further integrating citizen science observations based on the findings of this paper.