Machine learning approach for modeling daily pluvial flood dynamics in agricultural landscapes

Despite rural, agricultural landscapes being exposed to pluvial flooding, prior predictive flood modeling research has largely focused on urban areas. To improve and extend pluvial flood modeling approaches for use in agricultural regions, we built a machine learning model framework that uses remotely sensed imagery from Planet Labs, gridded rainfall data, and open-access geospatial landscape characteristics to produce a pluvial flood timeline. A Random Forest model was trained and daily flood timeline was generated for Hurricane Matthew (2016) at a 10-m resolution. The results show the model predicts pluvial flooding well, with overall accuracy of 0.97 and F1 score of 0.69. Further evaluation of model outputs highlighted that corn and soybean crops were most impacted by the pluvial flooding. The model may be used to identify agricultural areas susceptible to pluvial flooding, crops that may be potentially impacted, and characteristics of areas that experience pluvial flooding.