Application of Machine Learning to Model Wetland Inundation Patterns Across a Large Semiarid Floodplain. Issue 11 (8th November 2019)

• Record Type: Journal Article
• Authors: Shaeri Karimi, Sara; Saintilan, Neil; Wen, Li; Valavi, Roozbeh
• Is Part Of: Water resources research. Volume 55:Issue 11(2019)
• Page Start: 8765
• Page End: 8778
• Publication Date: 2019-11-08
• Subjects: machine learning -- downsampling -- sensitivity‐specificity sum maximizer -- inundation regime -- wetland -- environmental water; Hydrology -- Periodicals; 333.91
• DOI: 10.1029/2019WR024884
• Languages: English
• ISSNs: 0043-1397

Abstract

Abstract: Inundation is a primary driver of floodplain ecology. Understanding temporal and spatial variability of inundation patterns is critical for optimum resource management, particularly in striking an appropriate balance between environmental water application and extractive use. Nevertheless, quantifying inundation at the fine resolution required of ecological modeling is an immense challenge in these environments. In this study, Random Forest, a machine learning technique, was implemented to predict the inundation pattern in a section of the Darling River Floodplain, Australia, at a spatial scale of 30 m and daily temporal resolution. The model achieved very good performance with an average accuracy of 0.915 based on the area under the receiver operating characteristic curve over 10 runs of the model in testing data sets. Six variables explained 70% of the total contribution to inundation occurrence, with the most influential being landscape shape (local deviation from global mean elevation), elevation‐weighted distance to the river, the magnitude of river flow (10‐ and 30‐day accumulated river discharge), local rainfall, and soil moisture. This approach is applicable to other floodplains across the world where understanding of fine‐scale inundation pattern is for operational ecological management and scenario testing. Key Points: A machine learning technique (Random Forest) was successfully applied to predict daily floodplain inundation at a fine resolution of 30 m The downsampling method effectively improved model performance Topography and 30‐day cumulative discharge are the most influential parameters predicting inundation occurrence