Modelling the Hawaiian Shoreline and Wetland Prediction

Purpose: Utilizing Earth Observations to Delineate Wetland Extents, Model Sea Level Rise Inundation Risk, and Assess Impacts on Historic Hawaiian Lands

Supervisors: Dr. Roberta Martin (Arizona State University), Dr. Jiwei Li (Arizona State University), Dr. David Hondula (Arizona State University)

Internship Term: Summer 2022

Software: ArcGIS Pro, Wetland Intrinsic Potential Tool, R-Studio,

Skills: Machine Learning, Data processing with Python and R programming, Data Analysis with ArcGIS Pro, Research

Introduction

Hawaii Island is the largest island in Hawaii with ~200,000 people. Coastal areas of the island currently face significant erosion and flooding due to climate change-induced sea level rise (Kane et al., 2015). The County of Hawaii and the State of Hawaii Department of Land and Natural Resources (DLNR) are developing the Shoreline Setback Plan and Climate Adaptation Plan to protect historical sites, ecosystems, and private properties. The Shoreline Setback Plan utilizes tidal gauges and flood risk projections to assist with determining their setback delimitations, and the Climate Adaptation Plan’s goal is to utilize NASA Earth observations to guide their mitigation and adaptation efforts.










The team aimed to tackle three main issues:

1) Locate unmapped and probabilistically
determine the formation of new wetlands on Hawaii Island,

2) Calibrate local tidal
gauge data against NASA remote sensing data to enable future investigations of
historical local sea surface height,

3) Create a short-term flood risk map to identify vulnerable zones.

Satellites/ Sensors Used

Wetland Extent Map

To create a wetland extent map, we used the Wetland Intrinsic Potential tool. WIP was developed by Meghan Halabisky at the Remote Sensing and Geospatial Analysis Lab at the University of Washington in conjunction with Dan Miller at TerrainWorks. The study area is divided into five hydrographic units and the model is trained to predict wetland vs. Upland location. The training data is input into the tool which utilizes the predictive power of the Random Forest machine learning algorithm to predict wetland locations.  Lastly, a probability raster output is derived for each hydrographic unit which shows wetland extents and locations. Probability ranges from high probabilities of upland at 0 or wetland at 1.

Results - Probable Wetland Extent

  • To address sea level rise, we performed a feasibility study on using machine learning to classify flood risk and determine factors or features most influential to flooding.

  • For our model label, we generated flood risk indices for 5 known flood events in Hawaii from 2019 – 2021 using the Global Flood Mapper application in GEE. 

  • The data were computed at 20 m transect intervals that traversed the entire Hawaii coastline, and each transect had a data buffer 500 m inland and 15 km seaward. The raw data came primarily in the form of rasters, which we processed in ArcGIS and Python to generate a final data set of ~30,000 data points per flood event.

Sea-Level Inundation

Result - Sea Level Inundation Model

  • Results showed that all models were tuned to achieve an out-of-bag error below 10%, this means that each model for each hydrographic unit performed with a 90% accuracy or greater in predicting wetland and upland locations based on the provided training data set.

  • The large wetlands shown with high probabilities in the North and North East of the Island were confirmed by project partners to be a large bog (North) and a consistently very wet forested area (North East), this local verification bolstered our model performance confidence.