Authors: Caiyun Zhang*, Florida Atlantic University
Topics: Hazards, Risks, and Disasters, Remote Sensing, Coastal and Marine
Keywords: Coastal disaster, risk, remote sensing modeling, machine learning
Session Type: Paper
Start / End Time: 5:00 PM / 6:40 PM
Room: Taylor, Marriott, Mezzanine Level
Presentation File: No File Uploaded
Mangrove forests are productive ecosystems but are vulnerable to hurricanes. In this study, we quantified the damage of mangroves from Hurricane Irma at a large-scale using Landsat data, and characterized the vulnerability of mangroves to hurricanes using three internal/physical metrics (a vegetation index, ground elevation and distance to open ocean) and two external/hurricane-related metrics (hurricane track and storm surge inundation). From the assessment of damage and vulnerability we developed risk projections to model mangrove damages from future hurricanes. Four machine learning techniques including Artificial Neural Network (ANN), Support Vector Machine (SVM), Random Forest (RF), and k-Nearest Neighbor (k-NN) were examined and compared with the Multiple Linear Regression (MLR) method for risk model development. The models were calibrated and validated using data before and after Hurricane Irma. Machine learning algorithms had a better performance than the linear model, and RF achieved the best result with a correlation coefficient (r) of 0.89 in predicting mangrove damages. We applied object-based modeling and mapping techniques and produced a mangrove hurricane vulnerability map, mangrove damage maps from Irma and a worst-case scenario hurricane with an intensity of Category 5 and a track along the mangrove distribution in the southern coastal Everglades. A total of 332 km2 of mangroves were severely damaged from Irma, and 673 km2 would be devastated from the modeled scenario. The techniques developed here can be used for other mangrove forests over hurricane-prone regions.