Using Hyperspatial LiDAR and Multispectral Imaging to Identify Coastal Wetlands Using Gradient Boosting Methods

Abstract

Wetlands play a vital role in our ecosystems, preserving water quality, controlling flooding, and supplying aquifers. Wetlands are rapidly degrading due to threats by human encroachment and rising sea levels. Effective and timely mapping of wetland ecosystems is vital to their preservation. Unoccupied Aircraft Systems (UAS) have demonstrated the capability to access and record data from difficult-to-reach wetlands at a rapid pace, increasing the viability of wetland identification and classification through machine learning (ML) methods. This study proposes a UAS-based gradient boosting approach to wetland classification in coastal regions using hyperspatial LiDAR and multispectral (MS) data, implemented on a series of wetland sites in the Atlantic Coastal Plain region of North Carolina, USA. Our results demonstrated that Xtreme Gradient Boosting performed the best on a cross-site dataset with an accuracy of 83.20% and an Area Under Curve (AUC) score of 0.8994. The study also found that Digital Terrain Model-based variables had the greatest feature importance on a cross-site dataset. This study’s novelty lies in utilizing cross-site validation using Gradient Boosting methods with limited amounts of UAS data while explicitly considering topographical features and vegetation characteristics derived from multi-source UAS collections for both wetland and non-wetland classes. Future work is encouraged with a larger dataset or with semi-supervised learning techniques to improve the accuracy of the model.