Assessing Coastal Wetland Health: A Machine Learning Approach for Phragmites australis Detection

Keywords: Vegetation, Change Detection, Machine learning, NDVI, Phragmites australis
Abstract Type: Poster Abstract

Abstract

Detecting and mapping wetland ecosystems are challenging due to their complex morphology and varying vegetation types. Wetland mapping has been performed with optical remote sensing data for a long-time using machine learning algorithms, however, few studies have ever conducted a direct comparison of different approaches for Phragmites australis (Roseau cane) classification. This study compares the effectiveness of three machine learning classifiers: Maximum Likelihood Classification (MLC), Support Vector Machine (SVM), and Random Forest (RF) for identifying Roseau cane in the Rockefeller Wildlife Refuge using multispectral Landsat imagery from 2010 and 2022. Training and validation points were created based on NAIP images and classification was done in the ArcGIS Pro environment based on the created training samples and schema. The accuracies of the classified maps were assessed using overall accuracy (OA), Cohens’ kappa coefficient (k). From this research analysis, it was found that among the three classifiers, SVM outperformed MLC and RF achieving 90% OA with a Kappa value of 0.81 in 2010 and 86% OA with a Kappa of 0.76 in 2022. These results suggest that SVM technique provides a reliable approach for large-scale wetland classification. Additionally, an NDVI-based analysis showed a 34 sq. km reduction in vegetation over the study period, highlighting the ongoing vulnerability of wetland ecosystems and the need of continuous monitoring to mitigate ecosystem decline.