A Unifying Approach to Classifying Wetlands in the Ontonagon River Basin, Michigan, Using Multi-temporal Landsat-8 OLI Imagery

Accurate spatial information is critical to the assessment and protection of wetlands in the context of human intervention and global climate change. However, it is challenging to map and monitor wetland vegetation classes with satisfactory results because of their highly seasonal dynamics, spatial heterogeneity, and spectral similarity. This paper examines the effectiveness of various classification approaches commonly employed in wetland mapping, including the support vector machine (SVM) algorithm, maximum likelihood classifier (MLC), classification and regression tree (CART) and other remote sensing indices, by using multi-temporal Landsat-8 Operational Land Imager (OLI) spectral data and end-member fraction data as well as terrain data. These different mapping approaches were compared in the Ontonagon River drainage basin in upper Michigan, USA, where easily-confused wetland types such as forested wetland, palustrine scrub/shrub wetland, and palustrine emergent wetland are extensively distributed. The results show that multi-temporal data sets can reduce the classification omission caused by the highly seasonal dynamic of wetlands. The spatial heterogeneity can partly be characterized by using end-member fraction maps. Classification of the fraction maps had better results than that using the original spectral data, which implies that the selection of data inputs could be more important than the selection of classifiers. While each algorithm has its own capability of discriminating specific wetland types, CART showed relatively better classification results than the others due to its compatibility and lack of assumptions about data normal distribution. Finally, the authors developed a decision tree method (called DTM), which adopted the satisfactory resultant classification of specific wetland types using MLC and SVM classification, to update the wetland map of Ontonagon River Basin with acceptable accuracy (overall accuracy of 89% and kappa coefficients of 0.89).