Article

A support vector machine-based water detection analysis in a heterogeneous landscape using Landsat TM imagery

Surface water mapping is essential for studying global environmental changes in the quantity and quality of water bodies. This study explores the applicability of machine learning algorithm Support Vector Machine (SVM) in detecting open water surface from land. The study also compares the SVM based water extraction method with computationally simpler water index, modified Normalized Difference Water Index (mNMDI). The St. Croix Watershed Area is used as a test site for its humid environment with wetlands, built area, forest cover, and shadows as the background noise. The study uses Landsat TM data to generate spectral water index of mNDWI. Zero thresholding is used to generate binary images showing water and non-water areas. Two different SVM models, that is, Linear and Radial Basis Function (RBF) are also used to classify the Landsat TM image into water and non-water class. The accuracy of mNDWI and SVM classifiers are tested and compared using error matrices, Kappa coefficient, overall accuracy and McNemar chi-square test. The results show that the mathematically simpler mNDWI performed better than computationally complex SVM in terms of overall accuracy and Kappa coefficient. Furthermore, mNDWI accurately extracted water from narrow streams and wide rivers while SVM extracted water more accurately from locations in close proximity to urban areas such as reservoirs, boat launching ramps and locations with preponderance of wetlands.

Items in ScholarWorks are protected by copyright, with all rights reserved, unless otherwise indicated.