Surface water detection and delineation is an important and necessary step in change detection studies on water bodies using multispectral images. Commonly used techniques for surface water delineation from multispectral images are single band density slicing, spectral index based, machine learning based classification and spectral un mixing based methods. This paper presents a comparative study of commonly used machine learning algorithms viz. ANN, SVM, Decision Tree, Random Forest and K-means clustering for their suitability and effectiveness when applied on Landsat 8 images for surface water detection and delineation. The algorithms are compared for their classification accuracy and execution time. While all the above mentioned algorithms exhibited their usefulness in water detection, Decision Tree and Random Forest algorithms were found be faster in both training phase and testing phase and also yielded better accuracy with fewer miss-classifications. Though K-means clustering with more than four clusters yielded results comparable to that of supervised classification methods, it requires appropriate post-processing to obtain the output image with only two clusters; corresponding to water pixels and non-water pixels. Pierson's correlation co-efficient and Structural similarity Index (SSIM) are computed to compare the correlation and similarity of the output images yielded by the algorithms being studied.
Volume 12 | Issue 3
Pages: 207-216
DOI: 10.5373/JARDCS/V12I3/20201184