Abstract:

The mapping of water body areas such as rivers, streams, lakes and ponds is very important in terms of monitoring water resources, determining, and managing their change over time. Extracting water body areas is a complicated process that is influenced by a variety of factors. For the problem of identifying water and non-water areas, various multi spectral band satellite imagery and classification-based approaches are used. In this study, non-parametric (Support Vector Machines, k-Nearest Neighborhood and Decision Trees), probabilistic (Hidden Markov Model) and deep learning (Auto-Encoder) based supervised classification, which produce more successful results than index-based methods, were used to investigate the effectiveness of classification algorithms. Since multispectral high spatial resolution satellite images are costly, the water surface areas of Arıklar and Denizli ponds were determined by using only the red, green and blue bands of the Sentinel-2 satellite image for classification. Experimental results were compared using metrics obtained from the confusion matrix such as accuracy, specificity, precision, sensitivity, f-Score, and statistical tools such as mean square error, structural similarity index, peak signal-to-noise ratio, and Kohen's Kappa metric used in image quality determination. In the quantitative and qualitative experimental results obtained, while the deep learning-based auto-encoder was the most successful method statistically, it was determined that the decision trees method worked faster in terms of time comparison.

Keywords: