Clustering Based Flood Area Segmentation Benchmarking in Different Colour Spaces

Effective and quick response to natural disasters is crucial to save lives and reduce sufferings. Flooding is an annual concern in many countries with heavy rainfall, and monitoring systems are limited due to lack of data during flood events. Technological developments have increased interest in using computer vision techniques for automatic flood monitoring. One such technique, image segmentation, divides an image into multiple regions to extract the objects of interest. Methods include techniques such as thresholding, clustering, graph-based, superpixels, and watersheds. Here, we use a flood area segmentation dataset, comprising images of areas affected by floods and water region masks. The images, originally in the RGB format, are converted into <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathrm{L}^{*}\mathrm{a}^{*}\mathrm{b}^{*}$</tex>, HSV, and YCbCr colour spaces. We then apply a clustering method to extract flooding areas from 243 images within all four colour spaces and use Jaccard, Dice, and BF-score metrics to compare through k-means++ clustering. Finally, we applied the basic non-parametric bootstrap method to estimate the mean and its confidence interval for each metric result. This paper evaluates which colour space is more effective in segmenting flood areas using clustering based segmentation.