Comparative Evaluation of Image Segmentation Techniques for Flood Detection in High-Resolution Satellite Imagery

Abstract

Speedy reaction to natural disasters, such as floods, is critical to minimising loss of life and pain. Access to fast and reliable data is critical  for rescue teams. Satellite photography provides a wealth of data that may be analysed to assist pinpoint disaster-affected areas. The use  of segmentation to analyse satellite images is becoming increasingly important in environmental and climatic monitoring,  particularly in detecting and controlling natural disasters. Image segmentation improves pattern recognition, which divides a single  image into several homogeneous pieces. The efficiency of image segmentation techniques varies depending on the layout of objects,  illumination, shadow, and other variables. However, there is no one-size-fits-all method for successfully segmenting all imagery; specific  methods are more efficient than others. This report compares four different technologies. Commonly used image segmentation  techniques: K-means clustering (K.C.), Color thresholding (C.T.), Region-based Active Contour (R.A.C.) and Edge-based Active Contour  (E.A.C.) segmentation. These four techniques were used to detect, and segment flooded areas in high-resolution satellite imagery. The K.C. method had the best flood segmentation rate with a Jaccard Index of 0.8234, Dice of 0.9234, the precision of 0.9589, recall of 0.9078  and BFscore of 0.9327, which was higher than the other three segmentation technique and previous works.