Performance Evaluation of Thresholding-Based Segmentation Algorithms for Aerial Imagery

The effectiveness of various threshold-based segmentation algorithms is examined in this study utilizing aerial images, with a focus on accuracy metrics including intersection over union (IoU), variation of information (VI), objectness, and compactness. The investigation seeks to determine the best segmentation methods to handle the difficulties specific to aerial data. The study uses satellite data from the Mohammed Bin Rashid Space Centre (MBSRC) that includes aerial photography of Dubai, United Arab Emirates, to evaluate adaptive thresholding techniques with multi-Otsu and multilevel thresholding algorithms. To evaluate segmentation accuracy, the four segmentation algorithms are tested on individual mask pictures after being trained using 72 aerial photographs. Multilevel thresholding demonstrates superior performance when compared to multi-Otsu segmentation (with IoU values ranging from 0.6708 to 0.6854). This is substantiated by the observation of significantly higher Mean IoU values, which range from 0.8512 to 0.8533, suggesting enhanced segmentation accuracy. Furthermore, when contrasted with adaptive thresholding using the mean (with values spanning from 1.53E−06 to 4.90E−04) and Gaussian Kernel (ranging from 1.53E−06 to 2.77E−04), multilevel thresholding and multi-Otsu segmentation also exhibit superior mean accuracy (ranging from 1.53E−06 to 7.86E−04). This study shows that multilevel thresholding and multi-Otsu segmentation perform better than other techniques for segmenting aerial images in terms of accuracy and quality. These results help segmentation algorithms evolve by providing better boundary delineation and texture parameter selection. Such methods can improve the understanding of satellite data by helping to study the dynamics of natural objects using non-Euclidean fractal analysis.