Significant full reference image segmentation evaluation: a survey in remote sensing field

Image segmentation is a crucial step in remote sensing application, as it breaks down a larger image into smaller chunks, which contains useful information. Although there are several image segmentation algorithms available, evaluation of the algorithms is challenging. Furthermore, the evaluation of image segmentation can elucidate the best image segmentation algorithm for a single image or a group of image or a whole class of image. This paper explores and evaluates the benefits and the drawbacks of various qualitative and quantitative image segmentation evaluation metrics used in remote sensing applications. For all the metrics, a quantitative set of values for good and bad segmentation is provided. In addition, some image segmentation algorithms such as Multi Otsu Threshold, K Means clustering, Fuzzy C Means clustering, Improved K Means clustering (IFCM), Improved Fuzzy C Means clustering, Naïve Bayes classifier, K Nearest Neighbor, Decision Tree (DT) and Random Forest classifier are used in the experimental comparison of metrics. The qualitative and quantitative satellite image segmentation evaluation using the mentioned algorithms is measured. The results are analyzed to strengthen the impact of different metrics on the segmentation algorithms. In both qualitative and quantitative analysis, the IFCM outperformed the other unsupervised machine learning algorithms and the DT outperformed the other supervised machine learning algorithms. The effectiveness of the provided metrics in the remote sensing field is validated.