The purpose of our work is to define an original approach to determine the threshold of unimodal image histograms in a robust manner. Our proposed segmentation approach refers to Shannon entropy. Threshold estimation is based on the exploitation of the curve of the entropy loss quantity. The final expression of the sigmoid estimated function of the entropy error and derivatives are used to select the image's threshold segmentation. The proposed approach is evaluated on a set of SAS (Synthetic Aperture Sonar) images.