Adaptive isoline interval optimization for precise contour segmentation and instance-level detection of martian impact craters

Abstract

This study presents an adaptive method for optimizing isoline intervals for semantic segmentation and object detection of Martian impact crater contours based on Mars HRSC DEM data. The method uses an isoline interval division technique to improve the discernibility of terrain boundary features. This enhancement enables effective tracking and detection of closed contours of impact craters within the detection box. An energy function is designed for precise crater contouring, and a Bayesian optimization algorithm is applied to adaptively adjust contour intervals and optimize the contours. To address inaccuracies in the detection box, the paper proposes a progressively iterative correction method for refining impact crater target boxes. This method adjusts the width and height of the target box during iterations to make the bounding box tighten and fully enclose the impact crater. The position and size of the detection box are automatically optimized, thereby improving its accuracy. A pre-trained Segment Anything Model (SAM) is used to segment the morphological structures of the impact crater. Experimental results show that, for Martian impact craters larger than 3 km in the study area using Mars HRSC DEM data, the contouring accuracy achieved an mIoU of 69.75% and an mF1-Score of 80.51%. This substantially outperforms the performance of traditional watershed and depression filling contouring methods. The correction method improves the mIoU accuracy of bounding boxes within the study area by 7.49% for the YOLOv8 object detection network and 15.28% for the Robbins crater catalog database. The resulting high-precise crater contours, bounding boxes, and internal morphological features provide valuable inputs for deep space exploration and planetary science.