AUTOMATED CREST-LINE DETECTION AND ANALYSIS OF SAND DUNE PATTERNS

The pattern of dunes is a product of both external boundary conditions and the internal dynamics of pattern evolution. Extraction of morphometric parameters such as crest length, orientation, spacing, bifurcation and merging of crests from image data can reveal important information about dune-field morphological properties, development, and response to changes in boundary conditions, but manual methods are labor-intensive and time-consuming.

We are developing the capability to recognize and characterize patterns of sand dunes on planetary surfaces. Our goal is to develop a robust methodology and the necessary algorithms for automated or semi-automated extraction of dune morphometric information from image data.

We have evaluated the ability of a variety of approaches (e.g. appearance-based and gradient-based) to extract dune crest lines from image data of a range of linear and crescentic dune patterns on Earth and Mars. The most promising approach so far uses machine learning using the Support Vector Machine (SVM) and Gradient Boosted trees to train a classifier to recognize dune crest lines extracted with the Scale Invariant Feature Transform (SIFT). After training, each image pixel is classified and given a response score. Higher scores are kept as crest line candidates, filtered, and grouped to form dune crest lines. The quality of the results is measured against ground truth using precision-recall metrics, in which precision is the number of correct detections and recall is a measure of how much of the ground-truth was detected. Our results so far are promising but further work is needed to improve the filtering of false positives and to refine the detection of dune crest lines.