Automatic Mapping and Characterising Of Ripple Patterns on Sand Dunes using U-Net and 2D Semi-Variogram.

The shape, orientation, and spacing of sand ripples provide insights into local sand transport conditions. Consequently, mapping ripple patterns can enhance our understanding of wind regimes on Mars and Earth, especially in areas lacking local wind observations. However, manually mapping these ripple patterns is time-consuming and subjective, underscoring the need for automated methods to analyze large areas consistently. Our goal was to automatically quantify three types of ripple patterns—straight ripples, sinuous ripples, and complex textures—and to map their distribution over barchan dunes. We introduce two innovative and complementary methods for identifying these ripple patterns in high-resolution satellite imagery of Martian dunes. The efficacy of both approaches was assessed on 42 barchan dunes across 6 HiRISE sites.

The first method, a machine learning model known as U-Net, proved to be more reliable in classifying the ripple patterns, achieving 86% precision, 82% recall, and 82% F1-score for image tiles, as well as 82% precision, 77% recall, and 79% F1-score for complete dune mappings. The second method, a spatial autocorrelation analysis called the 2D semi-variogram, performed poorly in classifying ripple patterns over entire dunes, with precision reaching up to 56%, recall at 45%, and F1-score at 39%. However, it excelled in accurately measuring ripple spacing (R² = 0.78) and orientation (R² = 0.98). By combining the U-Net model's efficiency in ripple classification with the 2D semi-variogram’s precision in measuring spacing and orientation, we can conduct extensive analyses of ripples and local wind regimes across Mars. Furthermore, these methods hold potential for application in drone imagery of terrestrial dunes, paving the way for further research and exploration.