Automatic Detection and Segmentation of Barchan Dunes on Mars and Earth Using a Convolutional Neural Network

Abstract

The morphology of isolated barchan dunes on Mars and Earth may shed light on the dynamic conditions that form them, their migration direction and the physical properties of the sediments composing them. Prior to this study, dune fields have been largely analyzed manually from aerial and satellite imagery, as automatic detection techniques are often not sufficiently accurate in outlining dunes. Here, we employ an instance segmentation neural network to detect and outline isolated barchan dunes on Mars and Earth. We train and test the model on martian targets using Mars Reconnaissance Orbiter (MRO) Context Camera (CTX) images, and find it sufficiently accurate (mAP=77% on the test dataset) to characterize dune field dynamics. Using our trained model we detect and map the global distribution of barchan dunes relative to previously mapped dune fields, and find that barchan dunes are more abundant in the northern hemisphere than in the southern hemisphere. These contrasting abundances of barchans may reflect latitudinally dependent wind regimes, sediment supply, or sediment availability.

A distinct ripple-formation regime on Mars revealed by the morphometrics of barchan dunes

“…However, compiling such an extensive dataset has been challenging to date due to the large number of dunes on Mars and the time-consuming nature of manual measurements. Barchan dunes are ideal objects for machine-learning assisted detection because they appear as isolated units and their unique crescentic shape makes them more readily detectable by both humans and machines compared to other types of dunes 26 .…”

“…We elected to exclude connected barchanoidal ridges that are more difficult to analyze automatically. Using this dataset, we optimized the parameters of Mask R-CNN through training, until its accuracy (the mean average precision evaluated using a separate test dataset 26 ) reached 77%. Our dataset included 1074 images of barchan dunes, uniformly selected across the martian surface to obtain a well balanced dataset.…”

“…The full set of hyperparameters of the trained model can be found in ref. 26 . While training the model further would have likely resulted in higher accuracy and better outlier detection, we found this mAP value sufficient for our goal of characterizing the first-order morphometrics of barchan dunes on Mars.…”

“…A full discussion of training process and model evaluation is provided in ref. 26 , and a discussion of the errors associated with our statistical analysis is provided in the Supplementary Information.…”

“…S1 – S2 ) on Mars to reveal geospatial trends as they relate to the global near-surface atmospheric circulation and local environmental conditions. Our extensive dataset, which includes over a million barchan dunes previously outlined by an instance segmentation neural network and validated using a comprehensive manually processed dataset 26 (Fig. 2a and Figs.…”