Convolutional Neural Network (CNN) Based Martian Dune Detection

Abstract

Sand dunes are one of the most prominent Aeolian landforms present on the Martian surface. Accumulation and erosion of sand particles cause the formation of dunes, which possibly can influence the Martian climate too. For mapping such landforms over large areas of the Martian surface more effectively, automated detection of dunes has been brought out. For this a convolutional neural network (CNN) based detection approach has been implemented considering application of different models and assessing their respective performance using different sets of Orbiter images. CNN architectures such as U-net, ResUNet and ResUNet++ were used for segmentation of dunes over the Martian surface. CNN produced segmentation results with greater accuracy with advantage of designing new models and using different loss functions. Convolution neural networks such as U-Net, ResUNet, and ResUNet++ for detecting dunes on Martian surface used Context camera (CTX) and the High-Resolution Imaging Experiment (HiRISE) images of Mars Reconnaissance Orbiter (MRO) to generate the suitable models considering two different Martian sites, Gale crater and Nili Patera. The models thus generated were tested over Olympia Undae region of the Mars and all the architectures could produce more than 85% accuracy. The model created using CTX images performed well for Gale Crater region compared to the model created using HiRISE image. U-Net model created using CTX image performed well in case of low-quality images (coarse resolution noisy images) whereas, ResUNet ++ model created using HiRISE image performed well in case of good quality (fine resolution) images.