Presentation Information
[PPS07-P07]Derivation of Dust Optical Depth from Images Taken by a Mars Rover Using Deep Learning
Itsuki Kashimura1, *Takeshi Kuroda1,2, Tomohiro Sato2, Hironobu Iwabuchi1, Shohei Aoki1,3, Hiromu Nakagawa1, Naoki Terada1 (1.Department of Geophysics, Tohoku University, 2.National Institute of Information and Communications Technology, 3.Department of Complexity Science and Engineering, The University of Tokyo)
Keywords:
Mars,Deep Learning,Dust Cycle,Landing Exploration
The presence of Martian airborne dust and its temporal and spatial variability critically affect the weather system on Mars. Optical depth is one of the fundamental parameters to describe dust abundances. However, direct solar imaging, the standard method used by landers and rovers for measuring optical depth, requires high maintenance and dedicated operational resources in the missions. In contrast, sky imaging is less constrained operationally, but the resulting data are more challenging to analyze. Here we propose a simple and efficient Artificial Intelligence based algorithm for observing optical depth from sky images. Our neural network model was designed to derive optical depth from (i) sky radiance factors at 495 (blue), 554 (green), and 640 nm (red) and their ratios extracted from Martian landscape sky images, and (ii) the corresponding viewing and illumination geometry. The performance metrics identified the 14-layer neural network as the optimal model. It scored a Mean Absolute Error of 0.0458, a Mean Absolute Percentage Error of 5.59 %, and a coefficient of determination (R2 score) of 0.991. According to the Permutation Feature Importance analysis, deeper models tended to distribute importance relatively uniformly across the input variables, and the red-to-green ratio was prioritized. While a large uncertainty was observed for deriving low optical depth (<0.5), the model performed reliable regression when optical depth was greater than approximately 0.5. The deep-learning-based framework proposed in this study offers a promising path toward high-precision monitoring of Martian dust optical depth using a computationally efficient algorithm, which could substantially enrich the data available for future missions.