5:15 PM - 7:15 PM
[PPS09-P19] Derivation of dust optical depth from images taken by a Mars rover using radiative transfer model and deep learning method
Keywords:Mars, Dust, Deep learning, Radiative transfer model
Martian dust storms, which can be developed from local to global scales, are important meteorological phenomena that affect the temperature rise of the atmosphere and enhance atmospheric circulation. Dust storms increase the amount of dust in the Martian atmosphere, which reduces the amount of solar irradiation near the surface, resulting in poor visibility and reduced solar power generation. To date, various observations from orbiters and rovers, as well as simulation studies, have been conducted to elucidate atmospheric phenomena on Mars such as dust storms. However, the characteristics of the weather field at the time of occurrence of dust storms are not yet clear, and accurate reproductions of the occurrences and enhancement of dust storms have not yet been achieved. Thus, our research aims to establish the “weather forecasting on Mars" for the safe implementation of future Mars exploration activities.
To realize the weather forecasting on Mars, we should focus more on observations of the weather in the Martian boundary layer (the layer from the surface to several kilometers in altitude), which has not been deeply understood due to the lack of observations from the Martian surface and the difficulty of high-precision observations of the lower atmosphere from satellites. In addition, considering that AMeDAS (Automated Meteorological Data Acquisition System) has greatly contributed to the development of weather forecasting on Earth, it is desirable to make multi-point observations. Therefore, we have considered a new and much more convenient observational method of the amount of dust in the lower atmospheric layer at multiple points on surface. We are developing a deep learning model to estimate the amount of dust from images taken by a Mars rover, taking advantage of the fact that the visibility of the atmosphere varies depending on how much dust there is in the atmosphere.
We trained a convolutional neural network using images taken by the Mars rover Curiosity in MY31-36 and optical depths observed by Curiosity during the same period, and estimated dust optical depth from images (Kashimura et al., JpGU 2024). Note that it was just a simple analysis that only inputs images and other data such as geometry into the model, and there was still room for improvement in training data. Therefore, in this study, we added some innovations to training data so that the deep learning model can learn the effect of dust on images more directly. We calculated the sky radiances with a certain standard optical depth using the radiative transfer model DISORT (DIScrete Ordinate Radiative Transfer) and input them into the model together with observed images so that the model can learn how the sky image changes depending on the value of optical depth.
We aim to improve the results to be closer to the observations by reviewing the architecture of the model and the manners of feature engineering from images. The presentation will show the results obtained to date.
To realize the weather forecasting on Mars, we should focus more on observations of the weather in the Martian boundary layer (the layer from the surface to several kilometers in altitude), which has not been deeply understood due to the lack of observations from the Martian surface and the difficulty of high-precision observations of the lower atmosphere from satellites. In addition, considering that AMeDAS (Automated Meteorological Data Acquisition System) has greatly contributed to the development of weather forecasting on Earth, it is desirable to make multi-point observations. Therefore, we have considered a new and much more convenient observational method of the amount of dust in the lower atmospheric layer at multiple points on surface. We are developing a deep learning model to estimate the amount of dust from images taken by a Mars rover, taking advantage of the fact that the visibility of the atmosphere varies depending on how much dust there is in the atmosphere.
We trained a convolutional neural network using images taken by the Mars rover Curiosity in MY31-36 and optical depths observed by Curiosity during the same period, and estimated dust optical depth from images (Kashimura et al., JpGU 2024). Note that it was just a simple analysis that only inputs images and other data such as geometry into the model, and there was still room for improvement in training data. Therefore, in this study, we added some innovations to training data so that the deep learning model can learn the effect of dust on images more directly. We calculated the sky radiances with a certain standard optical depth using the radiative transfer model DISORT (DIScrete Ordinate Radiative Transfer) and input them into the model together with observed images so that the model can learn how the sky image changes depending on the value of optical depth.
We aim to improve the results to be closer to the observations by reviewing the architecture of the model and the manners of feature engineering from images. The presentation will show the results obtained to date.