Advance Dust Devil Detection with AI using Mars2020 MEDA instrument

Introduction
Mars’ dust cycle is a critical factor that drives the weather and climate of the planet. Airborne
dust affects the energy balance that drives the atmospheric dynamic. Therefore, for studying the present-day and recent-past climate of Mars we need to observe and understand the different processes involved in the dust cycle. To this end, the Mars Environmental Dynamics Analyser (MEDA) station [1] includes a set of sensors capable of measuring the radiance fluxes, the wind direction and velocity, the pressure, and the humidity over the Martian surface. Combining these observations with radiative transfer (RT) simulations, airborne dust particles can be detected and characterized (optical depth, particle size, refractive index) along the day. The retrieval of these dust properties allows us to analyze dust storms or dust-lifting events, such as dust devils, on Mars [2][3].
Dust devils are thought to account for 50% of the total dust budget, and they represent a
continuous source of lifted dust, active even outside the dust storms season. For these reasons, they have been proposed as the main mechanism able to sustain the ever-observed dust haze of the Martian atmosphere. Our radiative transfer simulations indicate that variations in the dust loading near the surface can be detected and characterized by MEDA radiance sensor RDS [4].
This study reanalyzes the dataset of dust devil detections obtained in [3] employing artificial
intelligence techniques including anomaly detection based on autoencoders [5] and deep learning models [6] to analyze RDS and pressure sensor data. As we will show, preliminary results indicate that our AI models can successfully identify and characterize these phenomena with high accuracy. The final aim is to develop a powerful tool that can improve the database for the following sols of the mission, and subsequently extend its use for other atmospheric studies.

Dataset
The dataset used in this study includes data from 365 Martian days, which represents half a Martian year of observations and it was collected with a temporal resolution of one sample per second (for 12 hours in average observations per sol). This dataset, which has been labeled by hand, contains 424 detected Dust Devil events. The duration of these events varies considerably, ranging from just a few seconds to several minutes. The precise manual labeling is crucial for training reliable machine learning models that can effectively recognize and predict these events. 

 

AI techniques
Deep learning is increasingly used to detect events in time-based signals, significantly enhancing accuracy and speed. Methods like CNN [7] and RNN [8] identify complex patterns effectively. These models learn from vast data, enabling real-time and predictive analytics.
Deep learning-based autoencoders are powerful tools for anomaly detection in temporal signals [5], offering a sophisticated method to capture complex patterns and identify outliers. Autoencoders, which are neural networks designed to reconstruct their input, learn to represent normal patterns during training. By minimizing the reconstruction error, these models learn to encode the regular, predictable aspects of temporal data. When an anomalous signal occurs, it typically results in a higher reconstruction error due to deviations from learned patterns.
The nature of the events and the low frequency of occurrence presents two main challenges for the Machine Learning algorithm:
• The database is highly unbalance. As it is said in Section Dataset, 424 events are detected over 365 Martian days. It means that less than 0.2% of the database corresponds to Dust Devil cases.
• Due to the spatial nature of the events, the Dust Devils manifests in different RDS sensors each time. This makes it difficult for the algorithm to generalize.

Results
For the experiments, the data is windowed and data augmentation techniques are used to try to correct the issue of class imbalance. The training set was selected randomly and intentionally balanced to ensure an equal number of samples per class. For testing the models, data from six randomly chosen suns are selected. The following results (Table 1) are obtained:  

Name	Train Accuracy	Test Accuracy
DNN	75.0%	65.7%
CNN	82.5%	78.9%
LSTM	73.1%	67.0%

At the time of writing this document, there are no consistent results for autoencoder’s approach.

Conclusions and Future Steps
This study demonstrates that data augmentation and advanced AI techniques can significantly improve the detection of dust devils on Mars. The use of deep learning, specifically convolutional neural networks (CNNs), has shown to outperform other models in accuracy both during training and testing phases.
Future work will focus on enhancing feature extraction, exploring new data augmentation techniques, and further developing autoencoders.

Acknowledgements
This work was funded by the Spanish Ministry of Science and Innovation under Project PID2021-129043OB-I00 (funded by MCIN/AEI/10.13039/501100011033/FEDER, EU), and by the Community of Madrid and University of Alcala under project EPU-INV/2020/003.

 

[1] Rodriguez-Manfredi at al. The mars environmental dynamics analyzer, meda. a suite of
environmental sensors for the mars 2020 mission. Space Science Reviews, 217, 04 2021.
[2] Lemmond et al. Dust, sand, and winds within an active martian storm in jezero crater. Geophysical Research Letters, 49(17):e2022GL100126, 2022. e2022GL100126 2022GL100126.

[3] Toledo et al. Dust devil frequency of occurrence and radiative effects at jezero crater, mars, as
measured by meda radiation and dust sensor (rds). Journal of Geophysical Research: Planets,
128(1):e2022JE007494, 2023. e2022JE007494 2022JE007494.
[4] Ap´estigue et al. Radiation and dust sensor for mars environmental dynamic analyzer onboard m2020 rover. Sensors, 22(8), 2022.
[5] Fatemeh Esmaeili, Erica Cassie, Hong Phan T. Nguyen, Natalie O. V. Plank, Charles P. Unsworth, and Alan Wang. Anomaly detection for sensor signals utilizing deep learning autoencoder-based neural networks. Bioengineering, 10(4), 2023.
[6] Alaa Sheta, Hamza Turabieh, Thaer Thaher, Jingwei Too, Majdi Mafarja, Md Shafaeat Hossain, and Salim R. Surani. Diagnosis of obstructive sleep apnea from ecg signals using machine learning and deep learning classifiers. Applied Sciences, 11(14), 2021.
[7] Naoya Takahashi, Michael Gygli, Beat Pfister, and Luc Van Gool. Deep convolutional neural
networks and data augmentation for acoustic event detection. CoRR, abs/1604.07160, 2016.
[8] Yong Yu, Xiaosheng Si, Changhua Hu, and Jianxun Zhang. A Review of Recurrent Neural
Networks: LSTM Cells and Network Architectures. Neural Computation, 31(7):1235–1270, 07
2019.