Extending Planet Four: machine learning applied to seasonal CO2 jet activity on Mars

Seasonal CO₂ jet activity is one of the most dynamic surface processes currently observed on Mars [1, 2, 3, 4]. During spring, a solid-state greenhouse effect occurs due to the translucent seasonal CO₂ ice and generates pressurized gas pockets beneath the ice layer. This eventually leads to jet-like eruptions that transport dust and regolith from the substrate onto the ice surface. These events produce characteristic dark fans and blotches observed in high-resolution imagery. Their morphology and evolution provide direct information about near-surface winds, ice properties, and present-day surface–atmosphere interactions.

Repeated observations by HiRISE (High Resolution Imaging Science Experiment) onboard the Mars Reconnaissance Orbiter [5] have documented seasonal activity across large parts of the south polar regions over ten Mars years, from MY28 to MY38. The Planet Four citizen-science project was set up to outline dark seasonal deposits and has produced a large catalog of manually identified fans and blotches within selected regions of interest (ROIs). This dataset enabled the first statistical analyses of seasonal activity [6, 7], including studies of seasonal timing, regional distributions, and interannual variability [8]. However, existing analyses remain incomplete because a) they are limited to selected ROIs, b) rely on the narrow central color swath of HiRISE images, and c) do not cover the most recent years of HiRISE observations. As a result, the large-scale distribution and variability of seasonal activity across the south polar region remain poorly constrained.

This work focusses on the first application of machine learning (ML) techniques to the full record of seasonal HiRISE observations of these dark fans and blotches. Expanding on a baseline study [9] for automatic feature detection for citizen-science labelled data in selected regions for two MY, ML models are being trained and validated using the existing Planet Four catalog and will then be applied to all relevant HiRISE observations in the southern polar regions, including the full red-channel width as well as images outside of the current Planet Four ROIs. The resulting dataset will provide large-scale fan and blotch distributions over the full extent of HiRISE observations. This expanded activity catalog will be used to address several key scientific questions. First, the project will investigate how seasonal activity is distributed across the south polar cap and whether currently-known active regions are representative of broader seasonal processes. Second, repeated observations will be used to determine whether CO₂ jet activity recurs at stable locations over time, potentially implicating persistent local conditions. Third, the project will examine the interannual variability of activity patterns and fan orientations and compare these observations with atmospheric circulation predicted by higher-resolution atmospheric models (expanding the study in [7] to additional years and locations). Fan orientations derived from machine-learning detections will be used as proxies for near-surface winds and compared with modeled wind fields to assess the relationship between observed activity and atmospheric dynamics.

Fig 1. Example comparison of detection of fans and blotches by Planet Four citizen science project (on the left) and ML model YOLO11 trained on Planet Four labeled dataset [10] (on the right).

References:

[1] Kieffer, H. H. 2007, Journal of Geophysical Research, 112, E08005

[2] Hansen, C. J., Thomas, N., Portyankina, G., et al. 2010, Icarus 205, 283

[3] Thomas, N., Pommerol, A., Hauber, E., Portyankina. G. et al. 2025, SSR 221, 3

[4] Hansen, C. J., Byrne, S., Calvin, W. M. et al. 2024, Icarus 419, 115801

[5] McEwen, A. S., Byrne, S., Hansen, C. J. et al. 2024, Icarus 419, 115795

[6] Aye, K. M., Schwamb, M. E., Portyankina, G., et al. 2019, Icarus, 319, 558

[7] Portyankina, G., Michaels, T. I., Aye, K.-M., et al. 2022, Planetary Science Journal, 3, 31

[8] Hansen, C. J., Aye, K.-M., et al. 2023, LPI Contribution No. 2806, id.2315

[9] McDonnell, M. D., Jones, E., Schwamb, M. E., et al. 2023, Icarus, 391, 115308

[10] Jocher, G. and Qiu, J., 2024, https://github.com/ultralytics/ultralytics