Unraveling the climate evolution on Mars and Earth with AI-driven surface mapping and explainable AI

Our research leverages state-of-the-art deep learning techniques to automate surface mapping and continuous monitoring on planetary bodies. We are also developing tools to analyze the model uncertainty and decision-making in AI models with evaluation in our surface mapping projects and beyond. We focus on one of the solar system's most dynamic Earth-analog environment on terrestrial planets - Mars' northern polar region, a repository of the planet's climatic history within its extensive ice-layered dome. We detect small blocks [1] and their sources yielding a reliable method for monitoring mass wasting activity with valuable present-day erosion rate results [2]. In parallel, we investigate and map polygonal patterns on both Earth and Mars to assess the global distribution of polygons and their potential as indicator for climate conditions and changes. On Earth, polygons are indicators of the volume of ground ice and provide insights into permafrost vulnerability to climate change. On Mars, similar young landforms could be linked to geologically recent freeze-thaw cycles. This would be conflicting with the current environment and would have implications for the recent hydrologic past of the planet. The distribution of polygonal ground on Mars can provide valuable information on the role of liquid water in the recent past by shedding light on the formation mechanism. We use AI models for automated surface mapping because they achieve highly complex decision-making. However, they are usually treated as Black-Box systems. To tackle this problem, we are developing software tools for analyzing model uncertainty and decision-making within an application-independent framework. Typical questions are why did the model produce exactly this response and how certain is it about the correctness of its results? References: [1] Martynchuk O. et al., 2024. EGU 2024. [2] Su S. et al., AGU 2023.