Automated Identification of Global Martian Stratigraphy Based on a Two-Stage Transfer Learning Framework

The identification and correlation of Martian stratigraphy are fundamental to establishing a global stratigraphic framework and reconstructing the paleoclimate and geological evolutionary history. While orbital missions have accumulated a petabytes of high-resolution remote sensing imagery of Mars, global-scale stratigraphic studies reveal a significant gap: a larger portion of these data remains effectively uninterpreted and undocumented, which severely constrains paleoclimatic reconstruction and cross-regional stratigraphic correlation of Mars' geological history. To address this critical issue, this study proposes an innovative two-stage transfer learning framework designed to achieve automated and large-scale identification of global Martian stratigraphic imagery. In the first stage, the model is pre-trained using the large-scale, continuous stratigraphic imagery from the North Polar Layered Deposits (NPLD) to extract universal stratigraphic texture features with cross-regional generalization. In the second stage, a few-shot fine-tuning strategy is introduced to achieve rapid regional adaptation for fragmented stratigraphic outcrops in mid-low latitudes regions. This approach effectively decouples general feature learning from local domain adaptation, mechanistically mitigating model performance degradation caused by latitudinal geomorphological variations. Preliminary experimental results demonstrate that compared with conventional single-stage training paradigms, this method significantly improves the recall rate in cross-regional stratigraphic identification tasks and exhibits the model's exceptional generalization stability. This research provides a feasible quantitative analytical pathway for fully unlocking the value of massive Martian remote sensing data and constructing a high-precision global Martian stratigraphic database.