MOMO: Mars Orbital Model Foundation Model for Mars Orbital Applications
Summary: arXiv:2604.02719v1 Announce Type: cross
Abstract
We introduce MOMO, the first multi-sensor foundation model for Mars remote sensing. MOMO uses model merge to integrate representations learned independently from three key Martian sensors (HiRISE, CTX, and THEMIS), spanning resolutions from 0.25 m/pixel to 100 m/pixel. Central to our method is our novel Equal Validation Loss (EVL) strategy, which aligns checkpoints across sensors based on validation loss similarity before fusion via task arithmetic. This ensures models are merged at compatible convergence stages, leading to improved stability and generalization.
Introduction
The exploration of Mars has significantly advanced over the past two decades, largely due to the availability of high-resolution orbital data. However, integrating information from multiple sensors has posed challenges in achieving meaningful insights. MOMO addresses this issue by leveraging a diverse set of data sources, providing a robust foundation for various Martian applications.
Methodology
MOMO employs a novel approach to model merging that utilizes:
- Model Merge: This technique integrates the strengths of different sensors to create a unified model.
- Equal Validation Loss (EVL): A strategy that aligns model weights based on the similarity of validation losses, ensuring consistent convergence stages.
- Task Arithmetic: This allows for the efficient fusion of models from different sensors, thus enhancing their performance.
Data and Training
The training process for MOMO involved a large-scale dataset comprising approximately 12 million samples, meticulously curated from various Martian orbital datasets. This extensive training corpus is pivotal for ensuring that MOMO can generalize well across different tasks.
Evaluation
MOMO was evaluated on nine downstream tasks from the Mars-Bench suite, demonstrating remarkable performance improvements across the board. The model particularly excelled in segmentation tasks, indicating its effectiveness in interpreting complex Martian surface features.
Results
The results of the evaluation revealed that MOMO outperforms several benchmarks, including:
- ImageNet pre-trained models
- Earth observation foundation models
- Sensor-specific pre-training methodologies
- Fully-supervised training approaches
These findings underscore the potential of model merging and optimal checkpoint selection strategies for building effective foundation models in multi-resolution contexts.
Conclusion
MOMO represents a significant advancement in the field of Mars remote sensing, offering a robust platform for future research and exploration. The model weights, pretraining code, pretraining data, and evaluation code are openly available at https://github.com/kerner-lab/MOMO, facilitating further research and development in this critical area.