[test]: Explore fusion strategies for Thor and TerraMind embeddings

[Dinghye]

We want to explore how to better combine Thor and TerraMind embeddings for segmentation and elevation regression. This direction is still open-ended, so the first goal is to test a few reasonable fusion strategies and identify promising ones.

Motivation

Thor and TerraMind may encode complementary information. A better fusion strategy could improve both segmentation IoU and elevation RMSE compared with using either embedding source alone or using a simple fusion baseline.

Tasks

• Review the current embedding inputs and fusion approach.
• Design a small set of fusion strategies, for example:
  • simple concatenation
  • weighted sum
  • separate encoder branches + late fusion
  • attention/gating-based fusion
• Implement the most feasible 2–3 strategies first.
• Run controlled experiments using the same training/validation setup.
• Compare against single-source and existing fusion baselines.

Deliverables

• Code/config support for tested fusion strategies.
• Experiment results comparing fusion methods.
• Short analysis of which fusion method looks most promising.

Acceptance Criteria

• At least 2 fusion strategies are implemented and evaluated.
• Results include IoU and RMSE comparison.
• The issue concludes with a recommendation for the next fusion direction.

Notes

Please keep experiments reproducible:

• record config changes
• record checkpoint path
• report validation IoU and RMSE
• attach qualitative visualization if possible