Cross-Modal Reconstruction Objectives in Multimodal Geospatial Foundation Models Can Appear to Learn Without Learning — Code and Figures
Sardor KarimovFergana State Technical University
2026
ABI
Annotatsiya
Code and figures accompanying the manuscript "Cross-Modal Reconstruction Objectives in Multimodal Geospatial Foundation Models Can Appear to Learn Without Learning" (submitted to IEEE Geoscience and Remote Sensing Letters, 2026). Contains the source-level fix (robust standardisation and Huber loss for the cross-modal term, per-direction validation logging), the direction-resolved evaluation with causal ablations, and the figure-generation scripts.
Hali tarjima qilinmagan
Identifikatorlar
Iqtiboslar va manbalar
0 ta iqtibos0 ta foydalanilgan manba