geomtl-vanilla

GeoMTL Custom Model (Trained from Scratch)

This repository contains the model weights for a custom multi-modal geospatial architecture (GeoMTL) trained completely from scratch.

🏗️ Architecture Details

Unlike fine-tuned or adapted foundation models, this checkpoint represents an end-to-end custom configuration featuring:

• Prithvi Vision Encoder: Processes multi-temporal satellite imagery.
• UperNet Decoder: used for generating a multiclass semantic segmentation masks.
• Q-Former (Querying Transformer): Bridges the visual and text modalities.
• Decoder: TinyGPT-2 decoder used for generating descriptive text and answering visual queries.

🗂️ Checkpoint Information

• File: mtl_scratch_best.pth
• Training Type: Scratch training (no pre-trained weights used for initialization of decoders).
• Task Domain: Remote sensing image captioning, segmentation alignment, and VQA.

📝 Citation

If you use this model checkpoint in your research, please cite our GeoMTL IGARSS paper:

@inproceedings{geomtl2026,
    author    = {[Authors]},
    title     = {[Title]},
    booktitle = {IEEE International Geoscience and Remote Sensing Symposium (IGARSS)},
    year      = {2026},
    note      = {Forthcoming}
}