PakFlood-LoRA

Flood mapping from Sentinel-1 radar with a frozen TerraMind v1 base encoder, a rank-32 LoRA adapter (5.7 M trained encoder weights) and a UNet decoder, trained on GEOID-Flood. Every 10 m pixel of a pre/post radar pair is labelled as dry land, permanent water or flood.

  • Official GEOID-Flood test split (2,417 tiles): flood IoU 0.525, equal to full fine-tuning of all 101 M weights (frozen encoder 0.433, LoRA r16 0.493).
  • Pakistan official test tiles (EMSR838, August 2025, 15 tiles): flood IoU 0.786.
  • The sample tiles in this Space are those 15 official test tiles; the model never saw them in training.

MS AI deep learning course project, NED University of Engineering & Technology. Author: Muhammad Shahwar Uddin. Supervisor: Dr. Sharaf Hussain.

Data: GEOID-Flood v1.1 (Chiriaco et al., Fondazione LINKS, 2026, CC BY 4.0), built from Copernicus Emergency Management Service activations. Model: TerraMind v1 base (Jakubik et al., IBM and ESA, 2025) through TerraTorch; LoRA (Hu et al., 2022) via Hugging Face PEFT.

Live demo: open PakFlood_LoRA_Colab_demo.ipynb in Google Colab (direct link), run the three cells, and use the public Gradio link it prints. Or run locally: pip install -r requirements.txt && python app.py.

Project code, results and figures: see code/, results/, figures/ and CODE_README.md in this repository (training and evaluation pipeline, per-tile results, all figures).

Source code on GitHub: https://github.com/ShahwarUddin/pakflood-lora

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for msherry/pakflood-lora

Adapter
(2)
this model

Dataset used to train msherry/pakflood-lora