Satellite-based oil spill detection, risk analysis, Arabic/English reports, and an intelligent assistant — built with DeepLabV3+, FastAPI, React, and Qwen2.5 (LoRA).
This repository contains source code only. Large datasets, trained weights, database dumps, and runtime uploads are not on GitHub; download them from Hugging Face (below).
For full project structure and setup, see PROJECT_DOCUMENTATION.md.
Backend
cd backend-rana-clean
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
# Configure .env (DB paths, model paths, API keys)
uvicorn main:app --reload --port 8000Frontend
cd oil-spill-app
npm install
npm run devTrained models and the tokenized dataset are hosted on Hugging Face under @ra-hk1.
| Resource | Description | Link |
|---|---|---|
| SpillVision DeepLabV3+ Keras Models | Semantic segmentation for oil spill detection from satellite imagery | spillvision-deeplab-keras |
| SpillVision Qwen Oil LoRA Adapter | Fine-tuned LoRA adapter for oil spill report generation | spillvision-qwen-oil-lora |
| Qwen2.5 0.5B Instruct (local base) | Base model used with the SpillVision LoRA adapter | spillvision-qwen2.5-0.5b-instruct-local |
| Resource | Description | Link |
|---|---|---|
| Oil Spill Tokenized Dataset | Tokenized dataset for fine-tuning and NLP tasks | oil-spill-tokenized-dataset |
pip install huggingface_hub
# DeepLab weights (example)
huggingface-cli download ra-hk1/spillvision-deeplab-keras --local-dir ./models/deeplab
# Qwen base + LoRA (example)
huggingface-cli download ra-hk1/spillvision-qwen2.5-0.5b-instruct-local --local-dir ./oil_llm_reporter/models/Qwen2.5-0.5B-Instruct
huggingface-cli download ra-hk1/spillvision-qwen-oil-lora --local-dir ./oil_llm_reporter/oil_qwen_lora_adapterPoint your .env / backend config to the downloaded paths (see PROJECT_DOCUMENTATION.md).
- Satellite imagery (
Oil/), user uploads, generated reports full_database.sqland local vector DBs.envsecrets
These stay local or on Hugging Face; see .gitignore.