CephasNzaana/KigeziOreSight

An open-source exploration into using computer vision to identify potential iron ore deposits in my home region of Kabale, Western Uganda. # Kigezi OreSight: An Open-Source AI Exploration **An open-source exploration into using computer vision to identify potential iron ore deposits in my home region of Kabale, Western Uganda.** --- ### 🌱 The Spark The story of this project started with these rocks right here. With the growing momentum around developing the iron ore industry in my home region, I became fascinated by a simple question: can modern AI help us understand our own land better? Instead of just talking about it, I went out and got my first data points. This project is my journey to answer that question, and I'm building it entirely in the open. ### 💡 The Idea My goal is to build **Kigezi OreSight**, a deep learning system that will learn to analyze publicly available satellite imagery and create a predictive heat map of potential iron ore deposits. By training a model to recognize the unique patterns of mineral-rich land, I hope to create an open-source tool that can help guide and de-risk initial exploration efforts for our local community. This isn't about creating a commercial product; it's about exploring the art of the possible and sharing the knowledge with everyone. --- ### ✨ The Vision for the Live Demo The ultimate deliverable will be a simple, interactive web application. Imagine this: you'll be able to select an area on a map of Kabale, and the AI model will overlay a beautiful heatmap in real-time. *(This will be a screen recording of the final Streamlit application.)* --- ### ⚙️ How It Will Work: The Technical Journey 1. **Data Collection:** I will gather multispectral satellite images from the European Space Agency's Sentinel-2 satellite via the Sentinel Hub API. 2. **Data Labeling:** I'll use existing geological survey maps of the Kigezi region to create labeled training data, marking areas of known iron ore presence. 3. **Model Training:** A custom U-Net segmentation model, built with **PyTorch**, will be trained on this labeled data. The model will lear …