An open-source exploration into using computer vision to identify potential iron ore deposits in my home region of Kabale, Western Uganda.
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.
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 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.)
- Data Collection: I will gather multispectral satellite images from the European Space Agency's Sentinel-2 satellite via the Sentinel Hub API.
- 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.
- Model Training: A custom U-Net segmentation model, built with PyTorch, will be trained on this labeled data. The model will learn to segment any given satellite image into areas of high, medium, and low probability for iron ore deposits.
- Inference & API: The trained model will be served via a lightweight FastAPI application, which will expose an endpoint to process new satellite images.
- Visualization: A simple web interface built with Streamlit will allow users to interact with the model, select an area, and view the resulting heat map visualization.
This project will follow a standard structure for a production-ready data science project.
kigezi-oresight/
│
├── app/ # The Streamlit web application
│ └── app.py
│
├── data/
│ ├── raw/ # Will contain raw satellite image downloads
│ │ └── sample_kabale_image.png # A sample image for testing
│ └── processed/ # Labeled and processed images for training
│
├── notebooks/ # Jupyter notebooks for exploration and analysis
│ └── 01_initial_data_exploration.ipynb
│
├── src/ # Source code for the project
│ ├── config.py # Project configuration variables
│ ├── data_loader.py # Scripts to download and process data
│ ├── model.py # The PyTorch model architecture (U-Net)
│ └── train.py # The main script for training the model
│
├── tests/ # Unit and integration tests
│ └── test_model.py
│
├── .gitignore
├── Dockerfile # To containerize the FastAPI/Streamlit app
├── LICENSE
├── README.md # You are here!
└── requirements.txt
To get a local copy up and running, follow these steps.
Prerequisites:
- Python 3.10+
- Poetry (for dependency management) or pip
Installation & Training:
# Clone the repository
git clone [https://github.com/your-username/kigezi-oresight.git](https://github.com/your-username/kigezi-oresight.git)
cd kigezi-oresight
# Install dependencies
pip install -r requirements.txt
# Run the training script (will use sample data in `data/raw`)
python src/train.pyRunning the Demo App:
# Run the Streamlit web application
streamlit run app/app.pyThis is an open project, and collaboration is highly encouraged. Please see CONTRIBUTING.md for guidelines on how to get involved.
# ========================================================================
# File: kigezi-oresight/app/app.py
# Purpose: The main file for the Streamlit web demo application.
# ========================================================================
import streamlit as st
import numpy as np
from PIL import Image
import os
# Placeholder for a function that would load our trained PyTorch model
def load_model():
# In a real app, this would load a .pt file
st.session_state['model_loaded'] = True
return "Dummy Model"
# Placeholder for a function that processes an image and returns a heatmap
def predict(model, image):
# Simulate a model prediction
# In a real app, this would run the image through the U-Net model
st.session_state['prediction_run'] = True
heatmap = np.random.rand(*image.shape[:2])
return (heatmap * 255).astype(np.uint8)
st.set_page_config(layout="wide", page_title="Kigezi OreSight")
# --- Page Header ---
st.title("🛰️ Kigezi OreSight")
st.markdown("An open-source exploration into using AI to identify potential iron ore deposits from satellite imagery.")
# --- Main Interface ---
st.sidebar.header("Controls")
uploaded_file = st.sidebar.file_uploader("Upload a satellite image of the Kigezi region", type=["jpg", "png"])
col1, col2 = st.columns(2)
with col1:
st.subheader("Input Image")
if uploaded_file is not None:
image = Image.open(uploaded_file)
st.image(image, caption="Uploaded Satellite Image", use_column_width=True)
else:
st.info("Please upload an image to begin analysis.")
# Display a sample image if it exists
sample_image_path = os.path.join('data', 'raw', 'sample_kabale_image.png')
if os.path.exists(sample_image_path):
sample_image = Image.open(sample_image_path)
st.image(sample_image, caption="Sample Satellite Image of Kabale", use_column_width=True)
else:
st.warning("Sample image not found. Please create 'data/raw/sample_kabale_image.png'")
with col2:
st.subheader("AI Prediction Heatmap")
if uploaded_file is not None:
with st.spinner('Koko the Pangolin is analyzing the terrain... Please wait.'):
model = load_model()
pil_image = Image.open(uploaded_file).convert('RGB')
np_image = np.array(pil_image)
heatmap_data = predict(model, np_image)
# Create a red overlay effect
heatmap_colored = Image.new("RGBA", pil_image.size)
overlay_color = (255, 0, 0, 100) # Red with some transparency
for x in range(pil_image.width):
for y in range(pil_image.height):
if heatmap_data[y, x] > 180: # Higher threshold for a less dense map
heatmap_colored.putpixel((x, y), overlay_color)
# Composite the original image with the heatmap
final_image = Image.alpha_composite(pil_image.convert("RGBA"), heatmap_colored)
st.image(final_image, caption="Predicted Iron Ore Probability (Red = High)", use_column_width=True)
st.success("Analysis complete!")
else:
st.info("Prediction will appear here once you upload an image.")
# --- Footer ---
st.markdown("---")
st.markdown("Built with ❤️ in Kabale. An open-source project by Cephas Nzaana.")
```python
# ========================================================================
# File: kigezi-oresight/src/train.py
# Purpose: A placeholder for the main model training script.
# ========================================================================
import torch
# from model import OreSightUNet # Assuming model architecture is in model.py
# from data_loader import get_dataloader # Assuming data loader is set up
def train_model():
print("Starting model training process...")
# 1. Load configuration
# config = load_config('config.yaml')
print("Configuration loaded.")
# 2. Get data loaders
# train_loader = get_dataloader(config.data.train_path)
# val_loader = get_dataloader(config.data.val_path)
print("Data loaders created.")
# 3. Initialize model, optimizer, and loss function
# model = OreSightUNet(in_channels=3, out_channels=1)
# optimizer = torch.optim.Adam(model.parameters(), lr=config.training.learning_rate)
# criterion = torch.nn.BCEWithLogitsLoss()
print("Model and optimizer initialized.")
# 4. Training loop
print("Entering training loop...")
# for epoch in range(config.training.num_epochs):
# --- Training Step ---
# for images, masks in train_loader:
# model.train()
# ... training logic ...
# --- Validation Step ---
# for images, masks in val_loader:
# model.eval()
# ... validation logic ...
# print(f"Epoch {epoch+1} completed.")
print("Model training complete!")
# 5. Save the trained model
# torch.save(model.state_dict(), 'kigezi_oresight_model.pt')
print("Trained model saved to 'kigezi_oresight_model.pt'")
if __name__ == "__main__":
train_model()
```python
# ========================================================================
# File: kigezi-oresight/requirements.txt
# Purpose: Lists the Python dependencies for the project.
# ========================================================================
streamlit
torch
torchvision
numpy
Pillow
# For data handling (add as needed)
# pandas
# geopandas
# rasterio