🛰️ Automatic Modulation Recognition (AMR) for DVB-S2X Waveforms

Beginner ML Project | Classification using Random Forest | Dataset: RadioML 2016.10A

---

📄 Abstract

Satellite communication systems operating under the DVB-S2X standard dynamically switch between multiple modulation schemes — including QPSK and 8PSK — based on prevailing channel conditions. Accurate and automatic identification of these modulation schemes at the ground receiver is essential for reliable signal synchronization. Traditional manual and rule-based modulation recognition methods are computationally expensive and error-prone under noisy conditions.

This project presents an Automatic Modulation Recognition (AMR) system that uses a Random Forest classifier to identify DVB-S2X modulation schemes from raw In-phase and Quadrature (I/Q) signal samples. The system is trained and evaluated on the RadioML 2016.10A dataset — a synthetic dataset generated using GNU Radio, containing 11 modulation types across SNR levels from -20 dB to +18 dB, with 1,000 samples per modulation-SNR pair. Statistical features including amplitude, phase, and kurtosis are extracted from I/Q samples and used for classification. Model performance is evaluated using classification accuracy and confusion matrices across multiple SNR levels.

---

📁 Project Structure

AMR-DVB-S2X/
│
├── data/
│   └── RML2016.10a.pkl          # Downloaded dataset (place here)
│
├── src/
│   ├── load_data.py             # Load and filter dataset
│   ├── feature_extraction.py   # Extract features from I/Q samples
│   ├── train_model.py          # Train Random Forest classifier
│   └── evaluate.py             # Accuracy + confusion matrix
│
├── models/
│   └── random_forest_model.pkl  # Saved trained model
│
├── results/
│   ├── confusion_matrix.png     # Output plot
│   └── accuracy_per_snr.png    # Output plot
│
├── requirements.txt             # Python dependencies
└── README.md                    # This file

---

🗄️ Dataset

Field	Detail
Name	RadioML 2016.10A
Source	DeepSig Inc. (GNU Radio synthetic)
Modulation Types	11 total (we use QPSK, 8PSK)
SNR Range	-20 dB to +18 dB (2 dB steps)
Samples per class/SNR	1,000
Total Samples	220,000
Format	.pkl (Python pickle)
Signal Format	Raw I/Q (In-phase & Quadrature) samples

📥 Download

Direct download:
http://opendata.deepsig.io/datasets/2016.10/RML2016.10a.tar.bz2

After downloading, extract and place RML2016.10a.pkl inside the data/ folder.

---

⚙️ Tech Stack

Layer	Technology
Language	Python 3.8+
Data Loading	pickle, numpy
Feature Extraction	numpy, scipy
ML Model	scikit-learn (Random Forest)
Evaluation	scikit-learn, matplotlib, seaborn
Model Saving	joblib

---

🚀 Setup & Installation

Step 1 — Clone the repository

git clone https://github.com/your-username/AMR-DVB-S2X.git
cd AMR-DVB-S2X

Step 2 — Install dependencies

pip install -r requirements.txt

Step 3 — Download and place dataset

# Download
wget http://opendata.deepsig.io/datasets/2016.10/RML2016.10a.tar.bz2

# Extract
tar -xvjf RML2016.10a.tar.bz2

# Move to data folder
mv RML2016.10a.pkl data/

Step 4 — Run the pipeline

python src/load_data.py
python src/feature_extraction.py
python src/train_model.py
python src/evaluate.py

---

🔬 ML Pipeline

Raw Dataset (RML2016.10a.pkl)
          ↓
  Load & Filter
  (Keep only QPSK, 8PSK)
          ↓
  Feature Extraction from I/Q
  - Mean Amplitude
  - Variance of Amplitude
  - Kurtosis
  - Mean Phase
  - Variance of Phase
  - Max Amplitude
          ↓
  Train/Test Split (80/20)
          ↓
  Train Random Forest Classifier
          ↓
  Evaluate:
  - Overall Accuracy
  - Accuracy per SNR level
  - Confusion Matrix
          ↓
  Save Model → models/random_forest_model.pkl

---

📊 Expected Results

Metric	Expected Value
Accuracy at high SNR (≥ 10 dB)	> 90%
Accuracy at low SNR (≤ -10 dB)	50–65%
Overall Accuracy	~75–85%
Classes	QPSK, 8PSK

---

📦 requirements.txt

numpy
scipy
scikit-learn
matplotlib
seaborn
joblib
pickle5

---

🤖 Using Claude Code for Further Development

This project is designed to be extended using Claude Code — Anthropic's agentic CLI coding tool.

What is Claude Code?

Claude Code is a command-line tool that understands your codebase and helps you develop faster using natural language. It can read files, write code, run commands, and handle Git workflows autonomously.

Installation

macOS / Linux / WSL:

curl -fsSL https://claude.ai/install.sh | bash

Windows PowerShell:

irm https://claude.ai/install.ps1 | iex

⚠️ Requires a Claude Pro, Max, Team, Enterprise, or Console account. The free plan does not include Claude Code access.

After installation, log in:

claude
# Follow the browser prompts to authenticate

Using Claude Code on This Project

Navigate to the project folder and launch Claude Code:

cd AMR-DVB-S2X
claude

💬 Example Prompts for Further Development

Explore the dataset:

"Load the RML2016.10a.pkl file and show me the structure,
available modulation types, and sample shapes"

Build feature extraction:

"Write a Python script that extracts amplitude mean, variance,
kurtosis, and phase statistics from I/Q signal samples in the dataset"

Train the model:

"Train a Random Forest classifier on the extracted features,
split 80/20 train/test, and print the accuracy"

Add more modulations:

"Update the pipeline to also include AM-DSB and WBFM
modulations from the dataset alongside QPSK and 8PSK"

Try other models:

"Replace the Random Forest with an SVM classifier using
an RBF kernel and compare accuracy with Random Forest"

Plot results:

"Generate a confusion matrix heatmap and an accuracy vs SNR
line plot, and save them to the results/ folder"

Improve accuracy:

"Add more features like spectral kurtosis and autocorrelation
to the feature extraction script and retrain the model"

Save and load the model:

"Save the trained Random Forest model using joblib to
models/random_forest_model.pkl and write a load function"

---

📅 Development Roadmap

Week	Task	Status
Week 1	Project planning, problem definition, tooling setup	✅ Done
Week 2	Functional modeling, use case diagrams, actor identification	✅ Done
Week 3	Dataset download, exploration, preprocessing	🔲 Pending
Week 4	Feature extraction from I/Q samples	🔲 Pending
Week 5	Model training (Random Forest)	🔲 Pending
Week 6	Evaluation, confusion matrix, accuracy vs SNR	🔲 Pending
Week 7	Model optimization, compare with SVM/KNN	🔲 Pending
Week 8	Final report and documentation	🔲 Pending

---

👥 Team Roles

Role	Responsibility
Project Manager	Planning, tracking, documentation
Signal Processing Engineer	Dataset handling, feature extraction
ML Engineer	Model training, evaluation, optimization
Testing Lead	Validation, results verification

---

📚 References

• T.J. O'Shea, "RadioML2016.10A", DeepSig Inc., 2016
• ETSI EN 302 307-2 — DVB-S2X Standard Specification
• Scikit-learn Documentation: https://scikit-learn.org
• Claude Code Documentation: https://docs.anthropic.com/en/docs/claude-code/overview

---

📝 Notes for Beginners

• You do not need to understand all the signal processing theory to run this project
• Claude Code can explain any part of the code to you — just ask it in plain English
• Start by running the scripts one by one and reading the output
• If anything breaks, paste the error into Claude Code and ask it to fix it

---

Project developed as part of academic coursework on Automatic Modulation Recognition for satellite communication systems.