Landmine Detection with Deep Learning

Code for my BEng thesis at the University of Edinburgh (2024): Deep Learning Driven Landmine Detection: Generalizing Real World Data from Synthetic GPR Training Sets.

Trained a convolutional neural network entirely on synthetic ground penetrating radar (GPR) data and tested it against real radar scans. The model achieved 80% accuracy on individual real world scans and 100% with weighted averaging across multiple scans.

IMechE Best Student Award

Pipeline

1. model_generate.py — gprMax input script. Generates randomised FDTD simulations of a GSSI 2 GHz antenna scanning over soil with buried PMN landmines, bullet casings, and rocks at varying depths and soil conditions.

2. process.py — Processes raw A-scans from the simulations: time-zero correction, normalisation, background removal, resampling, and dataset assembly with classification labels.

3. grid_search.py — Neural architecture search using AutoKeras. Runs 600 trials with up to 750 epochs to find the best model architecture.

4. retrain.py — Retrains the best model with weight resetting on specific layers, early stopping, and accuracy/loss plotting.

5. evaluate.py — Evaluates the trained model: ROC curves, confusion matrices, per-data-type accuracy breakdown (Target / Mix / False Alarm), and real data testing.

Config files

• dataset_definitions.txt — Simulation parameters (domain size, scan counts, training split)
• processing_definitions.txt — Processing parameters (PCA variance, scaling, batch config)

Dependencies

• Python 3.8+
• gprMax (FDTD simulation)
• TensorFlow / Keras
• AutoKeras
• NumPy, Pandas, Scikit-learn, SciPy, Matplotlib, Seaborn, h5py

Thesis

Full thesis available at louisbrouwer.com/thesis.html