Convolutional Layer Optimisation for GPR Anomaly Detection: Preliminary Experimental Evaluation

Convolutional Layer Optimization for GPR Anomaly Detection: Preliminary Experimental Evaluation

This study investigates the impact of convolutional layer depth on anomaly detection performance in Ground Penetrating Radar (GPR) data using unsupervised Convolutional Autoencoders (CAE). While GPR provides non-invasive subsurface imaging for infrastructure maintenance, interpretation remains largely manual and subjective. CAEs offer an efficient approach by learning from anomaly-free data alone, eliminating the need for labelled examples. Multiple architectures ranging from lightweight (VGG8-equivalent) to deep configurations (VGG16, VGG19, ResNet50) were systematically evaluated using road survey datasets to assess the relationship between architectural complexity and detection sensitivity. Preliminary results indicate that complex models like ResNet50 unexpectedly underperformed compared to simpler VGG-based architectures, with VGG19 and VGG16 showing similar detection capabilities despite their architectural differences. These findings suggest a threshold of complexity beyond which additional convolutional layers provide diminishing returns for GPR anomaly detection. This has significant implications for field applications, as simpler models require fewer computational resources when processing large volumes of survey data. While lacking comprehensive quantitative evaluation, this research establishes a framework for optimizing CAE architectures specifically for GPR interpretation rather than generic object recognition tasks.