BFDMCA: A Novel Approach for Classification of GPR Images with SHAP Explanations

Accurate classification of clutter-free images is crucial for precise analysis in ground penetration radar (GPR) imaging. However, significant challenges arise due to limited annotated datasets and the demand for efficient models capable of handling complex subsurface data. Existing methods often struggle to differentiate subtle features critical for accurate classification, where fine details are essential. To address these challenges, the proposed model employs an optimized architecture combined with a multiscale attention mechanism, achieving improved classification accuracy and robustness in GPR imaging. The approach begins with the collection of synthetic and experimental GPR images, followed by data augmentation through conditional generative adversarial networks (CGANs), which enrich data diversity and volume, strengthening classification accuracy. For classification, the model employs the Bidirectional Focal Tversky Dynamic Multiscale Convolutional Neural Network based on AlexNet (BFDMCA), which integrates multiscale convolutional feature extraction with Bidirectional Long Short-Term Memory (Bi-LSTM) and attention mechanisms to capture complementary spatial and sequential information from GPR images. Explainability is supported by Shapley additive explanation (SHAP), which provides insights into critical features and regions, ensuring model transparency and reliability. A lightweight web-based application is also implemented to facilitate real-time visualization of predictions, supporting practical use of the proposed framework.