Physics-Informed Genetic-Conditioned Network for Glacial Landform Classification

Recent advances in deep learning, particularly convolutional neural networks (CNNs), together with the increasing availability of high-resolution multibeam bathymetry data, have made automated classification of glacial landforms increasingly feasible. However, robust and scalable classification remains challenging. The main challenge is not only computational scale, but the mismatch between data-driven learning and the physics-governed nature of glacial terrain. To address this gap, we propose a physics-informed deep learning framework for automated classification of submarine glacial landforms from multibeam bathymetry. Unlike conventional CNNs that rely purely on data-driven features, our approach integrates physically meaningful constraints reflecting glacial geomorphology. Specifically, we develop a Physics-Informed Genetic-Conditioned Network (PI-GCNet) by integrating a genetic-conditioned layer into a standard CNN and introducing a physics-guided loss that enforces geomorphological consistency. Each landform class is represented by a learnable genome vector initialized from class-wise statistics of depth, slope, and curvature and trained jointly with the embeddings. A genetic attraction and repulsion mechanism structures the latent space, and classification is performed via an energy-based distance between embeddings and genomes. We further optimize a composite objective combining genetic Cross-Entropy, genome regularization for stable and interpretable representations, and a class-wise slope deviation loss penalizing departures of predicted mean slope from expected class-specific values. Together, these components enhance robustness and interpretability, enabling scalable and physically consistent mapping of submarine glacial landforms. We validate the proposed model using high-resolution multibeam bathymetry acquired in northern Greenland. We prepared a dataset of 1515 samples across 10 classes of glacial and glacimarine features identified through manual interpretation. Model inputs include bathymetry, slope, aspect, profile curvature, tangential curvature, and the rasterized outline of each feature. In addition, we compare PI-GCNet with state-of-the-art deep learning baselines to demonstrate reliability and improved generalization.