Multi-resolution reconstruction of longitudinal streambed footprints using embedded sparse convolutional autoencoders

This study introduces an embedded convolutional autoencoder (CAE) architecture designed for the multi-resolution reconstruction of longitudinal streambed footprints as sparse heatmaps. Three standalone but interrelated CAEs are trained to achieve double-upsampling, enhancing the heatmaps’ spatial resolution and data measurement resolution simultaneously. Transfer learning improves model training efficiency by incorporating a trained model into a larger model at the next level. Cascading the CAEs facilitates a direct pathway for enhancing data quality from the coarsest inputs and recovering fine-grained patterns. Systematic evaluations prove the CAEs’ reliability in working individually and collectively. Robustness analyses demonstrate the model’s ability to retain field reconstructive quality when subjected to various corrupted inputs, including bulk data loss and spiky noise interference with local measurements at different streambed sections. The model’s capacity benefitted from including attention mechanisms (convolutional block attention modules, CBAM) and the adaptive training strategy using crafted loss functions, ensuring efficient extraction and learning of sparse dense patterns and fast reconstruction of physically sound fields. The model architecture’s flexibility and scalability are highlighted, proving it suitable for more complex geophysical systems with higher dimensions. The proposed embedded CAE architecture provides a foundational tool for creating digital surrogates of river courses and similar entities, which often involve inherently sparsely distributive data in both spatial and temporal domains.