IEEE Access (Jan 2024)
Uncertainty-Aware Aerial Coastal Imagery Pattern Recognition Through Transfer Learning With ImageNet-1K Variational Embeddings
Abstract
Monitoring coastal landscape changes is crucial for understanding the impact of extreme weather events on coastal communities as well as longer term trends in landscape change. However, continuous monitoring remains challenging given the dynamic nature of these environments as well as the diversity of landscape characteristics (for example, not all beaches have the same features). While previous work has primarily focused on semantic segmentation of satellite imagery, oblique aerial imagery offers superior temporal and spatial resolution for coastal monitoring, which can help identify regions of greatest change. However, the variability in image quality and lack of consistent coverage hinder the application of semantic segmentation to oblique aerial imagery across broad geographic regions. In this study, we demonstrate the effectiveness of whole-image classification using transfer learning on a novel dataset of 8,800 oblique coastal images from the U.S. East, Gulf, and West Coasts. We evaluate the performance of fifteen convolutional neural networks and two vision transformers, including both deterministic and probabilistic models, with all networks achieving over 90% accuracy. After pre-training Bayesian variants of ResNet50 on ImageNet-1K and transferring them to our coastal dataset, we perform uncertainty decomposition analysis to enhance model explainability. We make our pre-trained Bayesian ImageNet-1K checkpoints, fine-tuned Bayesian checkpoints, and curated dataset publicly available to facilitate reproducibility and further research in this domain. The resulting models could later be used to classify and map coastal landscapes on a global level, which would allow for longer term determination of landscape change associated with climate variability.
Keywords
