PhD Advisor: Professor Matthew McCabe

Abstract:
Reliable marine measurements are difficult to obtain when refraction and caustics distort the view of the seabed and rare events such as oil spills leave few labelled examples. Combining successive drone frames recovered submerged object shapes with errors below about 5% under favorable low-altitude conditions, but the method required high-frame-rate video and substantial processing time. Those demands motivated a shift toward restoring a single image in under a second. Because paired training images were scarce, KAUSTics was developed to provide about 9,000 semi-synthetic image pairs for training a diffusion model. The same idea of generating training examples carried over to oil-spill monitoring, where at most 50 labelled radar patches were used to generate synthetic data for segmentation models tested on real radar imagery. Lastly, the focus returned to the seabed to evaluate whether multimodal large language models could both assess and restore shallow-water drone imagery. The best of five models reached 65.4% balanced accuracy in detecting distortions, but the models struggled to restore submerged scenes faithfully.

Bio:
Dario received his M.Sc. in Space Engineering from the University of Naples Federico II, Italy, in 2022, and is currently pursuing a Ph.D. in Environmental Science and Engineering at KAUST, Saudi Arabia. He has contributed to projects spanning CubeSat missions, UAV-based environmental monitoring, and synthetic dataset generation for scientific imaging applications. Before his Ph.D., he worked as a satellite operator at ICEYE in Poland, operating synthetic aperture radar (SAR) satellites.