Unexploded ordnance (UXO) presents ongoing risks to shallow marine environments and maritime infrastructure. Remediation of in-water UXO requires localizing and characterizing objects in cluttered environments where the marine environment obfuscates detection because of biofouling, sedimentation and other changes in UXO appearance and structure. The detection of UXO in littoral zones 10m and shallower, remains especially challenging owing to the limited ability of acoustic methods to operate over large areas in such shallow regimes and the effects of ocean wave distortion and caustics on optical sensing methods from aircraft or spacecraft. This project explores using several novel airborne aquatic remote sensing technologies for the automated airborne detection and localization of underwater military munitions in a complex marine environment. 

Here, multispectral (444-842nm) airborne Fluid Lensing, an airborne remote sensing technology capable of optical imaging through ocean wave distortion without refractive or caustic effects, is applied as well as active MiDAR Fluid Lensing, spanning ultraviolet to visible optical bands (375-675nm), to image underwater munitions of varying colors and conditions ranging in size from 2cm to 10cm in width and maximal linear dimension from 25.5cm to 66cm over a large marine environment using unpiloted aerial vehicles. Inert munitions were deployed underwater at the University of Miami’s Florida Keys Broad Key Research Station under a National Oceanographic & Atmospheric Administration/Florida Keys National Marine Sanctuary permit over a large area replete with high anthropogenic and natural clutter. Over the next two months, the targets were left to biofoul and accumulate sediment. Airborne Fluid Lensing campaigns were then conducted to detect and localize the targets prior to manually removing them.

A YOLO-based model is trained on 2,700 artificially augmented samples from nine UXO targets in the field site to detect the inert munitions from the airborne datasets. All 14 deployed UXO were detected and localized in three different Fluid Lensing modalities (passive 3-band high resolution (0.5-1cm ground sample distance [GSD]), passive 10-band multispectral (1-3cm GSD), and active 8-band MiDAR (0.3-1cm GSD)) at previously unknown locations. This instance–segmentation detector achieves high precision with moderate recall upon convergence (~200 epochs: ! ≈ 0.95, ' ≈ 0.71, mAP@0.5 ≈ 0.775, mAP@0.5: 0.95 ≈ 0.488), and cross-validated with a F1 (Dice) score within 0.83 – 0.89.

Active 8-band MiDAR Fluid Lensing was found to outperform passive 3-band and 10-band multispectral Fluid Lensing at comparable spatial resolution with precision values in the 0.8-0.9, 0.73-0.89, and 0.71–0.74 ranges respectively. Indeed, several active water-penetrating MiDAR bands were identified for these UXO targets that result in higher precision, even in the presence of decoy targets placed next to target UXO. Together, these results suggest airborne active MiDAR and passive Fluid Lensing combined with a pretrained convolutional neural network are viable solutions to large-scale UXO detection in cluttered marine environments; however additional campaigns and UXO target types are needed to scale the method more broadly and increase detector precision while reducing false positive rates across more heterogeneity in depth and benthic substrates. (Project Completion - 2025)