Smart Environmental Monitoring of Marine Pollution using Edge AI
Mohamed Moursi, Norbert Wehn, Bilal Hammoud · 2025
Read Smart Environmental Monitoring of Marine Pollution using Edge AI by Mohamed Moursi, Norbert Wehn, Bilal Hammoud (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-21759.txt.
Introduction
Mohamed Moursi, Norbert Wehn, and Bilal Hammoud put an oil spill detector on a drone in this 2025 paper. Spills at sea demand fast mapping of where the oil is and how thick it lies, since thick patches are skimmed first. Neural networks can read that from radar, but they are usually too heavy for small aircraft.
The team shrank a standard segmentation network hard, cutting blocks and channels until it was about two hundred seventy times smaller. Trained on synthetic radar scenes, the small model still scored about seventy-nine percent overlap on spill outlines and judged slicks up to ten millimeters thick. It runs on a programmable chip aboard the drone at about two watts, close to real time. The authors present the system as proof that sharp environmental sensing can live at the edge, on the device itself.
Catalog record
| CREATOR | Mohamed Moursi, Norbert Wehn, Bilal Hammoud |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 17 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-21759
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Smart Environmental Monitoring of Marine Pollution using Edge AI Thanks: The authors thank the funding support of the Carl Zeiss Stiftung under the Sustainable Embedded AI project (P2021-02-009). Mohamed Moursi, Norbert Wehn, Bilal Hammoud Abstract Oil spill incidents pose severe threats to marine ecosystems and coastal environments, necessitating rapid detection and monitoring capabilities to mitigate environmental damage. In this paper, we demonstrate how artificial intelligence, despite the inherent high computational and memory requirements, can be efficiently integrated into marine pollution monitoring systems. More precisely, we propose a drone-based smart monitoring system leveraging a compressed deep learning U-Net architecture for oil spill detection and thickness estimation. Compared to the standard U-Net architecture, the number of convolution blocks and channels per block are modified. The new model is then trained on synthetic radar data to accurately predict thick oil slick thickness up to 10 mm. Results show that our optimized Tiny U-Net achieves superior performance with an Intersection over Union (IoU) metric of approximately 79%, while simultaneously reducing the model size by a factor of ∼ \sim 269x compared to the state-of-the-art. This significant model compression enables efficient edge computing deployment on field-programmable gate array (FPGA) hardware integrated directly into the drone platform. Hardware implementation demonstrates near real-time thickness estimation capabilities with a run-time power consumption of approximately 2.2 watts. Our findings highlight the increasing potential of smart monitoring technologies and efficient edge computing for operational characterization in marine environments. I Introduction Oil spills represent a significant environmental hazard with high impacts on marine ecosystems and wildlife populations, constituting one of the most severe anthropogenic disturbances to maritime environments. It is imperative from both economic and ecological perspectives to mitigate the deleterious effects of these spills [ 1 ] . While state-of-the-art research has predominantly focused on oil detection capabilities, a detailed characterization of additional spill parameters like the spatial distribution of oil slick thickness helps implement effective containment strategies [ 2 ] . The development of advanced monitoring systems that spontaneously provide such detailed characterization enables more precise intervention approaches and ultimately improves remediation outcomes. This requires careful consideration of the selected platform and sensor as potential candidates for specific monitoring functionalities [ 3 ] . Drone platforms have received attention for environmental monitoring in the last years [ 4 ] , considering their relatively low cost, high resolution, and fast intervention. For oil spill monitoring, several drone-based approaches targeted oil detection using visual [ 5 ] , and infrared cameras [ 6 ] . Hyperspectral images were also analyzed to provide thickness estimation [ 7 ] . But the previous sensors cannot operate during the full daytime and in different weather conditions. For this, the radar sensor is more prominent for monitoring purposes, which explains why synthetic aperture radar (SAR) imaging have been extensively used for oil spill detection [ 8 , 9 , 10 ] . However, even when processing SAR imaging with advanced deep-learning models, the effectiveness of this technique is reduced in calm ocean conditions due to the loss in the radar backscattering [ 11 ] . Additionally, SAR systems are rarely used for spatial thickness estimation. Most state-of-the-art drone-based solutions lack on-site real-time processing due to the constraints imposed by the limited power budget of battery-powered drones. To address this, field-programmable gate array (FPGA) compute platforms have been widely explored in other applications [ 12 , 13 ] for executing complex algorithms at the edge. However, their use in oil spill monitoring remains largely unexplored. Contrary to the state-of-the-art techniques, we suggested a drone platform with radar sensor approaches for both, oil spill detection [ 14 ] and thickness estimation [ 15 ] , in calm ocean conditions. Additionally, we proposed extremely small deep learning models (few bytes in size) that are suitable for direct onboard processing [ 16 ] . But these TinyML models work only at extremely low wind speeds. Then, we improved the monitoring by covering moderate wind speed scenarios using a more complex U-Net model [ 17 ] . In this paper, we propose a drone-based smart monitoring system using an optimized Tiny U-Net model for on-site estimation of up to 10 mm oil slick thickness at both calm and moderate wind speeds. This work differs from that of [ 17 ] by optimizing the architecture of the original U-Net to further boost the system and model performance. By applying different compression techniques, the optimized Tiny U-Net outperforms the state-of-the-art in estimation performance, while reducing its size by a factor of 269x. Furthermore, by implementing the Tiny U-Net on the FPGA with memory constraints, the applied optimization showcases the efficient edge computing capability of the proposed system compared to [ 17 ] in near real-time with low power consumption. Fig. 1: The drone-based proposed system for oil spill monitoring with on-edge computing. II Monitoring System II-A Environmental Model For the marine environment, we follow the same model adopted in [ 17 ] . The current investigation focuses on floating thick slicks that occur at the early stage (within hours to 1-2 days) of the spill before emulsification, with a thickness range from 1 to 10 mm. The ocean surface roughness is mathematically modeled by the surface root-mean-square height. It depends on the wind speeds above the sea surface, which is varied between 2 and 8 m/s. The relative dielectric constant of the oil is assumed to be 3. The water temperature and salinity are assumed 20 ∘ and 35 ppt, respectively. II-B Monitoring System Model The target is to estimate the spatial thickness distribution of the oil spill directly on-site. Our system is composed of three main components, as shown in Fig. 1 : • The drone platform to carry the sensor and perform the scanning of potential oil spill scenes. • A wide band radar sensor that utilizes multiple frequency channels in C and X -bands (4-12 GHz), which are convenient considering the resulting antenna size and weight relative to the drone. The selected number of frequencies is 9, from 4 GHz to 12 GHz with an increment of 1 GHz. • An FPGA hardware compute platform as an embedded system mounted on the drone platform to perform on-edge data processing. During the scan of a potential oil spill scene, the drone-based radar operates as a nadir-looking system where the propagating electromagnetic waves are normally incident on the ocean surface. Thus, the radar specular backscattering is completely captured [ 18 ] . Afterwards, the radar backscattering is processed onboard on the FPGA. The synthetic radar dataset is based on the complete framework presented in [ 17 ] , and publicly available in [ 19 ] . III Tiny U-Net Model Fig. 2: U-Net internal architecture. III-A Input and Output Maps Once the wideband radar scans the scene, the collected data is forwarded to the U-Net model in a 3-D N_{w}\times N_{l}\times N_{c} matrix. N_{w} and N_{l} represent respectively the geometrical width and length of the scanned environment, and N_{c} represents the number of used radar channels. The model outputs the spatial distribution map of the estimated thickness of potential oil slicks. Considering the discrete range of thicknesses is from 0 mm to 10 mm, each pixel in this output map is