Full-Scale Aggregated MobileUNet: An Improved U-Net Architecture for SAR Oil Spill Detection

Yi-Ting Chen ¹, Lena Chang ^2,3, Jung-Hua Wang ^1,4

⁰Department of Electrical Engineering, National Taiwan Ocean University, Keelung 202301, Taiwan.

Sensors (basel, Switzerland) +

|

June 27, 2024

View abstract on PubMed

Summary

A new deep learning model, FA-MobileUNet, significantly improves oil spill detection in synthetic aperture radar (SAR) images. This advanced model accurately distinguishes oil spills from look-alikes, enhancing marine environmental monitoring.

Area Of Science

Remote Sensing
Marine Environmental Monitoring
Artificial Intelligence

Background

Oil spills pose a significant threat to marine ecosystems.
Synthetic Aperture Radar (SAR) images are crucial for detecting oil spills, but distinguishing them from look-alikes is challenging.
Class imbalance in SAR image datasets complicates the training of accurate multi-category classifiers for marine targets.

Purpose Of The Study

To develop a lightweight deep learning model for improved oil spill detection in SAR images.
To enhance the feature extraction capabilities for diverse marine targets, including oil spills and their look-alikes.
To address the challenges of class imbalance and accurate classification in multi-target SAR image analysis.

Main Methods

Proposed a novel lightweight U-Net-based model named Full-Scale Aggregated MobileUNet (FA-MobileUNet).
Utilized MobileNetv3 as the U-Net encoder backbone for efficient feature extraction.
Incorporated Atrous Spatial Pyramid Pooling (ASPP) and Convolutional Block Attention Module (CBAM) for multi-scale feature extraction and computational efficiency.
Aggregated full-scale features from the encoder to improve the network's feature extraction competence.

Main Results

The FA-MobileUNet achieved a mean intersection over union (mIoU) exceeding 80% for detecting five marine target types.
Achieved IoU scores of 75.85% for oil spills and 72.67% for look-alikes, surpassing the original U-Net model by 18.94% and 25.55%, respectively.
Demonstrated superior accuracy in classifying dark regions in SAR images as oil spills or look-alikes compared to other semantic segmentation models.
Validated superior detection performance and computational efficiency against existing models.

Conclusions

The FA-MobileUNet model effectively enhances the extraction and integration of multi-scale features for accurate marine target detection.
The proposed model offers a significant advancement in distinguishing oil spills from look-alikes in SAR imagery.
FA-MobileUNet provides a computationally efficient and accurate solution for marine oil spill monitoring using SAR data.

Keywords:

U-Net oil spills semantic segmentation models synthetic aperture radar (SAR)