[Paper Review] Oil Spill Segmentation using Deep Encoder-Decoder models

Abhishek Ramanathapura Satyanarayana, Maruf A. Dhali|arXiv (Cornell University)|May 2, 2023

Oil Spill Detection and MitigationEnvironmental Science3 citations

TL;DR

This paper proposes training deep encoder-decoder models directly on high-resolution Synthetic Aperture Radar (SAR) satellite images for oil spill segmentation, avoiding patch-based processing. The best-performing model—ResNet-50 encoder with DeepLabV3+ decoder—achieved a mean Intersection over Union (m-IoU) of 64.868% and a class IoU of 61.549% for oil spills, outperforming prior patch-based methods in detecting actual oil spills despite challenges with look-alikes.

ABSTRACT

Crude oil is an integral component of the world economy and transportation sectors. With the growing demand for crude oil due to its widespread applications, accidental oil spills are unfortunate yet unavoidable. Even though oil spills are difficult to clean up, the first and foremost challenge is to detect them. In this research, the authors test the feasibility of deep encoder-decoder models that can be trained effectively to detect oil spills remotely. The work examines and compares the results from several segmentation models on high dimensional satellite Synthetic Aperture Radar (SAR) image data to pave the way for further in-depth research. Multiple combinations of models are used to run the experiments. The best-performing model is the one with the ResNet-50 encoder and DeepLabV3+ decoder. It achieves a mean Intersection over Union (IoU) of 64.868% and an improved class IoU of 61.549% for the ``oil spill" class when compared with the previous benchmark model, which achieved a mean IoU of 65.05% and a class IoU of 53.38% for the ``oil spill" class.

Motivation & Objective

• To evaluate the feasibility of training deep encoder-decoder models directly on full-resolution SAR satellite images for oil spill segmentation.
• To compare multiple encoder-decoder combinations in detecting oil spills from high-dimensional SAR data.
• To improve oil spill detection performance by avoiding image patching, which increases memory and computational costs.
• To assess model robustness in distinguishing oil spills from look-alike features such as oil spill look-alikes and ships.
• To provide a baseline for future work using self-attention modules or alternative architectures.

Proposed method

• The study employs a U-Net-style encoder-decoder architecture with multiple backbone encoders (e.g., ResNet-50, EfficientNet-B0) and decoders (e.g., DeepLabV3+, PSPNet).
• High-resolution SAR images are used as input without patching, preserving spatial context and reducing computational overhead from patch management.
• The models are trained using supervised semantic segmentation with cross-entropy and Dice loss, optimizing for pixel-wise classification.
• A multi-class IoU metric is used to evaluate performance across five classes: sea surface, oil spill, oil spill look-alike, ship, and land.
• The best-performing model is selected based on validation m-IoU, with inference conducted on a test set of 102 images.
• Training is conducted on the Peregrine HPC cluster using PyTorch, with data augmentation and learning rate scheduling.

Experimental results

Research questions

• RQ1 Does training on full-resolution SAR images improve oil spill segmentation performance compared to patch-based approaches?
• RQ2 Which encoder-decoder combination yields the highest m-IoU and class IoU for oil spill detection?
• RQ3 How well do the models generalize to distinguishing oil spills from oil spill look-alikes and ships?
• RQ4 What is the impact of model architecture choice on detection accuracy for small and large oil spill regions?
• RQ5 Can high-dimensional input processing reduce memory and energy costs while maintaining or improving performance?

Key findings

• The ResNet-50 encoder with DeepLabV3+ decoder achieved the highest mean IoU of 64.868% on the test set, outperforming the benchmark model's 65.06% m-IoU.
• For the oil spill class specifically, the best model achieved a class IoU of 61.549%, significantly higher than the benchmark model's 53.38%.
• The model demonstrated strong performance on 'sea surface' (96.422% IoU) and 'land' (92.218% IoU), indicating robust background segmentation.
• Detection of 'ships' was improved, with a class IoU of 33.378% compared to 27.63% in the benchmark model.
• The model struggled with 'oil spill look-alike' regions, achieving 40.773% IoU—lower than the benchmark's 55.40%, indicating ongoing confusion between similar features.
• Qualitative analysis showed consistent detection of medium to large oil spills, but inconsistent performance on small or ambiguous regions, particularly near ships or look-alikes.