Oil Spill Semantic Segmentation

The Problem

Oil-spill monitoring from aerial imagery requires detailed pixel-level interpretation rather than simple image-level classification. The challenge is increased by severe class imbalance, visual similarity between surface phenomena, and large appearance variations within the same semantic class.

The Approach

• Six-class semantic segmentation covering Background, Emulsion, Oil, Oil Platform, Sheen, and Ship.
• Image preprocessing and augmentation using geometric and photometric transformations.
• Class-distribution analysis and weighting strategies to reduce the effect of class imbalance.
• A pretrained ResNet50 encoder combined with a custom decoder, skip connections, progressive bilinear upsampling, and channel/spatial attention.
• Training with weighted Cross-Entropy and Dice Loss, AdamW optimization, learning-rate scheduling, and early stopping.

Technical Decisions

• Class imbalance was addressed explicitly through distribution analysis and weighted learning rather than relying on architecture changes alone.
• Skip connections were used to recover spatial detail that can be lost during encoder downsampling.
• Channel and spatial attention were introduced to help the model emphasize relevant visual features and regions within aerial imagery.
• Qualitative mask inspection was treated as an important part of evaluation because segmentation quality depends on spatial structure, not only a single aggregate score.

Results

The project produced a complete multi-class semantic- segmentation pipeline covering dataset analysis, preprocessing, model development, training, and qualitative review of predicted segmentation masks.

Why It Matters

The project demonstrates how deep learning can be applied to complex environmental imagery where multiple surface phenomena may appear visually similar. Pixel-level segmentation can provide more detailed spatial information than image-level classification for environmental monitoring workflows.