Computer Vision · Deep Learning · 2026

MARIDA: Marine Debris Detection from Sentinel-2 Satellite Imagery

Built a PyTorch multi-label classification model to detect marine debris and ocean-surface features across 15 imbalanced environmental classes from satellite imagery.

PythonPyTorchDeep Learning

0.19

Macro F1

0.23

Micro F1

0.12

mAP

0.63

Best class (Marine Water) AP

Detection performance by metric

Aggregate scores (macro/micro F1, mAP) sit well below the best single class (Marine Water), which is the clearest picture of the class-imbalance problem.

Problem / Research Question

Detecting marine debris from multispectral satellite imagery is a multi-label classification problem with severe class imbalance: some environmental classes (like open marine water) are common, while debris-related classes are rare.

Dataset

Sentinel-2 multispectral satellite imagery labelled across 15 environmental classes, including marine debris and natural surface features.

Methodology

01Built a PyTorch multi-label classification model over multispectral satellite bands.
02Evaluated performance per-class to surface where class imbalance was limiting the model most.

Models

PyTorch multi-label CNN classifier

Key Findings

Marine Water was the strongest-performing class by a wide margin (AP 0.63), reflecting both its prevalence in the dataset and its distinct spectral signature.
Overall macro and micro F1 scores (0.19 and 0.23) show the model struggled broadly once averaged across all 15 classes.

Limitations

Class imbalance was the dominant limiting factor: rarer classes such as Dense Sargassum and Waves saw much weaker performance than common classes.
Low mAP (0.12) indicates the model is not yet reliable for rare-class detection in its current form; results should be read as a baseline, not a deployment-ready detector.