Research / Deep Learning

Marine Species Classification

Research-focused image classification work comparing CNN and Transformer models for endangered marine species.

Overview

This research initiative addresses the data scarcity and environmental distortions typical of underwater monitoring environments to identify endangered marine species.

The work benchmarks advanced deep visual backbones—including transformer and convolutional paradigms—against unified datasets to validate their viability for real-world automated oceanic exploration.

Tech Stack

Backbones

Swin-Tiny Transformer

ViT-Tiny

ConvNeXt Model Families

Frameworks

PyTorch Lightning

Hugging Face Hub

Visualizations

Grad-CAM Interrogation Maps

t-SNE Dimensionality Layouts

Features

+ Multi-repository underwater dataset unification and alignment

+ Comprehensive benchmark testing for visual backbones

+ Few-shot capability mapping across highly data-constrained classes

+ Interpretability analysis via visual attention spatial clustering

+ Reproducible Kaggle notebook execution architectures

Architecture

Unified Underwater Imagery Repositories

Domain-Specific Augmentation Layers

Deep Learning Training Matrix (Swin/ViT/ConvNeXt)

Interpretability Layers (Grad-CAM/t-SNE)

Macro F1 Cross-Validation Benchmarks

Challenges

Highly imbalanced dataset pools and water-turbidity noise distorted early attention configurations during training iterations.

Configured robust domain-specific transformations and evaluation parameters to force attention onto anatomical boundaries rather than environmental artifacts.

Balanced performance metrics carefully to ensure zero-class overlap artifacts did not inject classification bias.

Lessons Learned

Scientific validation and rigorous evaluation structures for machine learning systems
Structural trade-offs between global self-attention networks and local convolutional operations
Data handling and mitigation strategies for real-world environmental distortion challenges