🌊 AquaMorph — Microplastic Morphology Classifier

Theme: Marine Ecosystem Protection & Water Quality Monitoring

---

📽️ Demo Video

▶️ [Watch Demo Video]https://drive.google.com/file/d/1NN_VU7zDTB2pvGUqbP6io8ZcJAt-oihn/view?usp=sharing

---

🖥️ App Screenshots

Batch Analysis with ETI Scoring

Batch

---

📁 Datasets Used & Preprocessing

Datasets

Dataset	Source	Train Images	Val Images	Classes
microplastic_100	Roboflow Universe	210	30	fiber, film, fragment, pallet
Microplastic-v2	Roboflow Universe	1,461	140	bead, fiber, fragment
Merged	Combined	1,671	170	fiber, fragment, film, bead

📂 Full datasets + trained model: Google Drive

Class Standardization

Dataset 1 original	Dataset 2 original	AquaMorph unified class
fiber	fiber	fiber
fragment	fragment	fragment
film	(not present)	film
pallet	bead	bead

Preprocessing Pipeline

1. Merging — Images from both datasets merged into unified folder with remapped class labels
2. Resize — All images standardized to 640×640 for YOLOv8 input
3. Augmentation (applied by YOLOv8 during training):
   • Random horizontal flip (p=0.5)
   • HSV colour jitter (hue ±1.5%, saturation ±70%, value ±40%)
   • Random erasing (p=0.4)
   • Mosaic augmentation (4 images per tile)
4. Normalization — Pixel values normalized to [0, 1]
5. Class imbalance — Handled via augmentation; film class supplemented from Dataset 1

Size Estimation Preprocessing (OpenCV)

1. Grayscale conversion
2. Gaussian blur (5×5 kernel) for noise reduction
3. Otsu's adaptive thresholding to isolate particle from background
4. cv2.findContours() to detect particle boundary
5. Maximum pairwise distance across contour points = Feret diameter
6. Pixel → micrometre conversion at 0.25 µm/pixel (40× magnification)

---

🤖 Model Used & Performance Metrics

Architecture

![Screenshot 2]!

Property	Value
Model	YOLOv8n (nano)
Pretrained on	COCO dataset
Task	Object Detection + Classification
Parameters	3,006,428
Model size	6.2 MB
Input size	640×640
Training epochs	30
Optimizer	Auto (SGD → Adam)
Batch size	16
Hardware	Tesla T4 GPU (Google Colab)
Training time	~16 minutes

Performance Metrics (Final Validation)

Class	Precision	Recall	mAP50	mAP50-95
fiber	0.923	0.961	0.956	0.775
fragment	0.937	0.970	0.959	0.595
film	0.972	0.977	0.958	0.430
bead	0.801	0.825	0.816	0.515
Overall	0.908	0.933	0.922	0.578

• Overall mAP50: 92.2% — production-quality performance
• Training loss at epoch 30: box=0.912, cls=0.514, dfl=1.001
• Inference speed: 4.1ms per image on GPU

Training Output

Ecological Threat Index (

ETI) Formula

ETI = morphology_score + size_score

morphology_score:
  fiber     = 40  (highest — ingested by filter feeders)
  fragment  = 30  (leaches chemical additives)
  film      = 25  (smothers benthic organisms)
  bead      = 20  (point source, predictable)

size_score = 60 × (1 − min(feret_µm, 5000) / 5000)
  → particles < 100µm can penetrate biological tissues
  → particles < 10µm can cross cell membranes

ETI range: 0–100  (higher score = greater ecological threat)

---

✨ Key Features

Core Classification Features

Feature	Description
🔍 Morphology classification	Classifies into Fiber, Fragment, Film, Bead with confidence %
📏 Feret diameter estimation	Estimates longest particle dimension in micrometres (µm) via OpenCV contour detection
🌡️ Ecological Threat Index	Combined 0–100 risk score — ETI Score shown per particle
🚦 Threat level badge	HIGH RISK / MEDIUM / LOW colour-coded alert per prediction
🧪 Polymer source predictor	Maps morphology → likely polymer (e.g. Polyester/Nylon for fiber)
🌿 Degradation stage estimator	Classifies particle as Fresh / Lightly Weathered / Heavily Weathered

Analysis Modules (from sidebar)

Module	Description
🔍 Classify	Single image upload → full morphology + ETI analysis
📦 Batch Analysis	Upload multiple images → summary table with Class, Confidence, Feret Diameter, ETI Score, Threat Level, Degradation, Polymer → Download Full Results CSV
🗺️ Spatial Heatmap	Multi-particle field image → colour-coded particle distribution map
🎨 GAN Generator	Generate synthetic microplastic training images per class
🧪 Polymer Source	Detailed polymer origin analysis per morphology type
📍 Field Logger	Log GPS coordinates + sample source + date for each analysis
⚠️ Review Queue	Low-confidence predictions flagged for manual human review
📊 Dashboard	Summary stats — Images Processed, Avg ETI Score, High Threat count
ℹ️ About	Indian river context, model info, ETI formula explanation

System Status Panel

Status	Component
🟢 Ready	YOLOv8 model loaded
🔵 Ready	Contour Analyser (OpenCV)
🔴 Standby	DCGAN Generator

🇮🇳 Indian Environmental Context

AquaMorph is calibrated against published Indian river research:

• Pune — Mula River → 1,808 microplastic particles/L detected (2025 study); film-shaped MPs most abundant; particles as small as 25µm (our backyard)
• Ganga & Yamuna Rivers → Fibre-shaped MPs dominant at 300µm–5mm; polyamide and PAM primary polymers
• South Indian rivers → Fibers 64.1%, films 21.7%, fragments 12%, beads 2.2%

---

🏗️ Model Architecture

Input Image (640×640)
        ↓
YOLOv8n Backbone (CSPDarknet)
        ↓
Feature Pyramid Network (FPN)
        ↓
Detection Head
        ↓
┌─────────────────────────────────────────┐
│  Class: fiber / fragment / film / bead  │
│  Confidence %                           │
│  Bounding box coordinates               │
└─────────────────────────────────────────┘
        ↓
OpenCV Contour Detection → Feret Diameter (µm)
        ↓
Degradation Stage Estimator
        ↓
ETI Score (0–100) + Threat Level + Polymer Source
        ↓
Streamlit UI — Classify / Batch / Heatmap / GAN / Dashboard

---

🚀 How to Run AquaMorph

Step 1 — Clone the repo

git clone https://github.com/Saanidhi123/cummins_hackathon.git
cd cummins_hackathon
pip install -r requirements.txt

Step 2 — Download the model

Download best.pt from Google Drive and place it inside the model/ folder:

📁 AquaMorph Model + Datasets (Google Drive)

Step 3 — Run

streamlit run app.py

Open http://localhost:8501 in your browser.

---

📂 Repository Structure

cummins_hackathon/
├── app.py                  ← AquaMorph Streamlit web app
├── train.py                ← YOLOv8 training script
├── requirements.txt        ← Python dependencies
├── README.md
├── assets/                 ← Screenshots and training output images
│   ├── training_results.png
│   ├── screenshot_batch.png
│   └── screenshot_classify.png
├── utils/
│   ├── contour.py          ← OpenCV Feret diameter estimation
│   └── scoring.py          ← ETI score + polymer + degradation predictor
└── model/
    └── .gitkeep            ← Download best.pt from Drive link above

---

🛠️ Tech Stack

Layer	Technology
ML Model	YOLOv8n (Ultralytics)
Image Processing	OpenCV
GAN Generator	DCGAN (PyTorch)
Frontend / UI	Streamlit
Data Management	Pandas
Training Platform	Google Colab (Tesla T4 GPU)
Dataset Source	Roboflow Universe
Model Storage	Google Drive
Version Control	GitHub

---

👩‍💻 Team — NeuralNomads_PS3

AquaMorph

Microplastic Morphology Classifier

---

📄 License

Datasets used under CC BY 4.0 license (Roboflow Universe).