Sediment Particle Size Prediction (sedpred-project)

UE23CS352A — Machine Learning Mini Project

Team Members:

• C S Deepak
• Dareddy Devesh reddy
  Project Duration: Sept 29 – Oct 13, 2025
  Faculty: D uma

---

Project Overview

This project replicates the research paper

Machine Learning for Predicting Sediment Particle Size Distributions” — Galen Egan, Stanford University

• We apply machine learning models to predict:
• Median sediment particle diameter (d₅₀)
• Particle size distribution variance (σ²)
  using water-quality and hydrodynamic features such as salinity, temperature, and velocity.

---

Objectives

• Implement Random Forest and SVR models to estimate d₅₀ and σ²
• Analyze feature importance
• Compare model performance using R² scores
• Visualize prediction accuracy and model stability (OOB curve)

---

Dataset

Data Since the original dataset link is no longer available, we generated a synthetic dataset (~1648 samples) that simulates real marine sediment data.

Feature	Description
S	Salinity (ppt)
ub	Near-bottom wave velocity (m/s)
np	Particle refractive index
T	Water temperature (°C)
a676/a650	Organic peak ratio
a450/a676	Inorganic peak ratio
chl_a	Chlorophyll-a concentration
u	Mean tidal velocity

Targets:

• d₅₀: Median particle diameter
• σ²: Particle size distribution variance

---

⚙️ Model Pipeline

1️⃣ Data Preprocessing

01_EDA.ipynb

• Generated synthetic dataset → data/synthetic_data.csv
• Scaled and cleaned data using StandardScaler

2️⃣ Modeling

02_models.ipynb

• RandomForestRegressor (n_estimators=32, oob_score=True)
• Support Vector Regressor (SVR) (C=2048, epsilon=4)
• R² score for each target variable
• Out-of-Bag (OOB) score curve
• Feature importance analysis

---

Results

Model	R² (d₅₀)	R² (σ²)
Random Forest	0.84047	0.914699
SVR	0.826608	0.928971

Conclusion:
Both models perform comparably, with Random Forest slightly outperforming SVR and providing clearer feature importance insights.

---

Visual Outputs

All plots are saved in the figures :

• feature_importance_d50.png
• oob_score_curve
• predicted vs actual

---

Repository Structure

sedpred-project/
├─ data.zip         # raw and processed datasets (not pushed to repo if large)
├─ notebooks/
│   ├─ 01_EDA.ipynb      # exploratory data analysis
│   ├─ 02_models.ipynb   # baseline models (RF, SVR)
├─ figures.zip           # generated plots & figures
├─ Final_result .zip  
├─requirements.txt      # dependencies
├─ .gitignore
├─ README.md             # project documentation
└─ one_page_writeup.pdf  # final report (to be added)



---

##  Requirements

Install dependencies:
```bash
pip install -r requirements.txt