Rice Grain Classification Using a Feedforward Neural Network

A Binary Classification Study on Morphological Features of Jasmine and Gonen Rice Varieties

---

Overview

This project implements a feedforward neural network (FNN) from scratch using PyTorch to classify rice grains into two varieties — Jasmine and Gonen — based on morphological features extracted from grain images. The study demonstrates that lightweight neural network architectures can achieve high classification accuracy on well-structured agricultural datasets without the overhead of deep convolutional models.

Scope: This is a proof-of-concept implementation and learning exercise exploring the applicability of simple neural networks to agricultural grain classification tasks.

---

Dataset

• Source: riceClassification.csv — a tabular dataset of morphological measurements derived from rice grain images
• Classes: Binary — Jasmine (1) / Gonen (0)
• Features (10 morphological attributes):

Feature	Description
Area	Total pixel area of the grain
MajorAxisLength	Length of the major axis of the ellipse fitting the grain
MinorAxisLength	Length of the minor axis
Eccentricity	Eccentricity of the fitted ellipse
ConvexArea	Area of the convex hull of the grain
EquivDiameter	Diameter of a circle with the same area
Extent	Ratio of grain pixels to bounding box pixels
Perimeter	Grain boundary length
Roundness	Measure of circularity
AspectRation	Ratio of major to minor axis length

• Preprocessing: Max-absolute normalization applied to all features; id column dropped; missing values removed

---

Model Architecture

A simple feedforward neural network with the following structure:

Input Layer  →  (10 features)
Hidden Layer →  Linear(10 → 20) + [implicit linear activation]
Output Layer →  Linear(20 → 1) + Sigmoid

• Loss Function: Binary Cross-Entropy Loss (BCELoss)
• Optimizer: Adam (lr = 0.01)
• Batch Size: 32
• Epochs: 10

Note: No nonlinear activation (e.g., ReLU) is applied between the input and hidden layer in the current implementation, making the two linear layers effectively equivalent to a single linear transformation. This is a known architectural limitation and a potential area for improvement.

---

Data Splits

Split	Proportion	Purpose
Training	70%	Model training
Validation	15%	Hyperparameter monitoring
Test	15%	Final evaluation

---

Results

Metric	Value
Final Training Accuracy	~98.55%
Final Validation Accuracy	~98.61%
Test Accuracy	98.42%

Training and validation loss converged steadily within 10 epochs, with no significant signs of overfitting.

---

Limitations & Future Work

This project is intentionally minimal and has the following known limitations:

• Binary classification only — does not generalize to multi-class rice variety identification
• Shallow architecture — a single hidden layer with no nonlinear activation between layers limits representational capacity
• Tabular features only — morphological CSV features are used rather than raw image data; a CNN-based approach on images would be more generalizable
• Small epoch count — 10 epochs may not represent fully converged training
• No regularization — dropout or L2 regularization not applied

Potential improvements:

• Add ReLU activation after the hidden layer
• Extend to multi-class classification (e.g., Arborio, Basmati, Ipsala, Karacadag)
• Implement a CNN pipeline operating directly on grain images
• Apply cross-validation for more robust evaluation
• Compare performance against traditional ML baselines (SVM, Random Forest)

---

Requirements

torch
numpy
pandas
scikit-learn
matplotlib
torchsummary

Install with:

pip install torch numpy pandas scikit-learn matplotlib torchsummary

---

Usage

1. Clone the repository
2. Place riceClassification.csv in the project root
3. Open and run rice_classification.ipynb sequentially

For inference on a new sample, update the morphological values in the prediction cell at the bottom of the notebook.

---

Project Structure

├── riceclassification.ipynb   # Main notebook
├── riceClassification.csv      # Dataset
└── README.md

---

Academic Context

This project was developed as part of an independent study in applied machine learning, focusing on agricultural classification tasks. The implementation prioritizes interpretability and simplicity over performance optimization, making it a useful reference for understanding neural network fundamentals in a real-world domain.

---

Author

[Naosin Tabassum Lineya] [Jahangirnagar University] [lineyanaosin@gmail.com]