FertilNet

Hierarchical Knowledge Distillation Framework for Lightweight Real-Time Soil Fertility Prediction

Architecture

Teacher MLP (5 × 256) ↓ Assistant MLP (3 × 128) ↓ Student MLP (2 × 64)

🚀 Project Highlights

• 🎯 90.91% Test Accuracy
• ⚡ 51× Smaller Model than Teacher Network
• 🧠 Hierarchical Knowledge Distillation (Teacher → Assistant → Student)
• 🌱 Real-Time Soil Fertility Prediction
• 📦 ONNX Export for Edge Deployment
• 🐍 Built with PyTorch, Scikit-Learn, and XGBoost

Overview

FertilNet is a machine learning framework that predicts soil fertility using physicochemical soil attributes while remaining lightweight enough for deployment on resource-constrained devices.

The project introduces a Teacher–Assistant–Student (TAS) knowledge distillation pipeline that compresses a large neural network into a compact model suitable for real-time inference.

Why FertilNet?

Traditional soil fertility testing is often expensive, time-consuming, and inaccessible to smallholder farmers. While machine learning models can improve prediction accuracy, many high-performing models are too computationally intensive for deployment on low-power devices.

FertilNet addresses this challenge through a Hierarchical Knowledge Distillation framework that compresses a large neural network into a lightweight student model. By leveraging a Teacher–Assistant–Student (TAS) architecture, FertilNet achieves high prediction accuracy while remaining suitable for real-time inference and edge deployment.

This makes the framework a promising solution for precision agriculture applications in resource-constrained environments.

Features

• Soil fertility prediction
• Hierarchical Knowledge Distillation
• Teacher–Assistant–Student architecture
• ONNX model export
• Edge AI deployment support
• Random Forest baseline comparison
• Model compression for real-time inference

Dataset Features

The model uses 12 soil attributes:

• Nitrogen (N)
• Phosphorus (P)
• Potassium (K)
• pH
• Electrical Conductivity (EC)
• Organic Carbon (OC)
• Sulfur (S)
• Zinc (Zn)
• Iron (Fe)
• Copper (Cu)
• Manganese (Mn)
• Boron (B)

Technologies

• Python
• PyTorch
• Scikit-Learn
• XGBoost
• ONNX
• NumPy
• Pandas
• Matplotlib

Results

Model	Accuracy
Teacher MLP	87.50%
Student (HierKD)	90.91%
Random Forest	90.23%

Key Achievements

• Achieved 90.91% test accuracy using Hierarchical Knowledge Distillation.
• Reduced model size by approximately 51× compared to the teacher network.
• Exported the lightweight student model to ONNX for edge deployment.
• Enabled real-time inference suitable for resource-constrained devices.

The distilled student model achieved higher accuracy than the teacher while using approximately 51× fewer parameters.

Project Structure

FertilNet/
│
├── data/
│   └── soil_data.csv
│
├── images/
│
├── outputs/
│
├── main.py
├── requirements.txt
├── README.md
└── .gitignore

Installation

git clone https://github.com/ButtySaylee/FertilNet.git

cd FertilNet

pip install -r requirements.txt

Run

python main.py

Author

Butty Saylee

BTech Software Engineering
Delhi Technological University

License

MIT License