Crop Yield Prediction using ANN

This project implements an end-to-end regression model to predict crop yields (measured in hectograms per hectare) using various factors like location, crop type, year, rainfall, pesticide usage, and average temperature. It includes a data preprocessing pipeline, model training scripts, and a Streamlit-based web interface for real-time predictions.

Model Specifications

The predictive model is built using a Sequential Artificial Neural Network (ANN) implemented with TensorFlow/Keras. Below are the detailed specifications:

• Model Architecture:

• Input Layer: Accepts 15 input features (after preprocessing and encoding).

• Hidden Layer 1: 64 neurons with ReLU (Rectified Linear Unit) activation function.

• Hidden Layer 2: 32 neurons with ReLU activation function.

• Output Layer: 1 neuron (for regression) to predict the numerical yield value.

• Compilation Details:

• Optimizer: Adam.

• Loss Function: Mean Absolute Error (MAE).

• Metrics: Mean Absolute Error (MAE).

• Training Configuration:

• Epochs: 100.

• Callbacks: Includes Early Stopping (patience of 15 epochs) to prevent overfitting and TensorBoard for visualization.

• Data Split: 80% Training, 20% Testing.

• Performance: The model achieved a validation MAE of approximately 17,074 hg/ha after 100 epochs.

Dataset and Features

The model is trained on a dataset (yield.csv) containing the following features:

Feature Name	Type	Description
Area	Categorical	Country or region name (e.g., India, Brazil).
Item	Categorical	Crop type (e.g., Maize, Rice, Wheat).
Year	Temporal	Year of observation.
Average Rainfall	Numerical	Annual rainfall in mm per year.
Pesticides	Numerical	Amount of pesticides used in tonnes.
Average Temp	Numerical	Average yearly temperature in °C.
hg/ha_yield	Target	Crop yield in hectograms per hectare.

Preprocessing Pipeline

To prepare the data for the ANN, the following steps are performed:

1. Encoding:

• Item: One-Hot Encoded (transformed into 10 binary columns).
• Area: Target Encoded (using category_encoders.TargetEncoder to handle high cardinality).

2. Scaling: All numerical features are standardized using StandardScaler to ensure the neural network converges efficiently.
3. Artifacts: The trained encoders and scalers are saved as .pkl files in the artifacts/ folder for use during inference.

Project Structure

• app.py: The Streamlit web application providing the user interface.
• cleaning.ipynb: Jupyter notebook detailing data cleaning, encoding, scaling, and model training.
• predict.ipynb: Notebook containing the logic for loading the model and making test predictions.
• model.h5: The saved trained ANN model.
• artifacts/: Contains Age_encoder.pkl (Area encoder), onehot_encoder.pkl, and scaler.pkl.
• requirements.txt: Lists all necessary Python libraries (TensorFlow, Streamlit, Pandas, etc.).

Installation and Usage

1. Install Dependencies:

pip install -r requirements.txt

2. Run the Web Application:

streamlit run app.py

3. Prediction: Enter the required crop and environmental details in the Streamlit UI to receive an instant yield prediction.