Fertilizer Recommendation System

A machine learning approach to predict optimal fertilizer recommendations based on soil conditions and crop requirements using Random Forest classification.

Project Overview

This project tackles fertilizer recommendation as a multi-class classification problem, predicting the top 3 most suitable fertilizer types for given agricultural conditions. The model analyzes soil nutrient levels, environmental factors, and crop characteristics to generate recommendations.

Dataset

The project combines multiple data sources:

• Training Data: 850,000 samples (including external fertilizer dataset)
• Test Data: 250,000 samples
• Features: 27 engineered features derived from 7 base features

Input Features

• Soil Nutrients: Nitrogen (N), Phosphorus (P), Potassium (K)
• Environmental: Temperature, Moisture Content
• Categorical: Soil Type, Crop Type

Feature Engineering

The model implements several feature engineering techniques to capture agricultural relationships:

Nutrient Ratios

• N/P, N/K, P/K ratios (nutrient balance is often more critical than absolute values)
• Total nutrient content

Environmental Interactions

• Temperature² and Moisture² (polynomial features for non-linear relationships)
• Temperature × Moisture interaction terms

Categorical Encoding

• One-hot encoding for soil types and crop varieties
• Ensures consistent feature space across train/test sets

Model Architecture

Algorithm: Random Forest Classifier

• 200 estimators
• Max depth: 10 (prevents overfitting)
• Min samples per leaf: 5
• Stratified train/validation split (80/20)

Results

• Validation F1 Score: 0.1437 (weighted average)
• Output Format: Top 3 fertilizer recommendations per sample
• Model Performance: Indicates significant room for improvement

Key Challenges

1. Multi-class complexity: Predicting specific fertilizer products rather than nutrient requirements
2. Feature scaling: Mixed units across temperature, percentages, and ratios
3. Domain knowledge integration: Agricultural expertise needed for proper feature selection
4. Evaluation metrics: Gap between F1 scoring and top-3 prediction format

Technical Implementation

The pipeline includes:

• Data standardization and cleaning
• Feature engineering with agricultural domain knowledge
• Model validation with stratified sampling
• Top-3 prediction generation using probability rankings

Future Improvements

• Implement proper feature scaling and normalization
• Integrate soil pH and micronutrient data
• Add cross-validation for more robust evaluation
• Consider ensemble methods beyond Random Forest
• Incorporate seasonal and regional agricultural factors

Usage

# Load and preprocess data
train_df = create_features(train_df)
test_df = create_features(test_df)

# Train model
rf_model = RandomForestClassifier(n_estimators=200, max_depth=10, min_samples_leaf=5)
rf_model.fit(X_final, y)

# Generate predictions
predictions = rf_model.predict_proba(X_test_final)

Dependencies

• pandas
• numpy
• scikit-learn
• Standard Python data science stack

---

Note: This is a competition-style implementation focused on prediction accuracy. Real-world agricultural applications would require additional domain expertise and validation.