🌾 Oat Crop Classification

This project applies various classification models in R to predict whether oat crops are being grown on farmland using optical and radar-based remote sensing data. The dataset is derived from a modified version of the Winnipeg Crop Mapping Dataset, hosted by the UCI Machine Learning Archive.

Dataset

• WinnData.csv
• Source: UCI Machine Learning Archive
• Personal Subset (randomly sampled with seed = 33270961):
  • Observations: 5,000
  • Features: 20 variables from A01 to A30
  • Target: Binary variable Class
  • Class Distribution:
    • Oats: 14%
    • Other: 86%

Technologies Used

• Language: R
• IDE: RStudio
• Packages: tree, e1071, ROCR, randomForest, adabag, rpart, smotefamily, neuralnet, xgboost

Project Workflow

1. Exploratory Data Analysis

• Summary statistics, standard deviations, and correlation analysis
• Detection of outliers and assessment of multicollinearity

2. Preprocessing

• Log transformation of A26 to handle outliers
• Removal of highly correlated variables (A16 and A19)
• Synthetic Minority Oversampling Technique (SMOTE) for class balance
• Feature scaling to ensure stable convergence (when applicable)

3. Model Training

• Decision Tree
• Naive Bayes
• Bagged Trees
• AdaBoost
• Random Forest (RF)
• Hyperparameter-tuned RF on SMOTE-balanced data
• Artificial Neural Network (ANN) on SMOTE-balanced data
• XGBoost

4. Evaluation

• Accuracy, Precision, Recall, F1-Score, and AUC

Model Performance Summary

Model	Accuracy	Precision	Recall	F1-Score	AUC
Decision Tree	87%	53%	4%	7%	0.69
Naive Bayes	80%	25%	23%	24%	0.70
Bagged Trees	87%	67%	3%	6%	0.77
AdaBoost	87%	51%	24%	32%	0.73
RF	87%	71%	7%	14%	0.78
Tuned RF with SMOTE	80%	42%	53%	47%	0.72
ANN with SMOTE	67%	25%	72%	37%	0.73
XGBoost	77%	32%	61%	42%	0.78

Results

• Class imbalance had a major negative impact on model performance, particularly on recall for the minority class.
• Accuracy was not a reliable evaluation metric, as it was inflated by the majority class.
• The attributes A26, A02, A06, A07, A08, and A14 were consistently identified as the most significant across the "basic" models (Decision Tree, Naive Bayes, Bagged Trees, AdaBoost, Random Forest). These features were used in subsequent models to reduce overfitting and improve generalization.
• Tuned RF with SMOTE, ANN with SMOTE, and XGBoost achieved much higher recall and F1-scores than other models, with a slight trade-off in accuracy and precision due to more aggressive prediction of the minority class.
• Tuned RF with SMOTE provided the best overall balance with the highest accuracy and F1-score among the three. However, ANN with SMOTE was better for maximizing the identification of oat crops, achieving the highest recall while providing acceptable class discrimination with a fair AUC.

How to Run

1. Clone the repository or download the ZIP file from GitHub.
2. Open the project folder in RStudio.
3. Run the R script (oat_crop_classification.r) inside the RStudio environment.

Author

Developed by Juan Nathan.