Forest Cover Classification

Project re-upload notice (Jan 2026)

I got locked out of the original GitHub account that hosted this project. This repository is intended to be re-uploaded and maintained under my new GitHub account.

New canonical repo : https://github.com/agilap/Forest-Cover-Classification

Dataset Overview and Processing Pipeline

Dataset Details

• Records: 581,012
• Features: 54 (excluding the target variable class)
• Memory Usage: 243.80 MB
• Target Variable: class (7 classes: 1–7)

Feature Summary

• Includes topographic (e.g., Elevation, Slope), spatial (e.g., distances to hydrology/roadways/fire points), light-related (Hillshade), and categorical indicators (Wilderness Areas, Soil Types as one-hot encoded).
• All features are integer type.
• No missing values.

Target Class Distribution

Class 1: 211,840
Class 2: 283,301
Class 3: 35,754
Class 4: 2,747
Class 5: 9,493
Class 6: 17,367
Class 7: 20,510

Data Preparation Workflow

• Loading: Dataset loaded without issues.

• Label Encoding: class column encoded to 0–6.

• Feature Scaling: Applied to ensure normalized input values.

• Split:

  • Train: 406,707 samples
  • Validation: 58,102 samples
  • Test: 116,203 samples

Model Tuning Summary

• Best Validation Accuracy: 87.41%

• Optimal Hyperparameters:

  • Architecture: simple
  • Dropout Rate: 0.2
  • Learning Rate: 0.001
  • Batch Size: 64

Top Performing Configurations

1. Acc: 0.8741 | Dropout: 0.2 | LR: 0.001 | Batch: 64
2. Acc: 0.8725 | Dropout: 0.2 | LR: 0.001 | Batch: 128
3. Acc: 0.8700 | Dropout: 0.2 | LR: 0.001 | Batch: 32
4. Acc: 0.8635 | Dropout: 0.2 | LR: 0.005 | Batch: 128
5. Acc: 0.8529 | Dropout: 0.3 | LR: 0.001 | Batch: 64

Conclusion

The dataset was efficiently preprocessed and trained using a simple neural network architecture. The tuning process yielded strong performance, indicating that even basic models can achieve high accuracy with well-prepared features.