Using TensorFlow and deep learning, this project classifies forest cover types using terrain and environmental features. We use Keras Tuner for hyperparameter optimization and ensemble learning for performance enhancement.
- This project aims to predict forest cover types (7 classes) based on cartographic features.
- Using the UCI Covertype dataset, we trained a deep learning model (CNN + Bi-LSTM) and improved performance through hyperparameter tuning and ensemble methods.
- Keras Tuner was used to find the best combination of optimizer, learning rate, and dropout rate.
- Final models were evaluated using classification metrics, confusion matrices, training curves, and error analysis.
The repository is organized as follows:
deep-learning-cover-classification-project/
├── cover_data.csv # Dataset file
├── CoverType-classification-origin.ipynb # Baseline notebook (v1)
├── CoverType-classification-ImprovedPerformance.ipynb # Tuned & Ensemble notebook (v2)
├── cover_classifier_model.h5 # Final saved Keras model
├── my_tuner_dir/ # Keras Tuner trial logs
└── README.md # Project documentation
- Source: UCI Machine Learning Repository – Forest CoverType Dataset
- Target: Multiclass classification (7 forest cover types)
Features (subset):
- Elevation
- Aspect
- Slope
- Horizontal/Vertical distance to hydrology
- Hillshade (morning, noon, 3pm)
- Distance to roadways
- Soil types (one-hot encoded)
- Embedding Layer: Input projection
- 1D Convolution + Batch Normalization
- Bidirectional LSTM Layers: LSTM(128) + LSTM(64)
- Fully Connected: Dense(64) → Dropout → Dense(7, softmax)
- Used Keras Tuner (RandomSearch)
- Search Space:
- Optimizer:
adam,sgd - Learning Rate:
1e-4 ~ 1e-2 - Dropout Rate:
0.2,0.4
- Optimizer:
- Best Parameters:
optimizer: sgd learning_rate: 0.001615 dropout_rate: 0.2 - EarlyStopping applied (patience=3)
- GPU dynamic memory allocation activated with
tf.config.experimental.set_memory_growth - Best model saved to
cover_classifier_model.h5
- Optimizer:
adam, Learning rate: 0.0003, Dropout: 0.5 - Validation Accuracy: 68.5%
- Test Accuracy: 68.29%
- Macro F1-score: 0.47
- Optimizer:
sgd, Learning rate: 0.001615, Dropout: 0.2 - Validation Accuracy: 71.99%
- Test Accuracy: 72.00%
- Macro F1-score: 0.46
- Weighted F1-score: 0.71
Highlights:
- Class 2 (Spruce/Fir): 88% recall
- Class 1 (Lodgepole Pine): 80% recall
- Class 4/5: Low recall, needs future work
To analyze model weaknesses, we extracted and printed random misclassified examples.
# 예측값이 정답과 다른 인덱스를 뽑아 무작위로 추출
incorrect_indices = np.where(y_pred_best != y_test)[0]
sample_indices = np.random.choice(incorrect_indices, size=5, replace=False)
print("\n잘못 분류된 샘플 예시:")
for idx in sample_indices:
print(f"Index: {idx}, True Label: {y_test.iloc[idx]}, Predicted: {y_pred_best[idx]}")📌 출력 예시:
Index: 40500, True Label: 6, Predicted: 0
Index: 14617, True Label: 1, Predicted: 0
Index: 58282, True Label: 0, Predicted: 1
Index: 11860, True Label: 0, Predicted: 1
Index: 73515, True Label: 1, Predicted: 0
This helps identify which classes are most frequently confused and motivates further class-specific tuning.
- Per-Class Accuracy (Bar Chart):
- Class 2 shows ~90% accuracy
- Classes 4, 5, 6 show significantly lower accuracy
- Training Curve (Train vs Validation Accuracy over Epochs):
- Tuned model shows stable validation accuracy around 72%
- Training accuracy slightly lower than validation (suggests potential underfitting or regularization)
| Category | Baseline Model (v1) | Tuned + Ensemble Model (v2) |
|---|---|---|
| Optimizer | adam | sgd |
| Learning Rate | 0.0003 | 0.001615 |
| Dropout | 0.5 | 0.2 |
| Validation Accuracy | 68.5% | 71.99% |
| Test Accuracy | 68.29% | 72.00% |
| Macro F1-score | 0.47 | 0.46 |
| Ensemble Used | No | Yes (Soft Voting) |
- Address class imbalance (oversampling or loss weighting)
- Introduce TabTransformer or attention-based models
- Apply confidence thresholding in ensemble inference
- Fine-tune batch size and learning schedule
This project is part of a deep learning portfolio and focuses on model design, tuning, and evaluation.