1.9 KiB
Research: ResNet Phenology Classifier
Date: 2025-11-04 Feature: specs/1-phenology-classifier/spec.md
Research Tasks
- Research best ResNet variant for plant image classification
- Research data preprocessing techniques for botanical images
- Research evaluation metrics for multi-class phenological phase classification
- Research reproducibility practices for ML experiments
Findings & Decisions
ResNet Variant
Decision: Use ResNet50 as the base architecture Rationale: Provides good balance between accuracy and computational efficiency. ResNet50 has been proven effective for image classification tasks similar to ImageNet. Alternatives considered: ResNet18 (faster but lower accuracy), ResNet101 (higher accuracy but more compute-intensive)
Data Preprocessing
Decision: Use standard ImageNet preprocessing with augmentation Rationale: Random cropping, horizontal flipping, normalization to ImageNet mean/std. This is standard for transfer learning with ResNet. Alternatives considered: Custom augmentations for plant-specific features, but standard works well for general classification.
Evaluation Metrics
Decision: Primary: Accuracy, Secondary: F1-score per class, Precision, Recall Rationale: Accuracy for overall performance, F1-score to handle class imbalance in phenological phases. Alternatives considered: AUC-ROC (more for binary), but multi-class metrics are appropriate.
Reproducibility
Decision: Use random seeds, version data with DVC, log all hyperparameters Rationale: Ensures experiments can be reproduced. DVC for data versioning, MLflow or similar for experiment tracking. Alternatives considered: Manual logging, but automated tools are more reliable.
Resolved Clarifications
- Dataset format: Images in directory, labels in CSV with columns: image_path, phase
- Model output: Probabilities for each phase class
- Training hardware: GPU required for reasonable training time