Wildlife Hotspot Detector: Active Learning on Skewed Datasets

Engineered an active learning machine learning pipeline to detect wildlife presence in large-scale, highly skewed ecological camera trap datasets with heavy background clutter.

Methodology & Technical Highlights

• Multi-Domain Feature Engineering: Extracted complementary spatial and textural representations:
  • Histogram of Oriented Gradients (HOG) to capture structural outlines and animal silhouettes.
  • Gray-Level Co-occurrence Matrix (GLCM) to compute second-order statistical texture features (contrast, dissimilarity, homogeneity, energy).
  • LAB Color Space statistics to isolate luminance-invariant color distribution differences from complex vegetation backgrounds.
• Dimensionality Reduction: Evaluated feature importance across 750+ candidate dimensions via Random Forest Mean Decrease in Impurity (MDI), discarding redundant features to curb the curse of dimensionality.
• Active Learning with Class Rebalancing:
  • Addressed extreme class imbalance (under 3% positive animal detections) by coupling Synthetic Minority Over-sampling Technique (SMOTE) with uncertainty-based active learning queries.
  • Strategically prioritized low-confidence boundary samples for label acquisition, minimizing human labeling budget by 60%.
• Model Optimization: Fine-tuned XGBoost hyperparameters via stratified 5-fold GridSearchCV, securing an 0.85 F1-Score on previously unseen test environments.