Species Mapping Project

Overview

This project analyzes species distribution data across different geographic locations, combining environmental features to predict:

1. Classification Task: Which species is present at a given location
2. Regression Task: How many species observations occur at grid locations

---

Dataset

• File: species_with_country_final.csv
• Total Observations: ~270,000 (after cleaning)
• Unique Species: 500
• Geographic Coverage: Global (multiple countries)
• Features:
  • Geographic: latitude, longitude, country
  • Environmental: temperature (avg/max/min), precipitation, solar radiation, water vapor pressure, wind speed
  • Target: species_id, species_count (aggregated)

---

Project Structure

📁 Main Analysis Scripts

1. species_pipeline_augmented.ipynb

Purpose: Complete ML pipeline for species prediction (both classification and regression)

Key Features:

• Handles class imbalance using class_weight='balanced'
• Log transformation for skewed species counts
• Models tested:
  • Logistic Regression
  • Random Forest (best performer)
  • XGBoost
• Cross-validation for robust evaluation
• Comprehensive visualizations

Results (Global Data):

• Classification: Random Forest achieves ~33% accuracy (challenging due to 500 classes and imbalance)
• Regression: Random Forest R² = 0.13 (species count prediction)

Output Directory: plots_augmented/

---

2. species_pipeline_augmented_USA.ipynb ⭐ NEW

Purpose: Same analysis as above but filtered to USA data only

Key Differences:

• Data filtered to country == 'United States of America'
• ~86,000 USA observations (32% of total dataset)
• 116 unique species in USA
• No country feature in model (single country)
• Focused geographic analysis

Why USA Only?

• Homogeneous geographic region
• Potentially better model performance
• Regional species patterns
• Easier to interpret results

Output Directory: plots_USA/

---

3. species_imbalance_skewness_analysis.ipynb ⭐ NEW

Purpose: Deep dive into data quality issues affecting model performance

What It Analyzes:

A. Class Imbalance (Classification Problem)

• Imbalance Ratio: 40:1 (most common vs. least common species)
• Data Concentration: Top 169 species (34%) contain 80% of all data
• Rarity Categories:
  • Very Rare (<50 samples): 0% of species
  • Rare (50-200): 46% of species
  • Common (200-500): 23%
  • Very Common (500-1000): 20%
  • Abundant (1000+): 11%

Visualizations:

• Comprehensive 9-panel class imbalance analysis
• Cumulative distribution (Pareto principle)
• Top/bottom species frequency
• Category breakdowns

B. Skewness (Regression Problem)

• Skewness: 25.93 (extremely right-skewed!)
• Distribution:
  • Mean: 4.79 species/location
  • Median: 1.00 species/location (huge gap!)
• 98.8% of locations have <50 species
• Only 1.2% have 200+ species

Transformations Tested:

• Original: Skewness = 25.93
• Log(1+x): Skewness = 2.13 ✓ (best)
• Sqrt(x): Skewness = 8.71

Visualizations:

• Comprehensive 9-panel skewness analysis
• Q-Q plots (original vs. transformed)
• CDF and percentile analysis
• Transformation comparisons

Output Directory: imbalance_analysis/

Key Findings:

• ⚠️ Severe class imbalance makes classification very challenging
• ⚠️ Extreme skewness requires log transformation for regression
• ✓ Tree-based models (RF, XGBoost) handle these issues better than linear models

---

4. species_pipeline_deep_learning.ipynb ⭐ NEW

Purpose: Neural network approach as an alternative to traditional ML

Question Answered: Can deep learning improve performance over Random Forest and XGBoost?

Architecture:

Classification Model

Input (features) → Dense(512) → BatchNorm → Dropout(0.3)
                 → Dense(256) → BatchNorm → Dropout(0.3)
                 → Dense(128) → BatchNorm → Dropout(0.3)
                 → Dense(500, softmax) → Output (species)

Regression Model

Input (features) → Dense(256) → BatchNorm → Dropout(0.3)
                 → Dense(128) → BatchNorm → Dropout(0.3)
                 → Dense(64)  → BatchNorm → Dropout(0.3)
                 → Dense(1, linear) → Output (count)

Deep Learning Techniques Used:

• ✓ Batch Normalization: Stabilizes training
• ✓ Dropout: Prevents overfitting
• ✓ Class Weights: Handles imbalance
• ✓ Early Stopping: Prevents overfitting
• ✓ Learning Rate Scheduling: Adaptive learning
• ✓ Log Transformation: Handles skewness

Expected Performance:

• Neural networks typically excel with:

  • Very large datasets (millions of samples)
  • Complex feature interactions
  • Unstructured data (images, text)

• For this tabular dataset with moderate size:

  • Random Forest likely remains competitive
  • Deep learning may not significantly outperform
  • Tree models are naturally suited for tabular data

Output Directory: plots_deep_learning/

Models Saved:

• best_cls_model.h5 (classification)
• best_reg_model.h5 (regression)

---

5. detailed_analysis_and_viz.ipynb (Reference)

Purpose: Original exploratory data analysis (EDA)

Contains:

• Interactive visualizations (Plotly)
• Geographic distribution maps
• Country-level analysis
• Temperature vs. diversity patterns

Output Directory: detailed_analysis/

---

Key Challenges & Solutions

Challenge 1: Class Imbalance

Problem: Species distribution is highly imbalanced (40:1 ratio)

Solutions Implemented:

• ✓ class_weight='balanced' in scikit-learn models
• ✓ Stratified sampling (maintains class proportions)
• ✓ Weighted loss in neural networks
• ✓ Focus on weighted F1-score, not just accuracy

Not Implemented (computationally expensive):

• SMOTE (Synthetic Minority Over-sampling)
• Data augmentation

---

Challenge 2: Extreme Skewness

Problem: Species counts are extremely right-skewed (skewness = 25.93)

Solutions Implemented:

• ✓ Log1p transformation: y_transformed = log(1 + y)
• ✓ Inverse transform for predictions: y_pred = exp(y_pred_log) - 1
• ✓ Tree-based models (robust to skewness)

Why Log1p:

• Handles zero values (log(0) is undefined, but log(1+0) = 0)
• Reduces skewness from 25.93 → 2.13
• Allows use of MSE/RMSE metrics

---

Challenge 3: High Dimensionality

Problem: 500 species classes + multiple countries

Solutions Implemented:

• ✓ One-hot encoding for country features
• ✓ StandardScaler for numerical features
• ✓ Dimensionality reduction via feature selection
• ✓ Ensemble methods (reduce overfitting)

---

Model Comparison

Classification Results (Global Data)

Model	Test Accuracy	CV Accuracy	Training Time	Notes
Logistic Regression	0.162	0.161	~22 min	Linear model struggles
Random Forest	0.331	0.327	~3 min	Best performer
XGBoost	0.097	0.084	~10 min	Underperforms unexpectedly
Neural Network (MLP)	TBD	TBD	TBD	See deep_learning notebook

Why Random Forest Wins:

• Handles class imbalance well with class_weight='balanced'
• Robust to feature scaling
• Can model complex non-linear relationships
• Less prone to overfitting than XGBoost on this dataset

---

Regression Results (Global Data)

Model	Test R²	CV R²	RMSE	Training Time	Notes
Linear Regression	-0.011	0.051	16.18	~1s	Fails on skewed data
Random Forest	0.126	0.224	15.05	~38s	Best performer
XGBoost	0.076	0.183	15.47	~2s	Decent performance
Neural Network (MLP)	TBD	TBD	TBD	TBD	See deep_learning notebook

Why Random Forest Wins:

• Robust to outliers and skewness
• Handles spatial patterns well
• No need for complex feature engineering
• Fast training on moderate-sized datasets

---

Can Deep Learning Help?

Expected Scenarios:

✅ Deep Learning May Help If:

1. Very large dataset (millions of samples)
   • Current: ~270k samples → moderate size
2. Complex feature interactions (non-linear, high-order)
   • Current: mostly environmental features (temperature, precipitation)
3. Unstructured data (images, text, time series)
   • Current: tabular data
4. Need for transfer learning
   • Current: no pre-trained models applicable

❌ Deep Learning May NOT Help Because:

1. Tabular data: Tree models (RF, XGBoost) are state-of-the-art
2. Moderate dataset size: Not enough data to train deep networks effectively
3. High class imbalance: Neural networks struggle more than tree models
4. Interpretability: Tree models provide feature importance easily

Verdict:

Traditional ML (Random Forest) likely remains the best choice for this specific dataset. However, the deep learning notebook is provided as:

• An alternative approach to explore
• A learning exercise in neural network architectures
• A baseline for future improvements (e.g., with more data)

---

How to Use This Project

1. Run the Notebooks in Order

For Complete Analysis:

1. detailed_analysis_and_viz.ipynb         # EDA (optional)
2. species_imbalance_skewness_analysis.ipynb  # Understand data issues
3. species_pipeline_augmented.ipynb        # Train ML models (global)
4. species_pipeline_augmented_USA.ipynb    # Train ML models (USA only)
5. species_pipeline_deep_learning.ipynb    # Try deep learning (optional)

For Quick Results:

1. species_pipeline_augmented.ipynb        # Global ML models
OR
2. species_pipeline_augmented_USA.ipynb    # USA-only ML models

---

2. Dependencies

# Core libraries
pip install pandas numpy matplotlib seaborn scikit-learn

# Advanced ML
pip install xgboost

# Deep learning (optional)
pip install tensorflow

# Interactive visualizations (optional)
pip install plotly

---

3. Expected Outputs

Each notebook generates:

• PNG plots (static visualizations)
• HTML dashboards (interactive, where applicable)
• TXT reports (model performance summaries)
• CSV summaries (statistics and comparisons)

Output Directories:

• plots_augmented/ - Global ML results
• plots_USA/ - USA-only ML results
• imbalance_analysis/ - Data quality analysis
• detailed_analysis/ - EDA outputs
• plots_deep_learning/ - Neural network results

---

Key Insights

1. Data Quality Issues Are Major Challenges

• Class imbalance (40:1 ratio) severely limits classification accuracy
• Extreme skewness (25.93) makes regression difficult
• Top 34% of species contain 80% of data

2. Random Forest Is the Best Traditional ML Model

• Classification: 33% accuracy (given 500 classes and imbalance, this is decent)
• Regression: R² = 0.13 (moderate performance)
• Robust to imbalance and skewness

3. Feature Engineering Could Help

• Spatial features (k-NN neighbors, distance to biodiversity hotspots)
• Temporal features (season, year)
• Interaction terms (temperature × precipitation)

4. More Data Would Help

• Current: 50-2000 samples per species
• Ideal: 1000+ samples per species for rare classes
• Additional features: soil type, elevation, habitat characteristics

---

Recommendations for Improvement

For Classification:

1. Focus on top N species (e.g., top 100 with most data)
   • Reduces classes, improves accuracy
2. Hierarchical classification (family → genus → species)
   • Break down the problem into stages
3. Ensemble methods (stacking, blending)
   • Combine multiple models
4. More features (habitat type, ecosystem)
   • Richer information for predictions

For Regression:

1. Add spatial features (neighboring cell counts)
   • Species cluster geographically
2. Quantile regression (instead of mean prediction)
   • Better handles outliers
3. Time series analysis (if temporal data available)
   • Seasonal patterns in species counts
4. Geographically weighted regression
   • Local models for different regions

---

File Descriptions

Notebooks (.ipynb)

• species_pipeline_augmented.ipynb - Main ML pipeline (global)
• species_pipeline_augmented_USA.ipynb - ML pipeline (USA only) ⭐ NEW
• species_imbalance_skewness_analysis.ipynb - Data quality analysis ⭐ NEW
• species_pipeline_deep_learning.ipynb - Neural network approach ⭐ NEW
• detailed_analysis_and_viz.ipynb - EDA and visualizations
• species_pipeline_augmented_latest.ipynb - Alternative version
• species_pipeline_original.ipynb - Original baseline

Data Files

• species_with_country_final.csv - Main dataset (with country info)
• species_train.npz - Training data (numpy format)
• species_test.npz - Test data (numpy format)

Output Directories

• plots_augmented/ - Global ML results
• plots_USA/ - USA-only results ⭐ NEW
• imbalance_analysis/ - Data issue analysis ⭐ NEW
• plots_deep_learning/ - Neural network results ⭐ NEW
• detailed_analysis/ - EDA outputs

---

Frequently Asked Questions

Q1: Why is classification accuracy only 33%?

A: With 500 species classes and severe imbalance (40:1), this is actually reasonable:

• Random guessing would give 0.2% accuracy (1/500)
• 33% means the model is learning meaningful patterns
• Focus on top-k accuracy (top-5, top-10) for better metrics

Q2: Why use log transformation for regression?

A: The target (species count) is extremely skewed:

• Mean (4.79) >> Median (1.00)
• Skewness = 25.93 (very high)
• Log1p reduces this to 2.13, enabling better model training

Q3: Should I use deep learning or Random Forest?

A: For this dataset, Random Forest is recommended because:

• Tabular data (tree models excel here)
• Moderate dataset size (not big enough for deep learning)
• Class imbalance (trees handle better)
• Faster training and better interpretability

However, try deep learning if:

• You have more data (millions of samples)
• You want to learn neural network techniques
• You can leverage GPUs for faster training

Q4: Why analyze USA data separately?

A: Several reasons:

1. Homogeneous region: Similar climate/geography patterns
2. Data concentration: USA has 32% of all observations
3. Simpler model: No need for country feature
4. Better performance: Potentially higher accuracy in focused region

Q5: Can I apply this to other species datasets?

A: Yes! The methodology is generalizable:

1. Handle class imbalance (class weights, stratified sampling)
2. Transform skewed targets (log1p)
3. Use tree-based models (Random Forest, XGBoost)
4. Evaluate with appropriate metrics (weighted F1, R²)

---

Citation

If you use this project, please cite:

Species Mapping Project
Applied Machine Learning Mini Project
November 2025

---

Contact

For questions or issues, please refer to the notebook comments or reach out via the project repository.

---

Version History

• v1.0 (Nov 2025) - Initial release with augmented pipeline
• v2.0 (Nov 2025) - Added USA-specific analysis, imbalance/skewness analysis, and deep learning approach

---

Happy Modeling! 🌿🐾