Dhenkanal Counterfactual Deforestation Modeling

This project implements a comprehensive workflow for estimating deforestation and measuring additionality in the Dhenkanal region of Odisha, India. It combines satellite imagery analysis with machine learning and counterfactual estimation techniques.

Workflow Overview

1. Data Extraction from Google Earth Engine (GEE)

Status: [PLACEHOLDER - Add data extraction details here]

• Source: Google Earth Engine
• Data extracted for Dhenkanal region
• Output directory: data/GEE_exports_Dhenkanal/
• Key datasets:
  • Forest cover rasters (2005, 2010, 2015)
  • Training data with labeled pixels
  • District and jurisdiction boundaries

2. Model Training

Script: model_tif.py

Train Random Forest models on labeled training data:

• Input: Training data from GEE with labeled deforestation pixels
• Process:
  • Load training data with multiple bands (spectral indices, forest cover, etc.)
  • Split data into features and labels
  • Train RF models for different time periods (ex-ante and ex-post)
  • Apply SMOTE for class imbalance handling
  • Save trained models to models/ directory
• Output:
  • rf_smote_x_2000_05_y_2005_10_ex_ante.joblib
  • rf_smote_x_2010_y_2010_15_ex_post.joblib

3. Counterfactual Estimation

Script: counterfactual.py

Generate counterfactual scenarios to estimate what deforestation would have occurred without intervention:

• Input: Trained RF models and raster data
• Process:
  • Create counterfactual predictions using ex-ante models on ex-post time periods
  • Compare actual deforestation (from models) vs. counterfactual scenarios
  • Estimate additionality (avoided deforestation)
• Output: Counterfactual prediction rasters saved to outputs/predictions/

4. Prediction Generation

Script: pred_raster.py

Generate deforestation probability maps for the modeling region:

• Input: Trained RF models and full raster data
• Process:
  • Apply models to full raster extents (or specific sites)
  • Generate probability maps for different scenarios (ex-ante, ex-post)
  • Optionally adjust predictions to match observed deforestation quantities
  • Support both site-only and full-raster predictions
• Output:
  • Deforestation probability rasters: outputs/predictions/deforestation_prob_*.tif
  • Adjusted/density maps
  • Class predictions (binary deforestation)

5. Evaluation, Visualization & Analysis

Area Calculations

Script: calculate_forested_area.py

Calculate forested and deforested areas for each Odisha site:

• Input:
  • Dhenkanal_2010.tif (forest extent)
  • training_data_x_2010_y_2010_15.tif (ground truth deforestation)
  • Odisha_sites.csv (site geometries)
• Process:
  • Total area: All pixels from Dhenkanal_2010 within site geometry
  • Forested area: Pixels with value 1 from Dhenkanal_2010
  • Deforested area: Pixels with value 1 from "9_deforestation" band
• Output: forested_and_deforested_area_by_region.csv

Visualization Scripts

Ground Truth Deforestation

Script: visualize_gt_deforestation.py

Visualize the ground truth deforestation patterns:

• Input: Ground truth raster with "9_deforestation" band and site geometries
• Process:
  • Reproject to high-resolution grid (0.0001 degrees)
  • Mask to individual site boundaries
  • Count deforested pixels and calculate area
  • Generate binary deforestation maps
• Output:
  • PDF and JPG visualizations: figures/gt_deforestation_*.{pdf,jpg}
  • Console output of deforested pixel counts per site

Forest Cover Visualization

Script: visualize_forest_cover.py

Visualize forest extent from Dhenkanal rasters for different time periods.

Model Prediction Visualization

Script: visualize_odisha_sites.py

Visualize RF model predictions and counterfactual scenarios:

• Generates prediction maps for all sites
• Supports different model types (RF model, counterfactual)
• Output: figures/model_prediction_*.{pdf,jpg}, figures/cf_prediction_*.{pdf,jpg}

Additionality Estimation

Script: additionality.py

Calculate additionality metrics (avoided deforestation):

• Input: Counterfactual scenarios and actual predictions
• Process:
  • Compare counterfactual vs. actual deforestation
  • Calculate additionality for each site
  • Generate summary statistics
• Output: additionality_results.json, additionality_table.csv

Additional Analysis

Scripts:

• area_estimation.py: Estimate deforested areas for different model runs
• evaluate_predictions.py: Evaluate model performance metrics
• run_all_evaluations.py: Run comprehensive evaluation suite
• artificial_sites.py: Analysis for artificial/synthetic sites
• file_analysis.py: Data quality checks and file analysis

Directory Structure

CCModeling/
├── data/
│   ├── images/               # Raster data and training data
│   ├── GEE_exports_Dhenkanal/  # Raw exports from GEE
│   └── graph/                # Intermediate data
├── models/                   # Trained RF models (.joblib)
├── outputs/
│   └── predictions/          # Generated prediction rasters
├── figures/                  # Visualization outputs (PDF, JPG)
├── modeling/                 # Python virtual environment
├── archives/                 # Legacy/archived code
└── [Python scripts]          # Main processing scripts

Key Files

• Input Data: data/images/Dhenkanal_*.tif, training_data_*.tif, Odisha_sites.csv
• Models: models/rf_smote_*.joblib
• Results: additionality_table.csv, forested_and_deforested_area_by_region.csv
• Visualizations: figures/gt_deforestation_*.{pdf,jpg}, figures/model_prediction_*.{pdf,jpg}

Usage

1. Extract data from GEE (see placeholder section above)
2. Train models: python model_tif.py
3. Generate counterfactuals: python counterfactual.py
4. Generate predictions: python pred_raster.py
5. Calculate areas: python calculate_forested_area.py
6. Visualize results: python visualize_gt_deforestation.py, python visualize_odisha_sites.py
7. Run full evaluation: python run_all_evaluations.py

Key Metrics

• Total Area: All pixels in Dhenkanal_2010 within site geometry
• Forested Area: Pixels with value 1 (forest)
• Deforested Area: Ground truth deforestation from "9_deforestation" band
• Pixel Size: 0.09 hectares per pixel
• Additionality: Estimated avoided deforestation from counterfactual analysis

Configuration

• Pixel Area: 0.09 ha (based on ~30m resolution native data)
• Coordinate System: EPSG:4326 (WGS84)
• High-resolution grid for visualization: 0.0001 degrees (~10m)
• Model Parameters: RF with SMOTE for class balancing

Requirements

See requirements.txt for Python dependencies (rasterio, geopandas, scikit-learn, matplotlib, etc.)

Notes

• All areas are calculated in hectares
• Ground truth data spans 2010-2015 period
• Models trained for both ex-ante (baseline) and ex-post (intervention) scenarios
• Counterfactual analysis estimates what would have happened without intervention