Above-Ground Biomass Prediction Dataset for Northeastern India

Satellite-Based Pixel-Level Estimates for Assam and Mizoram Forests

This dataset contains satellite-based above-ground biomass (AGB) predictions for two locations in northeastern Indian forests. It is specifically optimized for the forest regions in Assam and Mizoram. The predictions are generated using a deep learning model trained on ecologically similar forest sites across South and Southeast Asia. This work was developed by vertify.earth as part of the digital Monitoring, Reporting, and Verification (dMRV) for Himalayasproject, funded by Lacuna.

Dataset Structure

The dataset contains two locations. Each location includes four files.

File Descriptions

1. Input Images (input_image.tif)

• Multi-sensor satellite data used for biomass prediction

• Includes Sentinel-1 (SAR), Sentinel-2 (optical), and DEM data

• Spatial resolution: 10 to 40 meters, depending on the sensor

• Format: Multi-band GeoTIFF with geospatial referencing

2. Predicted Biomass Maps (predicted_biomass.tif)

• Pixel-level biomass estimates in Mg/ha (megagrams per hectare)

• Generated using StableResNet deep learning architecture

• Model performance:

  • R² = 0.87

  • RMSE = 28.7 Mg/ha

  • MAE = 19.5 Mg/ha

• Format: Single-band GeoTIFF

• Typical value range: 40 to 460 Mg/ha

3. Visualization Maps (visualization.png)

• Color-coded biomass maps for visual reference

• PNG format, suitable for reports and presentations

• Optimized color scales and legends for biomass density

4. Summary Statistics (statistics.txt)

• Summary statistics of predicted biomass values

• Includes mean, median, standard deviation, and min/max

• Includes spatial distribution analysis

Model Information

Training Data

The model was trained on forests in:

• India: Yellapur, Betul, Achanakmar

• Thailand: Khaoyai

Sites were chosen for ecological similarity to northeastern India:

• Monsoon-influenced climate

• Comparable forest types (evergreen, semi-evergreen, moist deciduous)

• Hilly terrain and similar biomass density range

Model Architecture

Custom StableResNet with:

• Residual connections

• Layer normalization

• Designed for pixel-level regression stability

Feature Engineering

Features used include:

• Spectral indices (NDVI, EVI, NDWI)

• Texture features (LBP, GLCM)

• Spatial gradients

• PCA components

Regional Applicability

This dataset is tailored for:

• Primary Use: Biomass mapping in Assam and Mizoram

• Also Applicable To: Other northeastern Indian states

• Forest Types: Tropical and subtropical, monsoon-affected

• Terrain: Hilly and mountainous forest regions

Project Context

• Organization: vertify.earth

• Project: Digital Monitoring, Reporting, and Verification (dMRV) for Himalayas

• Funding: Lacuna Fund

• Purpose: Forest carbon monitoring and REDD+ support

Technical Specifications

Usage Applications

• Forest carbon stock assessments

• REDD+ monitoring and reporting

• Forest management planning

• Climate change and biodiversity studies

• Research on tropical forest dynamics

Data Quality and Validation

• Validated against LiDAR-derived biomass data

• Cross-validation across multiple forest sites

• Performance Metrics:

  • R² = 0.87

  • RMSE = 28.7 Mg/ha

  • MAE = 19.5 Mg/ha

• Includes quality flags for limited satellite coverage areas

Citation

If you use this dataset, please cite: