This repository contains the complete code and supporting material for our study on individual tree-level estimation of Above Ground Biomass (AGB) using multi-modal remote sensing data and deep learning. The proposed model does not rely on LiDAR data for prediction, making it suitable for operational scalability across large forested regions.
Accurate AGB estimation plays a crucial role in forest carbon accounting and ecological monitoring. Traditional approaches rely heavily on field data or LiDAR, which are costly and difficult to scale. This study introduces DeepBioFusion, a CNN-based architecture that fuses:
- 🛰️ Optical imagery (e.g., RGB or multispectral)
- 📡 Synthetic Aperture Radar (SAR) data (X, C, and L bands)
- 🌳 Tree species maps derived from the Norwegian Forest Inventory
- 🚫 LiDAR-free prediction: Model predicts biomass without needing LiDAR input at inference.
- 🌐 Scalable design: Adaptable to any region with satellite access.
- 🌲 Individual tree resolution: Focused on single-tree level predictions.
- 🔁 Reproducible workflow: Includes complete pipeline for training and evaluation.
The end-to-end pipeline includes segmentation, patch extraction, species labeling, and biomass estimation using a multi-branch CNN.
This diagram summarizes the complete DeepBioFusion pipeline — from LiDAR segmentation to AGB prediction using CNNs and multi-modal fusion.
Estimated AGB values are visualized spatially, reflecting biomass distribution across a Norwegian forested region.
Each individual tree is classified into species groups (e.g., Spruce, Pine, Deciduous), forming the categorical input for our model.
LiDAR-derived terrain and canopy height models used for initial tree crown segmentation and quality assessment.
Scatter plot comparing predicted biomass values with field-based measurements. The high correlation supports model validity.
git clone https://github.com/Ci2Lab/DeepBioFusion.git
cd DeepBioFusion
pip install -r requirements.txt