Aboveground Biomass Density Estimation Using Deep Learning: Insight from NEON Ground-Truth Data and Simulated GEDI Waveform

Abstract

Accurate estimation of forest Aboveground Biomass Density (AGBD) is essential for understanding and managing Earth's carbon cycle and informing climate change mitigation strategies. NASA's Global Ecosystem Dynamics Investigation (GEDI) mission has advanced global forest structure mapping; however, its AGBD estimates exhibit significant variability across forest types and regions due to limitations in model transferability and training data. This study aims to enhance AGBD estimation by leveraging deep learning techniques to model complex, non-linear relationships in ecological data, using extensive ground-truth data from the National Ecological Observatory Network (NEON) and simulated GEDI waveforms. We developed and compared various deep learning models, including one-dimensional convolutional neural networks (1D CNNs), long short-term memory (LSTM) networks, and pre-trained convolutional neural networks (CNNs), against traditional linear regression and random forest models. Using data from 2,948 NEON plots across diverse ecological settings in the United States, with AGBD calculated from NEON's Vegetation Structure data, and simulated GEDI waveforms generated from NEON's LiDAR point clouds, we found that deep learning models, particularly pre-trained CNNs, significantly outperformed traditional methods. The ResNet152 model achieved an R² of 0.68 for live biomass estimation, compared to 0.58 for linear regression, demonstrating up to 17% improvement. However, deep learning model performance declined with reduced dataset sizes, emphasizing the need for large, diverse training data. Stratification by plant functional types revealed that different models were more effective for specific vegetation categories, underscoring the importance of tailoring model selection to distinct forest ecosystems. Our findings indicate that deep learning models offer significant potential for improving AGBD estimation accuracy, contributing to more precise assessments of forest carbon stocks and enhancing the effectiveness of climate change mitigation efforts.

Table of Contents

Location	Description	Files
Models	This repository contains various models designed to predict Aboveground Biomass density using Simulated GEDI Waveform data. The models include OLS, 1D-CNN, CNN, LSTM, and Random Forest approaches. Each model has been implemented in Python using PyTorch or other relevant libraries, with detailed scripts for training and evaluation. For detailed research refer to our research paper (under review) - Aboveground Biomass Density Estimation Using Deep Learning: Insight from NEON Ground-Truth Data and Simulated GEDI Waveform	- 1D-CNN-Model.py - CNN-Model.py - LSTM-Model.py - OLS-Model.ipynb - RF-Model.ipynb
NEON	This repository contains several Python notebooks that were used for processing and extracting data for developing and training Deep learning models to estimate AGBD, such as PFT (Plant Functional Types), GEDI simulated waveforms, and more. Each script handles specific tasks such as clipping, normalization, and feature extraction from NEON and simulated GEDI datasets.	- Clip-DTM-NEON.ipynb - Normalize-withDTM-NEON.ipynb - PFT-Extractor.ipynb - RH-Extractor.ipynb - Remove-Blackborders-fromClipped-NEON.ipynb - Simulated-GEDI-Waveform-Visualizer.ipynb