Advancing Mapping and Monitoring of Forest Structure and Aboveground Biomass using Multisource Earth Observation Data

Abstract

Accurate estimation of forest aboveground biomass (AGB) is essential for understanding forest carbon dynamics and informing climate change mitigation strategies. The availability of spaceborne light detection and ranging (lidar) data from the Ice Cloud and Land Elevation Satellite-2 (ICESat-2) and the Global Ecosystem Dynamics Investigation (GEDI) have provided unprecedented opportunities for large area mapping of forest canopy height (CH) and AGB. This study aimed to advance methodological frameworks for CH and AGB estimation by leveraging recent advances in spaceborne lidar capabilities from both ICESat-2 and GEDI missions. The primary objectives of these studies were to: (1) investigate how spaceborne lidar has progressed over the past twenty years to advance forest AGB mapping, (2) create a data-driven cloud-based workflow for statewide AGB mapping using GEDI data in Alabama, (3) understand the synergistic potential of ICESat-2 and GEDI for advancing CH and AGB estimation and (4) develop spatially explicit time-series estimates of CH and AGB to enable annual monitoring of Alabama forests (2015–2024). We employed data from a range of multispectral, synthetic aperture radar (SAR) (e.g., Sentinel-1/2, Landsat-8/9, ALOS PALSAR) and ancillary variables. Six Machine Learning (ML) and Deep Learning (DL) algorithms, including random forest (RF), ridge regression, light gradient boosting machine (LGBM), extreme gradient boosting (XGBoost), categorical boosting (CatBoost), and convolutional neural networks (CNN), were evaluated. In Chapter 1, we conducted a systematic review of spaceborne lidar studies over the past two decades, following Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. The results of the review highlighted the need for improved understanding of the synergistic use of spaceborne lidar, cloud computing, high-performance computing (HPC), and time series analysis. In Chapter 2, we assessed the incorporation of readily available discrete GEDI AGB for wall-to-wall AGB estimation using RF ML in Google Earth Engine (GEE) and discovered that the ecoregion-specific model substantially outperformed statewide models. Additionally, a key limitation identified was the scarcity of samples when utilizing a single spaceborne lidar. Subsequently, in Chapter 3, we investigated a robust framework to advance CH and AGB estimations by combining GEDI and ICESat-2 in HPC while evaluating six learning algorithms. Results showed that XGBoost achieved the highest accuracy for CH estimation and LGBM for AGB estimation, when incorporating CH in the analysis (coefficient of determination (R²) increased from 0.59 to 0.62, and the root mean squared error (RMSE) decreased from 13.03 Mgha-1 to 12.60 Mgha-1 when CH was included in AGB estimation). In Chapter 4, we employed the established framework in Chapter 3 for time series analysis of CH and AGB over the last decade (2015 - 2024). The time series CH and AGB maps achieved good correlations (Pearson correlation coefficients for CH ranged from 0.66 to 0.74, and for AGB, from 0.54 to 0.74) when evaluated using the National Ecological Observatory Network CH models and the European Space Agency's Climate Change Initiative AGB product. This study advances current understanding of spaceborne lidar applications for mapping of forest structure and productivity. The developed frameworks and mapped products demonstrate effective data-driven methodologies and will facilitate cost-effective and time-efficient decision-making in forest management.