Spatially-aware machine learning predictions of tropical forest structure using GEDI metrics and multi-source earth observation data
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
The raster layers provided here were gererated from wall-to-wall predictions of GEDI metrics forest structure characteristics (canopy height and cover) of a diverse tropical landscape in Papua New Guinea (PNG) using spatially-aware machine learning modelling, synergising multi-source earth observation (EO) data particularly, Sentinel-1 SAR, Sentinel-2 Optical, AlphaEarth Satellite Embedding, environmental variables and synthetic spatial features. In addition, the CSV files contain the out-of-fold (OOF) prediction data, overall prediction metrics, and the feature importance for each model prediction. A key susequent post-process was the creation of a holistic Forest Structure Index (FSI) from the model predictions of multiple GEDI metrics forest structure characteristics using a performance-weighted summation approach. Apart from PNG, we demonstrate the operational scalability and cross-biome transferability of the FSI framework for broad pantropical landscapes particularly in the Congo basin and Amazon tropical forest ecosystems. We scaled FSI for a tropical coastal landscape in the Republic of Congo and an inland tropical landscape of Bolivia and their FSI raster layers are available here together with that of the PNG tropical landscape. Overall, we verified that the FSI is a stable ecosystem index given its standard consistency of index value range across all tropical landscapes tested in this study. The paper is currently under review by the GIScience & Remote Sensing Journal, Taylor & Francis. As part of the Journal's Data Sharing Policy (Taylor & Francis open + FAIR data policy), this dataset is being openly shared under CCBY & CC0 Licenses. Paper Title: Tropical forest structure mapping using spatially-aware machine learning with GEDI metrics and multi-source earth observation data Paper Highlights Introduction of a mean GEDI canopy height metric (RHm) to minimize inter-percentile sensitivity Development of a Spatially-Aware LightGBM (SA-LGBM) optimization framework (with nested Bayesian Optimizer) Local Outlier Factor thresholding Eigenvector Spatial Filtering Buffered 5-fold Spatial Block Cross-validation Creation of a comprehensive forest structure index (FSI) from GEDI multi-metric model predictions First spatially-explicit GEDI-derived forest structure wall-to-wall mapping effort specific to PNG Connect on LinkedIn: Glen R Yali
Fields of Study & Subjects
Remote Sensing and LiDAR Applications
Remote Sensing in Agriculture
Forest ecology and management
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 27 March 2026
- Language
- EN
- Document type
- Dataset