Main Article Content
Abstract
Accurate, high-resolution mapping of
Aboveground Biomass or AGB in complex tropical
landscapes remains a critical challenge for climate change
mitigation and carbon accounting. This is primarily due to the
signal saturation of widely-used optical Sentinel-2 and C-
Band radar Sentinel-1 sensors in dense forests. This study
presents a robust data-fusion methodology to overcome this
limitation by integrating L-Band radar from ALOS PALSAR
and testing a suite of advanced machine learning models in the
heterogeneous landscape of Yogyakarta, Indonesia. We fused
GEDI L4A AGBD data from 2020 with a comprehensive
feature dataset derived from Sentinel-2 optical data, Sentinel-
1 C-Band texture, ALOS PALSAR L-Band data, and SRTM
topography. This study empirically demonstrates the failure
of standard sensors. A model trained on only Sentinel-1 and
Sentinel-2 data was unable to explain AGBD variance,
achieving a coefficient of determination of approximately
0.18. However, by fusing this with L-Band radar and
topographic data, the model's performance more than
doubled to a coefficient of determination of approximately
0.34, proving L-Band is an essential predictor. A competition
between four models, MLR, RF, SVR, and MLP, was
conducted on a feature-selected dataset. The Multi-Layer
Perceptron, was found to be the most accurate predictor with
a final R-squared of 0.3389 and an RMSE of 74.26 tons per
hectare. While this R-squared value is moderate, it highlights
the inherent noise of the GEDI L4A product and the extreme
complexity of the fragmented study area. We conclude that L-
Band data are critical components for advancing AGB
mapping in the tropics.