Optical Remote Sensing and Machine Learning Enable Scalable Carbon Mapping in Himalayan Forests Despite Spectral Saturation Challenges
Abstract
Accurate aboveground biomass (AGB) estimation is critical for climate change mitigation and carbon accounting in mountain regions, yet spatially explicit mapping remains limited by reliance on conventional field methods and underexplored machine learning approaches in the Himalaya. Bhutan, a carbon-negative nation maintaining 70% forest cover, faces urgent demand to develop robust monitoring, reporting, and verification (MRV) systems for carbon market participation under the Paris Agreement Article 6. This study developed a random forest (RF) model that integrates Sentinel-2A multispectral imagery, nine vegetation indices, topographic variables, and field-measured AGB from 99 plots across 244.6 hectares in the Kyentshen Community Forest. Using recursive feature elimination, 10 key predictors were identified from 27 candidate variables: Band 11, SLAVI, Band 12, WDRVI, NDVI, elevation, NDWI, GNDVI, band ratio (r3), and Band 9. The optimized RF model achieved R² = 0.34 and RMSE = 67.98 ± 19.79 t/ha on training data (Ntree = 500, Mtry = 3). However, test-set performance declined substantially (R² = 0.42, RMSE = 86.14 t/ha), indicating overfitting and reflecting limitations of spectral saturation in high-biomass stands (>300 t/ha). A Wilcoxon signed-rank test revealed systematic overestimation (predicted median: 118.36 t/ha vs. observed: 105.15 t/ha; p = 0.03). Despite moderate accuracy, this study demonstrates that freely available Sentinel-2 data combined with machine learning offers a cost-effective, scalable alternative to traditional field mapping for rapid AGB assessment across Bhutan's temperate forests. Vegetation indices, particularly SLAVI and WDRVI, proved superior to raw spectral bands, offering improved sensitivity to dense canopies. The spatial-transect methodology and 10-predictor framework provide a replicable foundation for advancing Bhutan's forest carbon accounting systems. We recommend integrating synthetic aperture radar (SAR) or LiDAR to overcome optical saturation, enabling higher accuracy for robust MRV compliance and carbon credit monetization. This work directly supports the maintenance of Bhutan's carbon-neutral status and establishes the technical baseline for scaling AGB monitoring across mountain regions where engineered infrastructure capacity is limited, but climate action urgency is paramount.
Related
- P-T-t-d history of the Greater Himalayan Sequence in the Zanskar Shear Zone, NW India
- Modern Himalayan marmot faeces analyses: Implications for palaeodietary, palaeoecological, and palaeoclimate analysis
- The Current State of Research on Deformation, Strain, and Velocity across the Alpine-Himalayan Belt Using Transcontinental Sentinel-1 InSAR and GNSS
- Transboundary outburst floods from giant supraglacial lakes over the Himalayan China–Nepal border
- Microbial quotient responses to altitudinal and edaphic variability in Himalayan apple orchard soil
BibTeX
@misc{wangchuk2026optical,
title = {{Optical Remote Sensing and Machine Learning Enable Scalable Carbon Mapping in Himalayan Forests Despite Spectral Saturation Challenges}},
author = {Namgay Wangchuk and Yonten Dorji},
howpublished = {EcoEvoRxiv},
year = {2026},
month = sep,
doi = {10.32942/x29d5f},
url = {https://ecoevorxiv.org/repository/view/14866/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Optical Remote Sensing and Machine Learning Enable Scalable Carbon Mapping in Himalayan Forests Despite Spectral Saturation Challenges
AU - Namgay Wangchuk
AU - Yonten Dorji
PY - 2026
DA - 2026/09/09
PB - EcoEvoRxiv
DO - 10.32942/x29d5f
UR - https://ecoevorxiv.org/repository/view/14866/
AB - Accurate aboveground biomass (AGB) estimation is critical for climate change mitigation and carbon accounting in mountain regions, yet spatially explicit mapping remains limited by reliance on conventional field methods and underexplored machine learning approaches in the Himalaya. Bhutan, a carbon-negative nation maintaining 70% forest cover, faces urgent demand to develop robust monitoring, reporting, and verification (MRV) systems for carbon market participation under the Paris Agreement Article 6. This study developed a random forest (RF) model that integrates Sentinel-2A multispectral imagery, nine vegetation indices, topographic variables, and field-measured AGB from 99 plots across 244.6 hectares in the Kyentshen Community Forest. Using recursive feature elimination, 10 key predictors were identified from 27 candidate variables: Band 11, SLAVI, Band 12, WDRVI, NDVI, elevation, NDWI, GNDVI, band ratio (r3), and Band 9. The optimized RF model achieved R² = 0.34 and RMSE = 67.98 ± 19.79 t/ha on training data (Ntree = 500, Mtry = 3). However, test-set performance declined substantially (R² = 0.42, RMSE = 86.14 t/ha), indicating overfitting and reflecting limitations of spectral saturation in high-biomass stands (>300 t/ha). A Wilcoxon signed-rank test revealed systematic overestimation (predicted median: 118.36 t/ha vs. observed: 105.15 t/ha; p = 0.03). Despite moderate accuracy, this study demonstrates that freely available Sentinel-2 data combined with machine learning offers a cost-effective, scalable alternative to traditional field mapping for rapid AGB assessment across Bhutan's temperate forests. Vegetation indices, particularly SLAVI and WDRVI, proved superior to raw spectral bands, offering improved sensitivity to dense canopies. The spatial-transect methodology and 10-predictor framework provide a replicable foundation for advancing Bhutan's forest carbon accounting systems. We recommend integrating synthetic aperture radar (SAR) or LiDAR to overcome optical saturation, enabling higher accuracy for robust MRV compliance and carbon credit monetization. This work directly supports the maintenance of Bhutan's carbon-neutral status and establishes the technical baseline for scaling AGB monitoring across mountain regions where engineered infrastructure capacity is limited, but climate action urgency is paramount.
ER -