Modeling forest aboveground biomass in cloud-prone tropical mountains using temporally encoded Sentinel-2 time series and synthetic training data

Accurate modeling of forest aboveground biomass (AGB) in cloud-prone tropical mountains remains constrained by irregular optical remote-sensing observations and limited field-reference data. This study proposes a 10 m AGB modeling framework that combines Sentinel-2 time-series data with regression training based on synthetic samples. Sentinel-2 observations from 2019 to 2021 were folded into target-year annual sequences and encoded as weekly spectral tensors to preserve irregular cloud-free observation patterns. Field-derived AGB samples from northern Vietnam were spatially aggregated to 10 m resolution and used to generate approximately 128,000 synthetic training samples through weighted spectral–biomass mixing. A one-dimensional convolutional neural network (1D-CNN) regression model was then trained using the synthetic samples and evaluated against independent field-derived test samples. The synthetic-data model (Model-S) achieved MAE = 1.19 kg m⁻², RMSE = 1.61 kg m⁻², and R² = 0.87 across all validation locations, improving MAE by 32%, RMSE by 31%, and R² by 18% relative to the model trained only with real samples (Model-R). The resulting 10 m AGB map revealed moderate regional biomass levels and localized high-biomass patches in the western mountainous areas. This framework provides a scalable approach for biomass mapping in regions where field plots are sparse and Sentinel-2 optical time series are irregular.