SCE-BiLSTM: A Hybrid Deep Learning Model for Regional Forest Biomass Estimation with Spatial-Channel Attention and Extreme Learning

Aboveground biomass (AGB) is crucial for ecosystem monitoring, forest surveys, and management. Accurate and efficient AGB estimation remains challenging, as large-scale machine learning methods often sacrifice accuracy, while deep learning models enhance precision but struggle with efficiency and generalization. To address this, we propose an advanced deep learning framework (SCE-BiLSTM) for regional AGB inversion, integrating spatial (SAM) and channel attention mechanisms (CAM) to improve feature extraction. An extreme learning machine (ELM) enhances efficiency by randomly learning weights and thresholds. Using 11 remote sensing features from Luoyang forests and GEDI L4A data, the model outperforms CNN-BiLSTM, reducing MAE by 3.59 Mg/ha, RMSE by 6.46 Mg/ha, and increasing R² to 0.9052, with runtime reduced by 19 seconds. Validation in the Yellow River region shows strong generalization, achieving an MAE of 11.48 Mg/ha, RMSE of 14.72 Mg/ha, and R² of 0.8335. A time-series analysis from 2015 to 2023 reveals spatial and temporal AGB variations, highlighting influencing factors. These results demonstrate the framework’s potential for accurate, scalable biomass assessments, providing valuable insights for sustainable forest management.