Accurate estimation of aboveground carbon stocks in tropical rainforests is fundamental for global climate change mitigation, sustainable forest management, and carbon accounting. Traditional field-based carbon assessment methods are often time-consuming, labour-intensive, and spatially limited, making them unsuitable for large-scale or high-frequency monitoring. Recent advances in remote sensing have provided access to diverse data sources such as optical imagery, radar backscatter, and LiDAR point clouds, yet integrating these heterogeneous datasets effectively remains a major challenge. This study proposes a deep learning framework using an auto-encoder architecture to estimate forest carbon stocks by fusing multi-source remote sensing data. The framework systematically optimizes activation functions, learning rates, and encoder-decoder configurations to enhance model accuracy, stability, and computational efficiency. Experiments were conducted using multi-sensor datasets from tropical rainforests in Borneo, incorporating field-measured biomass data as ground truth. A comprehensive evaluation was performed using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Index of Agreement (IOA) metrics. The results indicate that ReLU, LeakyReLU, and Softplus activation functions deliver consistent performance across experiments, while RMSprop and Adam optimizers achieve the best convergence stability. The optimized encoder-decoder configuration significantly improves prediction accuracy compared to baseline models. Compared to the baseline encoder-decoder configuration, the proposed optimized architecture achieves a 2.7−3.2% reduction in RMSE, indicating incremental but consistent improvements attributable to architectural balancing rather than increased depth alone. The findings underscore the importance of multi-sensor fusion and systematic hyperparameter optimization for large-area carbon estimation. The proposed framework is scalable, adaptable to different forest ecosystems, and reproducible for other regions worldwide. This research contributes to the advancement of AI-driven environmental monitoring and provides a robust tool to support policy implementation, carbon credit verification, and sustainable forest resource management under Malaysia’s and global climate action agendas.
Khoo et al. (Fri,) studied this question.