A hybrid deep learning model based on CNN-GRU-BiLSTM for predicting the carbon removal capacity of the living standing tree using multi-source variables

Abstract

Accurate prediction of the carbon removal capacity of the living standing tree holds significant importance in understanding plant growth mechanisms, carbon balance, and environmental adaptability. This study investigated the entire Radermachera sinica trees using an enhanced static assimilation chamber integrated with IoT technology for automated, continuous monitoring. A novel self-designed stem water content sensor was introduced to capture the dynamics of stem water content, while simultaneously monitoring climate and soil factors. The relationship between the carbon removal capacity and multi-source variables was analyzed. A hybrid deep learning model based on Convolutional Neural Network (CNN), Gated Recurrent Unit (GRU), and Bidirectional Long Short-Term Memory (BiLSTM) (CNN-GRU-BiLSTM) was developed to predict the carbon removal capacity, with performance evaluated against traditional machine learning and simpler deep learning models. The results indicate that: 1) The carbon removal capacity exhibited a distinct "U-shaped" diurnal variation pattern. 2) In most cases, there was close alignment between the transition moments of positive and negative values of the carbon removal capacity and stem water change rate, showing a significant positive correlation. 3) The hybrid model can effectively capture key influences of multi-source variables on carbon removal capacity, markedly improving regression prediction accuracy over traditional and alternative models. The RMSE, MAE, MAPE, and R² values can reach 1.1221 %, 0.75879 %, 0.29614 %, and 0.91191, respectively. The findings underscore the pivotal role of stem water dynamics in carbon removal prediction and provide a foundational model framework for assessing carbon sequestration and ecological responses in plant communities under variable environmental conditions.