Abstract:

Objective: Based on continuous monitoring data of soil respiration and environmental factors in the fixed sample plots of stratified mixed stand of Eucalyptus spp. and Manglietia glauca, over a period of one year, a multi-factor prediction model for plantation soil respiration was constructed and screened, aiming to clarify key environmental drivers influencing the spatiotemporal variation of soil respiration in the plantations of the region, and provide a scientific basis for improving the accuracy of carbon emission simulations in plantations and calibrating large-scale predictive models. Method: The stratified mixed stand of E. spp. and M. glauca in Leizhou Peninsula was taken as the study object. Six machine learning algorithms (Random Forest, Time Convolutional Neural Network, Long and Short-Term Memory Network, Support Vector Regression, Extreme Learning Machine, and BP Neural Network) and two empirical models (Q₁₀ and Gamma) were introduced to simulate soil respiration changes at 1-hour and 24-hour scales. The accuracy evaluation metrics of the models were compared to select the optimal model algorithm suitable for the study area. Result: The soil respiration in the mixed forests was higher in the rainy season than in the dry season. Cumulative fluxes of soil respiration were 616.83 g·m^?2 in rainy season, 319.81 g·m^?2 in dry season, and 936.64 g·m^?2 throughout whole year. Six machine learning algorithms and two empirical models were able to successfully simulate soil respiration changes in the mixed forests, but the machine learning models outperformed empirical models. The Random Forest model achieved the highest consistency, with R² values of 0.89 (training set) and 0.76 (test set) using soil temperature and moisture as inputs. When the input variables increased soil electrical conductivity, soil heat flux, air temperature, air relative humidity, total solar radiation, and photosynthetically active radiation, R² values increased to 0.99 (training set) and 0.93 (test set). In addition to soil temperature and humidity, soil electrical conductivity significantly affected soil respiration. Conclusion: The soil respiration in a stratified mixed stand of E. spp. and M. glauca exhibits distinct temporal variation, machine learning algorithms are more advantageous than traditional empirical models in predicting soil respiration changes, among which the random forest model performs best. The predictive ability of the random forest model can be greatly improved by adding input variables such as soil electrical conductivity, and the addition of these factors can be a better prediction of soil respiration changes, which can provide a reliable basis for assessing the carbon sequestration status of plantations.

Key words: soil respiration, prediction model, random forest, Eucalyptus, stratified mixed stand