关键词: 时间卷积神经网络, NEE, 长白山红松阔叶林

Abstract:

Objective: Analyze the seasonal differences of net ecosystem exchange (NEE) of carbon in the study area(the forest of Korean pine and broad leaved trees in Changbai Mountain) and the response of the NEE to meteorological factors, reveal the dynamic impacts of meteorological factors on the NEE at a monthly basis, and provide a theoretical guidance for regulating the carbon budget in the study area. Meanwhile, the application of temporal convolution network (TCN) in the simulation of the NEE was studied to explore a new method of carbon flux simulation. Method: Based on 3 min data obtained at the flux observation station in temperate forest of Korean pine and broad-leaved trees in Changbai Mountain from 2007 to 2010, the seasonal differences of the NEE and five meteorological factors were analyzed, and correlation between the meteorological factors and the NEE was analyzed. The random forest model was used to calculate the importance scores of factors affecting NEE, five meteorological factors with higher scores were selected as inputs to the NEE simulation: latent heat flux, sensible heat flux, air humidity above the canopy, water vapor pressure above the canopy and net radiation; five NEE models were developed respectively based on TCN, long short term memory (LSTM), artificial neural network (ANN), support vector regression (SVR) and extreme learning machine (ELM). The prediction precision and stability of different models were examined by calculating coefficient of determination (R²), mean absolute error (MAE) and root mean square error (RMSE). Result: The annual total NEE was -74.777 3 gCO₂·m^-2a^-1 revealing an overall carbon sink for the whole year, but seasonally with a carbon sink in summer and a carbon source in winter. The NEE was extremely significantly (P < 0.000 1) negatively correlated with latent heat flux, vapor pressure over canopy, net radiation and air humidity over canopy, but not significantly correlated with sensible heat flux (0.077 0). The RMSE and the R² of TCN were respectively 0.110 5 mgCO₂·m^-2s^-1 and 0.821 4, the RMSE was decreased by 0.024 8, 0.022 4, 0.022 2, 0.006 8 mgCO₂·m^-2s^-1 and the R² was increased by 0.080 6, 0.077 7, 0.068 6, 0.022 3 compared with ELM, SVR, ANN and LSTM, respectively. According to 10 tests of the above 5 models, the standard deviation of RMSE of TCN obtained by calculation was 0.000 4 mgCO₂·m^-2s^-1, which was 0.001 4, 0.001 3, 0.000 2 mgCO₂·m^-2s^-1 lower than that of ELM, ANN and LSTM, respectively. Conclusion: The NEE of the flux observation station was a carbon sink overall, with obvious seasonal differences. The NEE was extremely significantly (P < 0.000 1) negatively correlated with latent heat flux, vapor pressure above the canopy, air humidity above the canopy and net radiation, but not significant (P=0.141 8) with the sensible heat flux. The long-term NEE prediction showed good precision and strong stability by the simulation model based on the TCN, providing a feasible basis for the application of the TCN in the field of ecological simulation. The study provides a theoretical guidance for regulating the carbon budget of the forest of Korean pine and broad-leaved trees in Changbai Mountain.

Key words: temporal convolution Network(TCN), net ecosystem exchange(NEE), forest of Korean pine and broad leaved trees in Changbai Mountain