Abstract:Aiming to apply deep learning theory and technology to model the net ecosystem exchange (NEE) of grassland ecosystems in the alpine region, based on the FLUXNET CO2 flux data of the grassland in Duolun County, Xilinguole League, Inner Mongolia Autonomous Region, from 2007 to 2008, the attentionbased encoderdecoder framework in deep learning was used to simulate NEE, and the random forest model was used to calculate the importance score of the relationship between NEE, which included the photosynthetic photon flux density (PPFD), soil temperature (Ts), air temperature (Ta), precipitation (P) 〖JP3〗and soil moisture content (SWC) and water pressure difference (VPD), and their seasonal differences in the relationship were analyzed. The result showed that the root mean square error (RMSE) of the deep learning model was 0.28μmol/(m2·s), which was declined by 0.14μmol/(m2·s) and 0.08μmol/(m2·s), respectively, compared with ANN and SVM. The coefficient of determination was 0.93, which was increased by 0.29 and 0.34, respectively. With the attention mechanism, the RMSE standard deviation predicted by 10 times training was 0.002μmol/(m2·s), which showed a reduction of 0.005μmol/(m2·s) and 0.036μmol/(m2·s) compared with Encoder-decoder network model and long short-term memory (LSTM). The attention mechanism model was more competitive in predicting stability. The importance score calculated by random forest model showed the variations of photon flux density PPFD (335) and saturated vapor pressure VPD (30.0) played a leading role in the variations of NEE from March to April. During the period from May to June after the growing season, soil water content SWC (50.5) was the main influencing factor of NEE variations. The precipitation P (3.8) showed a process of decrease in July, and the photon flux density PPFD (26.8) and the soil water content SWC (60.1) were the collaborative decision of NEE. In August, PPFD (2.8) and SWC (6.9) were relatively abundant with plentiful rainfall. The saturated vapor pressure difference VPD (41.5) and rainfall P (42.7) became the main factors affecting NEE. After September, the photon flux density PPFD and rainfall were both decreased sharply and remained stable. The temperature coefficient Q10 was increased slightly compared with the growing season and reached a maximum value of 5.96 in January, so temperature was the decisive factor affecting NEE through plant respiration in non growing season, which is 44.6 between January to March and 542 between October and December. In conclusion, radiation, temperature and moisture remarkably affected NEE in the growing season of grassland ecosystems in alpine region, temperature was the main factor in nongrowing season. Comparing with traditional machine learning algorithms such as support vector machines, deep learning theory and technology had better application prospects in the field of ecological model simulation.
