Estimation Of Ecosystem Respiration In The Grasslands Of Northern China Using Deep Learning

Accounting for about 40%of the Earth's land surface and 20%of the global carbon stocks in both vegetation and soil,grassland ecosystems play important roles in the terrestrial carbon cycle.Northern China's grasslands are mainly distributed in the Tibetan Plateau?TP?and Inner Mongolian Plateau?IM?,occupy more than 70%of the grasslands in China,and represent two significant grassland types worldwide?i.e.,alpine and temperate grasslands?.Moreover,they are sensitive to climate changes due to their unique plateau topography,the extreme cold,arid and semi-arid environment and the high soil carbon density.Ecosystem respiration?RE?is a major flux in the global carbon cycle.Small changes in RE can have a significant impact on the atmospheric CO2 concentration and thus be a potentially positive feedback mechanism to the warming climate.However,the respiration process is so far poorly understood due to the complex interactions among chemical,physical,and biological processes.Numerous studies at regional and global scales indicate that RE estimates remain rather uncertain.Therefore,accurately assessing the spatial and temporal dynamics of RE and its influencing factors in the grasslands of northern China is not only essential to the understanding of terrestrial carbon budget and its response to future global changes,but also important to the improvement of China's terrestrial carbon sink capacity and livestock production.A variety of environmental factors influencing RE,including temperature,moisture,plant productivity and soil organic carbon?SOC?.Among them,SOC storage not only represents the quantity of soil carbon substrate of soil respiration,which is a major component of RE,but also regulates the reference respiration rate.However,the three major approaches for regional RE estimation,process models,semi-empirical models and empirical models are all lack of considering the influence of SOC,which may lead to some uncertainty.Empirical models based on machine learning?ML?algorithms are driven by observational data without complex assumptions and a large number of parameters,thus have the potential to incorporate the influence of SOC.However,ML models are affected by uncertainty factors such as the representativeness of flux observation network,environmental variables and model structure.Meanwhile,as a new branch of ML models,deep learning?DL?models have been successfully applied to many fields in Earth system science,which need to predict in complex system.However,few attempts have been made to apply DL models in the estimation of regional RE,the suitability of DL models in quantifying RE at the regional scale remains unclear.This study is based on the integration of flux,meteorological,remote sensing and soil map data.Firstly,we assessed the spatial representativeness of the grassland ecosystem flux observation network in northern China by taking RE as the observation target and calculating the environmental similarity distance.Then we quantified the contribution of each flux site,and analyzed the effect of different grassland classification level to the spatial representativeness evaluation.Secondly,we developed four ML models for estimating RE in northern China's grasslands.The four ML models include three traditional ML models respectively named the back propagation artificial neural network?BP–ANN?model,the support vector regression?SVR?model and the random forests?RF?model,and a DL model named the stacked autoencoders?SAE?model.Then we compared the performance of the four ML models in estimating RE,evaluated the effects of different training strategies on the model performance,and analyzed the effects of different environmental variables on the RE estimation.Finally,based on the deep learning model?SAE?,we developed a regional RE estimation approach with incorporating the influence of SOC,examined the magnitude and spatial patterns of RE in the grasslands of northern China during 2001-2015,and analyzed the spatial control of environmental factors on RE in northern China's grasslands.The main results are as follows:?1?When RE was taken as the observation target,the grassland ecosystem flux observation network represented the environmental conditions of the northern China's grassland ecosystem very well.Under the two grassland classification levels,the well-represented areas covered 84.37%and 62.57%of the total area,respectively,while the under-represented areas were mainly distributed in the western of TP.The flux observation network well represented most areas of each grassland type,but the representativeness also varied with the grassland type.The contribution of different flux sites to the spatial representativeness differed largely,and the representative spatial extend of flux sites in each grassland type was consistent with the spatial distribution of each grassland type while with stronger spatial heterogeneity.There were significant differences in the results of spatial representativeness evaluation under different grassland classification levels.In the second grassland classification level,the spatial representativeness of flux observation network in the southwest of TP was significantly weakened.?2?All four ML models estimated RE in northern China's grasslands fairly well,while the SAE model performed best(R²=0.858,RMSE=0.472 g C m^-2 d^-1,MAE=0.304 g C m^-2 d^-1).In the three traditional ML models,the SVR model performed best,followed by the RF model and the BP-ANN model.Models trained in the two strategies had almost identical performances,while performing better in the alpine grasslands(R²>0.88,RMSE<0.43 g C m^-2 d^-1,MAE<0.30 g C m^-2 d^-1)than in the temperate grasslands(R²<0.72,RMSE>0.64 g C m^-2 d^-1,MAE>0.40 g C m^-2 d^-1).In addition,the four models performed better when using the enhanced vegetation index?EVI?than using the normalized difference vegetation index?NDVI?,and performed best when using both the NDVI and EVI.The EVI and soil organic carbon density?SOCD?were the two most important environmental variables for estimating RE in the grasslands of northern China.Air temperature was more important than the growing season land surface water index(LSWI_GS)in the Tibetan alpine grasslands,while the LSWI_GS was more important than air temperature in the Inner Mongolian temperate grasslands.?3?The deep learning approach proposed in this study accurately simulated the spatial and temporal variation of RE(R²=0.87,RMSE=0.45 g C m^-2 d^-1,MAE=0.29g C m^-2 d^-1)and SOCD(R²=0.61,RMSE=1.75 kg C m^-2,MAE=1.28 kg C m^-2)in the grasslands of northern China,while the uncertainty was also reduced in the regional RE estimation by considering the influence of SOC.RE in the grasslands of northern China presented a heterogeneous geographical pattern.The mean annual RE of northern China's grasslands during 2001-2015 was 436.71±8.14 g C m^-2 yr^-1.There were significant differences in the spatial pattern of RE between the alpine and temperate grasslands:the RE exhibited a clear decreasing gradient from the southeast to the northwest in the alpine grasslands,while exhibited a clear decreasing gradient from the northeast to the southwest in the temperate grasslands.The primary and the secondary regulators of the spatial pattern of RE were the mean annual growing season EVI and SOCD,which represented the plant productivity and SOC,respectively.The main climatic regulator in the alpine and temperate grasslands was temperature?mean annual air temperature?and moisture?mean annual growing season LSWI?,respectively.For each environmental factor,the intensity of influence on the spatial pattern of RE largely varied with the area in the grasslands of northern China.In conclusion,we proposed a deep learning approach for RE estimation with considering the influence of SOC in this study.This approach can accurately simulate the spatial and temporal variation of RE by extracting high-level features in environmental variables,and reduce the uncertainty in regional RE estimation by considering the influence of SOC and its cooperative restriction with moisture to the carbon substrate supply in the grasslands of northern China.At the regional scale,by fully considering the influence of climatic,vegetation and soil factors,we found that the spatial pattern of RE in the grasslands of northern China was mainly regulated by plant productivity and SOC,while the main climatic regulator in the alpine and temperate grasslands was temperature and moisture,respectively.For each environmental factor,the intensity of influence on the spatial pattern of RE largely varied with the area in the grasslands of northern China,which was mainly caused by the complex interactions between different environmental factors and their cooperative effect on respiration process.This study improved the accurate assessment of regional RE dynamics and their response to environmental change in a grassland ecosystem,which may further promote the better understanding of terrestrial carbon cycle.