Research On The Method Of Vegetation Productivity Estimation Based On Data Driven

Vegetation productivity is a significant indicator of terrestrial system carbon cycle and is one of the most important parameters of global climate change,such as gross primary productivity?GPP?,net ecosystem production?NEP?.Although there are many methods to predict the value of vegetation productivity now,the uncertainty of estimation between different models still exists.A global site-scale vegetation productivity estimation method based on machine learning was established in this study,combing the remote sensing meteorological data based on Google Earth Engine with the data products of 212 global flux sites based on the eddy covariance technology,the new method significantly improves the inversion accuracy of the vegetation productivity estimation.The paper used long-time series of remote sensing product and the random forest algorithm.Through compared with the corresponding flux site data product,the result of research verified the advantages of machine learning method in the inversion of terrestrial ecosystem productivity,and it provided a reference and direction for future research on large-scale carbon cycle.The main conclusions are as follows:?1?Enhanced vegetation index?EVI?is the most important input parameter in the model of GPP prediction.The analysis of contribution in predicting the productivity of different plant functional types analysis all show that the EVI is the most important driver among input variables.?2?Using the data product from 16 global deciduous broad-leaved forest flux sites EVI,land surface temperature,short wave radiation and precipitation data,a random forest model was established.The performance of the model was evaluated by predicting the GPP values of the other 8 deciduous broadleaf forest sites.The results show that,compared with the MOD17A2H product,this prediction method improves modeling of GPP,with an R² of 0.82 and an RMSE of 1.93 g C m^-2 d^-1.?3?Using the data product from 212 sites in global,a long-time series data sets of training and test of all flux sites were constructed.Compared with the MODIS-GPP product based on traditional process model,the results show that this GPP prediction method based on big data and machine learning had a great improvement in GPP estimation.In the prediction results of unclassified sites,the best results were obtained with an R² of 0.67 and RMSE of 2.28 g C m^-2 d^-1;and in the prediction results of classified sites,the best results were obtained at the deciduous broad-leaved forest sites with an R² of 0.81 and RMSE of 2.02 g C m^-2 d^-1.?4?The application of random forest and data product from 212 global sites were extended to the study of simulating net ecosystem exchange,ecosystem respiration and NEP.Among the classified global sites,the results are better in deciduous broadleaf forest and cropland ecosystem than other sites.