Applications Of Machine Learning Methods In Modeling Carbon And Water Fluxes Of Terrestrial Ecosystems

Investigating the carbon and water cycles in terrestrial ecosystems has been a key issue in the study of global change during recent decades.These studies contribute to predicting future climate change,assessing the response and feedback of terrestrial ecosystems to global change and providing scientific basis for addressing the effective management and control of terrestrial ecosystems.At present,both observation and simulation are considered the two main approaches for estimating the pattern and variability of global carbon and water cycles and understanding their key processes and control mechanisms.With the recent availability of large amounts of data accumulated from global flux tower network with the eddy covariance technique,how to reasonably interpret these data and extract the useful knowledge as well as develop effective methods for simulating the carbon and water fluxes has become a scientific issue which urgently needs to be solved.Therefore,based on the measurements from global flux network,this study used machine learning techniques to model the carbon and water cycles in terrestrial ecosystems,and also respectively utilized the intelligent optimization algorithms and data pre-processing technique to improve these applied models.The main findings of this dissertation can be concluded as follows:(1)The newly presented machine learning methods,including generalized regression neural network(GRNN),extreme learning machine(ELM),adaptive neuro-fuzzy inference system(ANFIS)and group method of data handling,had great ability of dealing with the nonlinear problems in modeling and predicting the carbon and water cycles in terrestrial ecosystems.These models provided satisfactory estimates comparable to traditional artificial neural network(ANN)and support vector machine(SVM).Regarding the computational time required for learning these models,ELM has great superiority over other models,while SVM requires too much time to find optimized parameters.According to the comparison among different types of ecosystems,these models generated the highest accuracy at evergreen needleleaved forests and deciduous broad-leaf forests,whereas offered the lowest accuracy at croplands.(2)Four different machine learning methods(ANN,ELM,ANFIS and SVM)were compared in terms of their potential and effectiveness to simulate and predict daily carbon and water fluxes in eight forest ecosystems across different geographic regions.In addition,to determine an optimal model for each approach,this dissertation also examined how the predictive accuracy of these approaches was affected by their respective internal parameters.The results indicated that the forecasting accuracy of each method strongly depended on their respective internal parameters.For ANN method,the generalization performance was limited by its training function.There did not exist a single universal model with the same training function that could guarantee the most accurate estimates at the eight sites.Accordingly,the best training algorithm can be acquired through the trial and error procedure.The SVM model with the radial basis function kernel algorithm considerably outperformed the SVM models with the polynomial and sigmoid kernel algorithms for the three carbon fluxes at the eight sites.For ELM approach,the sine and sigmoid transfer functions achieved similar estimates in most instances and consistently outperformed the hard limit function.In general,both the subtractive clustering and fuzzy c-means algorithms for ANFIS method were superior to the most popular grid partitioning algorithm.Considering the grid partitioning algorithm,no single membership function was best at all the sites.(3)Both genetic algorithm(GA)and particle swarm optimization(PSO)were used to optimize the machine learning methods(ANN,ANFIS and SVM).The results suggested that the two optimization algorithms were capable of improving the performance of these methods with respect to the estimation of carbon and water fluxes.For the estimation of each flux,there existed significant difference between GA and PSO algorithms for enhancing the machine learning methods among different sites.Regarding all the applied cases,there did not exist a single optimization algorithm that consistently performed better than the other algorithm.(4)Wavelet decomposition as a data pre-processing technique was employed to improve the machine learning methods in terms of estimating the carbon and water fluxes in a forest ecosystem.The results demonstrated that the modeled accuracy of ANN,SVM,GRNN,ELM,ANFIS models for estimating carbon and water fluxes was considerably improved by wavelet decomposition technique,especially for net ecosystem exchange estimates.There existed appreciable difference among a variety of wavelet basis functions in terms of improving the machine learning methods.Consequently,when integrating the wavelet decomposition technique into machine learning approaches,it is of vital importance to select the best wavelet basis function.