Insights Into The Approaches Of Extracting Climate-growth Information

Tree ring has been widely used as an alternative proxy in the quantitative paleoclimatic research.Currently,combining multiple tree ring proxies?radial growth width,stable isotopes,density,etc.?,multi-dimensional climate factors?temperature,precipitation,relative humidity,heat index,ENSO,SPEI,sc PDSI,NDVI,etc.?,multiple mathematical analysis methods?linear,nonlinear,etc.?and multi-spatiotemporal scale?year,season,month,day,global,regional,local,etc.?,were increasingly used to enhancing the climate signals of the past.Considering the same datasets,some studies have found that,different methods of analysis will produce different results.How to scientifically and efficiently analyze the variance and relationships between the target datasets,and extract climate change signal with reasonable accuracy is an important issue in the field of dendroclimatology.As a contribution to the development of research methods in dendroclimatology,the simulated data and the real datasets were taking into account.The real datasets were used including tree ring width?TRW?and stable carbon isotope data of 286 Pinus massoniana Lamb.?Masson pine?collected from six different subtropical regions in China,as well as the publicly available tree ring datasets and international tree ring data bank?ITRDB?.In order to provide methodological suport and strategy selection for extracting more accurate and wealth climate-growth information,with R-language-oriented,this paper emphasizing on the art of data mining methods was conducted.Finally,some new understandings about choosing and applying multiple tenable tools are obtained as follows.1. In tree-ring measurements,once crossdated,it is necessary to remove thebiological growth trends.However,the typically detrending approaches have their own drawbacks,such as the negative exponential curve standardisation technique that can suffer from fitting failures and remove the low and medium frequency climate-driven signal;Spline,signal-free RCS and RCS methods are faced with the problem of?segment length curse?.The complementary ensemble empirical mode decomposition?CEEMD?,which do not need to supervise the function form,decompose a raw time series into the sum of other different frequency signals.This method alleviated the problem of?low-frequency limitations?with the simulated data.It is worth repeating that the variance over time of master ring width chronologies generated by CEEMD and other detrending methods?e.g.Spline,signal-free RCS,etc?,is quite stable,with a slight increase in the mean correlation coefficient of climate-growth.Based on the large scale high solution climate grid datasets and their neighbor TRW sites,the results of point-point principal component regression analysis indicate that the range of the CEEMD in different screening probability criteria?VRSQ,VRE and VCE?are remarkably consistent across other detrending methods.In the field of the time series signal theory,this data-driven method has a good potential for application in dendroclimatology.2. The traditional quantitative correlation methods?Pearson,Kendall,Spearman?usually benifit only on linear or monotonic correlation,are susceptible to outliers and sample size,and have low robustness.In this paper,the traditional and modern correlation methods?mutual information,distance correlation,maximum information coefficient,Hoeffding's D,etc.?were compared and analyzed by using simulated data and published tree ring data.It was found that for the same data set,the correlation coefficients quantified by different correlation analysis mehtods have certain differences.Distance correlation and mutual information mthods not only have roghly equivalent performance to Pearson method in quantifying the linear climate-growth correlation,but are also suitable for nonlinear correlation analysis and can be used as an adjunct to climate-growth correlation anaalysis.3. The non-stationary tree-ring-climate time series and collinearity betweenclimate variables may cause spurious regression problems and redundant information of feature variables.In this case,dendroclimatologic regression model will produce unreliable and spurious results.Based on the non-stationary test of 67414 grid data sites?CRUts4.03?,it was found that mean temperature series in about 58%of sites show significant non-stationary variation,while other grid sites with significant stationary mean temperature time series were mainly concentrated between40°–70°N;the significant non-stationary precipitation time series were about 15%across global scale,and the grid sites with significant trends are mainly concentrated in the northern hemisphere domain.Therefore,when conducting climate-growth time series regression analysis,attention should be paid to the stationarity testing and stationarize the time series to reduce the uncertainty.Addressing the feature variable selection are demonstrated using simulated data and the TRW data of the upper Colorado River.Input variable selection was conducted by the partial mutual information?PMI?correlation algorithm with information entropy.Results indicated that PMI method can not only reduce the information redundancy problem,but also select the non-linear feature variables with better fit than stepwise regression methods.4. Based on linear?MLR,MT?and nonlinear?BRNN,RF?methods,climaticfactors with different temporal resolutions?day-wise,monthly?were used to fit the climate-growth response functions.The comparison showed that BRNN and MLR methods were superior to MT and RF methods.Compared to the monthly climate data,the maximum variance interpretation of the regression model is somewhat improved using the daily data analysis.5. In the face of global scale tree ring and climate grid data,it is necessary tosolve for distance between neighbors.But for a large data sets,how to solve them quickly and efficiently is a problem that needs to be solved.This topic introduces the Hash algorithm,the codes written in R language that not only solve the large dataset quickly,but also with high accuracy.6. The monthly,seasonal correlation and time stability characteristics of the threedifferent TRW metrics?total ring width,earlywood,latewood?and the stable carbon isotopes?whole wood,?cellulose,holocellulose?of different components from Masson pine and different climatic factors were analyzed.This study provides a reference for the selection of proxy indicators in dendroclimatelogy in the subtropic humid zone of China.The results indicated that,in the same region,the sign of correlation coefficient between the chronologies of three different TRW metrics and the same monthly climatic factor is coincident.The using of earlywood and latewood width has little benefit on extracting more climate information.In different study areas,the climatic factors that most significantly affect the TRW of Masson pine are different.The three components displayed uniform year-to-year variations and common significant climatic signals,indicating that?¹³C of whole wood from Masson Pine was enough to reconstruct inter-annual paleo-climate information without cellulose isolation for its high climate sensitivity and less time consuming.The mean value of?¹³C of the three components showed a significant negative correlation with the relative humidity in summer and autumn in all study areas.When going large scale,under the different climate environments,in the same time span,rather than the ring width,the stable carbon isotope response of Masson pine is stronger and more coincident,and the significant climate signals stored by the tree-ring width are different in a larger region.These new understandings provide a new basis and method for the scientific and efficient extraction of climate-growth information,and have important theoretical and practical significance for promoting the research of dendroclimatology.