Application of machine learning methods for paleoclimatic reconstructions from leaf traits. (18th November 2020)

• Record Type: Journal Article
• Authors: Wei, Gang; Peng, Changhui; Zhu, Qiuan; Zhou, Xiaolu; Yang, Bin
• Is Part Of: International journal of climatology. Volume 41(2021)Supplement 1
• Page Start: E3249
• Page End: E3262
• Publication Date: 2020-11-18
• Subjects: leaf physiognomy -- machine learning -- paleobotany -- paleoclimate -- precipitation -- temperature; Climatology -- Periodicals; Climat -- Périodiques; Climatologie -- Périodiques; 551.605
• DOI: 10.1002/joc.6921
• Languages: English
• ISSNs: 0899-8418

Abstract

Abstract: Digital leaf physiognomy (DLP) is considered as one of the most promising methods for estimating past climate. However, current models built using the DLP data set still lack precision, especially for mean annual precipitation (MAP). To improve predictive power, we developed five machine learning (ML) models for mean annual temperature (MAT) and MAP respectively, and then tested the precision of these models and some of their averaging compared with that obtained from other models. The precision of all models was assessed using a repeated stratified 10‐fold cross‐validation. For MAT, three combinations of models ( R 2 = .77) presented moderate improvements in precision over the multiple linear regression (MLR) model ( R 2 = .68). For loge (MAP), the averaging of the support vector machine (SVM) and boosting models improved the R 2 from .19 to .63 compared with that of the MLR model. For MAP, the R 2 of this model combination was 0.49, which was much better than that of the artificial neural network (ANN) model ( R 2 = .21). Even the bagging model, which had the lowest R 2 (.37) for loge (MAP), demonstrated better precision ( R 2 = .27) for MAP. Our palaeoclimate estimates for nine fossil floras were also more accurate, because they were in better agreement with independent paleoclimate evidence. Our study confirms that our ML models and their averaging can improve paleoclimatic reconstructions, providing a better understanding of the relationship between climate and leaf physiognomy. Abstract : The relationship between leaf physiognomy and climate is highly complex and nonlinear. Current models, such as linear regression models and artificial neural network models, lose precision when estimating some signals of paleoclimate. Our study confirms that tree‐based models and support vector machine models could improve the prediction accuracy. For MAT, the simple averaging of RF, SVM and GBRT models have strong power to improve predictive accuracy and can help to better understand the relationship between climate and leaf physiognomy.