The application of machine learning for evaluating anthropogenic versus natural climate change

Abstract

Time-series profiles derived from temperature proxies such as tree rings can provide information about past climate. Signal analysis was undertaken of six such datasets, and the resulting component sine waves used as input to an artificial neural network (ANN), a form of machine learning. By optimizing spectral features of the component sine waves, such as periodicity, amplitude and phase, the original temperature profiles were approximately simulated for the late Holocene period to 1830 CE. The ANN models were then used to generate projections of temperatures through the 20th century. The largest deviation between the ANN projections and measured temperatures for six geographically distinct regions was approximately 0.2 °C, and from this an Equilibrium Climate Sensitivity (ECS) of approximately 0.6 °C was estimated. This is considerably less than estimates from the General Circulation Models (GCMs) used by the Intergovernmental Panel on Climate Change (IPCC), and similar to estimates from spectroscopic methods.

Introduction

The past two decades have seen an unprecedented international focus on climate change, and particularly the perceived relationships between increasing global temperatures and emissions of greenhouse gases, referred to as anthropogenic climate change. The Intergovernmental Panel on Climate Change (IPCC) is a body under the auspices of the United Nations, set up at the request of member governments, to provide scientific information on climate change and its political and economic impacts. The IPCC publishes Assessment Reports at regular intervals, with the current viewpoint, based mainly on application of physical models, particularly General Circulation Models (GCMs). These models attribute over 90% of the global warming since 1900, and virtually 100% of the global warming since 1970, to anthropogenic climatic forcings, particularly industrial emissions of carbon dioxide and methane [71], [87].

Instrumental temperature records extend back a little over a century. To understand how climate has varied over much longer periods, over hundreds and thousands of years, various types of proxy records have been assembled. These are derived from measurements associated with biological and geological phenomena that can leave evidence of past climate, particularly temperatures. The most familiar proxy records are derived from annual rings of long-lived tree species. Other proxies include measurements from corals, stalagmites, and sediments. These types of records provide evidence for periods of time over the past several thousand years (the late Holocene) that were either colder, or experienced similar temperatures, to the present, for example the Little Ice Age and the Medieval Warm Period [40], [43], [44].

Examination of many of these proxy temperature records shows they typically consist of complex oscillations or cycles about a mean value, with the amplitude and structure of the temperature signal depending on the geographical location considered. In the pre-industrial era, these oscillations represent the compound effect of natural phenomena both internal (e.g. North Atlantic Oscillation, El Nino Southern Oscillation) and external (e.g. solar, volcanic activity).

Since about the mid-nineteenth century, with the growth of industrialisation, there is the possibility that there is also a contribution to climate change from anthropogenic greenhouse gases, particularly carbon dioxide and methane. However, the relative contributions of natural cycles and anthropogenic effects is far from certain [94] and there is continuing interest in attempting to answer this question of relative contributions [10], [35], [64], [67].

There is an extensive literature examining the occurrence of periodic cycles within proxy temperature reconstructions, through application of spectral analysis [32], [78], [88]. Many of these studies also discuss possible relationships between these cyclic patterns in temperature profiles and natural phenomena that may affect causation, particularly those associated with solar cycles [14], [27], [40], [56], [74], [81], [82], [97]. For example, in the southern hemisphere, Nordemann et al. [74] undertook spectral analysis using tree ring data from Brazil and Chile, providing evidence for associations with solar cycles, particularly the Suess (∼200 year), Gleissberg (∼90 years), Hale (∼ 22 years) and Schwabe (11 years) cycles. Rigozoa et al. [79] examined tree ring widths in Chile, and found an association with solar activity with 11 and 80 year periodicities.

In the northern hemisphere, Raspopov et al. [78] performed spectral analysis of long-term dendrochronological data from Central Asia and demonstrated an approximate 200-year climatic periodicity, showing a high correlation with solar periodicity for the de Vries period (∼210 years). Ogurtsov et al. [75] reported spectral analysis of tree periodicity and discussed the association with the modulation of regional climate in Northern Fennoscandia by the Gleissberg solar cycle (∼90 years).

Moffa-Sánchez et al. [68] examined marine sediments for isotopic signals in the shells of the planktonic foraminifera over the past 1000 years. Spectral analysis showed a 200-year periodicity, identified with de Vries solar cycle (∼210 years). Galloway et al. [32] generated a late Holocene temperature record based on diatoms from a sediment obtained from British Columbia, Canada. Spectral analysis shows significant periodicities at 42–60, 70–89, 241–243, and 380 years, and inferred relationships to sunspot number variation. Tan and Liu [89] produced a 2650-year temperature reconstruction from annual layers of a stalagmite from China, with spectral analysis indicating significant periodicities at 206 and 325 years.

Cyclic variations have also been associated with large-scale internal climate oscillatory modes [15], that may themselves in turn be influenced by solar activity [49], [65], [91], [93], [100], [102]. For example, Wilson et al. [97] examined tree ring widths to enable a reconstruction over 1300 years for the Gulf of Alaska: identifying oscillatory modes at 90, 38, 24, 50.4 and 18.7 years related to changes in sea surface temperature pattern. In addition to the decadal and centennial cyclic periodicities referred to above, there is evidence of cycles on millennial time scales, for example the Bond and Dansgaard-Oeschger (DO) cycles [70], [101].

These studies indicate that temperatures have oscillated at the regional and global scale with the resultant signals able to be decomposed into component parts. If a set of components can be identified, and given the very large datasets, at least in theory, the oscillations could be used to forecast future climate using machine learning techniques.

Artificial neural networks (ANNs) are a form of machine learning. The ANN technique has been widely applied to simulation and forecasting of climatic and meteorological variables including temperatures [18], [26], [76], [80], rainfall [3], [4], [5], [6], [24], [72], [99] solar radiation [8], [20], and wind speed [19], [29], [57]. In this study, proxy temperatures from the Canadian Rockies, Switzerland, Tasmania (Australia), New Zealand, southern South America and a composite representing the Northern Hemisphere, were decomposed into sets of sine waves. There are many such studies in the literature. However, what is unique with our study is that we have then used the resultant components to train ANNs and make projections of future temperature. The divergences between these projections based on the natural cycles, and actual temperature measurements from the twentieth century was used as an indication of the extent of anthropogenic influences contributing to global warming.