Quantitative pollen-based reconstructions of climate characteristics and forest coverage for Northern Central Siberia: evaluation of different techniques

Abstract

The paper presents the application of four quantitative methods for pollen-based paleoclimatic reconstruction (the Best Modern Analogue technique (MAT), weighted averaging (WA) and weighted averaging partial least squares (WAPLS) transfer functions and Random Forest algorithm (RF)) to derive temperatures, precipitation and total woody coverage (in 5, 10, 20 and 50 km radii) for the poorly studied Central Siberian region, and verifies their accuracy. The modern pollen dataset of 243 surface samples from tundra, forest-tundra and taiga covers most of Central Siberia (north of 60°N). The results of the cross-validation of the models revealed that all methods can reconstruct the mean July temperature (R² = 0.64–0.91, RMSEP = 1.15–1.71 °C) and the mean annual temperature (R² = 0.49–0.91, RMSEP = 1.04–1.8 °C) most accurately, with the perspective methods being WAPLS and RF. Reconstructions of mean January temperature (R² = 0.43–0.87, RMSEP = 2.45–4.47 °C), mean annual precipitation (R² = 0.25–0.89, RMSEP = 54–95 mm/yr) and total woody coverage (R² = 0,51-0,82, RMSEP = 18–30 %) are much less reliable. The application of four methods for climatic reconstructions using fossil pollen assemblages from the fast-growing peatland located in 20 km SE of the Tura settlement (Evenki Region, Krasnoyarsk Krai, Russia) showed that only the RF reconstructions for T_ann and woody coverage for 5, 20 and 50 km radii are statistically significant with p ≤ 0.05. The comparison of the obtained results with the instrumental observations at the Tura weather station (1929–2019) exhibited similar trends and close absolute values in 1929–1960 and 1970–1987, and opposite trends in 1960–1970 and after 1987.