Predicting the spatial distribution of stable isotopes in precipitation using a machine learning approach: a comparative assessment of random forest variants

Abstract

Stable isotopes of hydrogen and oxygen are important natural tracers with a wide variety of environmental applications (e.g., the exploration of the water cycle, ecology and food authenticity). The spatially explicit predictions of their variations are obtained through various interpolation techniques. In the present work, a classical random forest (RF) and two of its variants were applied. RF and a random forest version employing buffer distance (RFsp) were applied to each month separately, while a random forest model was trained using all data employing month and year as categorical variables (RFtg). Their performance in predicting the spatial variability of precipitation stable oxygen isotope values for 2008–2017 across Europe was compared. In addition, a comparison was made with a publicly available alternative machine learning model which employs extreme gradient boosting. Input data was retrieved from the Global Network of Isotopes in Precipitation (GNIP; no. of stations: 144) and other national datasets (no. of stations: 127). Comparisons were made on the basis of absolute differences, median, mean absolute error and Lin’s concordance correlation coefficient. All variants were capable of reproducing the overall trends and seasonal patterns over time of precipitation stable isotope variability measured at each chosen validation site across Europe. The most important predictors were latitude in the case of the RF, and meteorological variables (vapor pressure, saturation vapor pressure, and temperature) in the case of the RFsp and RFtg models. Diurnal temperature range had the weakest predictive power in every case. In conclusion, it may be stated that with the merged dataset, combining GNIP and other national datasets, RFsp yielded the smallest mean absolute error 1.345‰) and highest Lin’s concordance correlation coefficient (0.987), while with extreme gradient boosting (based on only the GNIP data) the mean absolute error was 1.354‰, and Lin’s concordance correlation coefficient was 0.984, although it produced the lowers overall median value (1.113‰), while RFsp produced 1.124‰. The most striking systematic bias was observed in the summer season in the northern validation stations; this, however, diminished from 2014 onward, the point after which stations beyond 55° N are available in the training set.