Soil Respiration Variability: Contributions of Space and Time Estimated Using the Random Forest Algorithm
I. A. Smorkalov
Abstract
Soil respiration modeling (i.e., simulation of carbon dioxide emissions from the soil surface) makes it possible to analyze and forecast changes in the carbon cycle in terrestrial ecosystems. Along with classical regression models, researchers currently use machine learning methods based on neural networks or regression tree ensembles. However, models produced using these methods are often used as ‘black boxes’, which hinders the analysis of process mechanisms. This paper demonstrates capabilities of the Random Forest algorithm that can be successfully used to estimate effects exercised by various factors on soil respiration based on features importance measurements. Using variance separation, predictors have been classified either as spatial (biotope type, soil type, vegetation type, and soil moisture content) or temporal (soil and air temperature, NDVI, LAI, FPAR, and SPEI). Several models were produced based on 5670 respiration measurements performed during five growing seasons (2012–2016) on 30 sampling plots in south-taiga pine forests and meadows that feature different vegetation and soil types but are confined to the same small area. Different models include different sets of predictors (all predictors, temporal predictors only, spatial predictors only, and temperature and humidity only), and their accuracy reaches R2 = 0.88 (MSE = 0.47). It is established that soil respiration depends primarily on temporal factors whose importance ranges from 76 to 91%. In forests, the effect of spatial factors on respiration is stronger than in meadows.