Estimating carbon storage through machine learning algorithms

Abstract

Forest ecosystems have an important place in carbon conversion by transforming the CO2 they receive from the atmosphere and storing it in large quantities on the earth. Standard models are established and carbon calculations are made for biomass estimations of forest trees. While regression equations are frequently used in the prediction of biomass, estimations made with machine learning algorithms using stand parameters are rarely tested. In this study, it is evaluated whether the parameters of the stand type can be used without using the standard models or equations in the biomass estimation procedure. Verification of biomass estimates via kNN (Kernel Nearest Neighbor), RF (Random Forest) and RPART (Recursive Partitioning and Regression Trees) from machine learning algorithms for coniferous, broad-leaved and mixed stands in the Amasra, Arıt and Kurucaşile Sub-district Directorates of Bartın Forestry Directorate have been carried out. Amasra, Arıt and Kurucaşile regions, where the work was carried out, have 121, 79 and 121 stand types respectively. Carbon calculations for five diametric classes were carried out using the data for the tables of stand identification. Total carbon stocks were found to be 111 tons/ha, 115 tons/ha and 179 tons/ha for Amasra, Arıt and Kurucaşile regions respectively. Carbon stock values calculated by regression equations; it can be estimated as 40%, 85%, 99% with the KNN algorithm, 42%, 57%, 85% with the RPART algorithm and 71%, 78%, 80% with the RF algorithm in the mixed, coniferous and broad-leaved stands, respectively.