Prediction of the fate of potentially toxic elements using the Random Forest model
Conference Object Open access M34
2026
Radojević, Ana 
Bogdanović, Anđelija
Kalinović, Tanja
Kalinović, Jelena
Nujkić, Maja
Jordanović, Jelena
Serbian Chemical Society
Book of abstracts - 10th Symposium Chemistry and Environmental Protection, EnviroChem2026, with International Participation, Bajina Bašta, 21-24th June 2026
45
English
M34 - Conference paper at an international conference, printed as an abstract
A Random Forest (RF) model, rapid field assessment tool in machine learning, uses a collection of decision trees to make predictions. RF is widely used in biomonitoring for high-dimensional, non-normally distributed datasets, to reduce overfitting, and to identify key predictors of environmental stress. The Bor area is well known for pollution resulting from miningmetallurgical production of copper. Sampling of plant and soil material was conducted before the modernization of the smelter and reflects a high, long-term environmental load. The dataset included total concentrations of potentially toxic elements (As, Cr, Cu, Mn, Ni, Pb, Zn) in soil, roots, and washed leaves/needles (µg/g dw) across three plant species (oak, hazelnut, spruce) and 12 sampling sites, along with soil properties (pHKCl, pHH2O, %SOM), distance of the sites from the smelter (m), and directions. In the dataset, 20% comprised random testing data, and the remainder was used for training. Concentrations in leaves/needles were used as response variables, while other parameters served as predictors. The model was built with 200 trees and evaluated using Leave-One-Out Cross-Validation (LOO-CV), while performance was assessed using root mean square error (RMSE), and coefficient of determination (R²). Data were standardized using StandardScaler, with scaling parameters calculated on the training set within each cross-validation fold to avoid data leakage. RF models were built using scikit-learn library in Python 3.12.3. Predictive performance was categorized as weak (R²<0.25), moderate (0.25–0.50), good (0.50–0.75), and high (>0.75). Overall, the model showed weak to moderate predictive power (R2<0.3), resulting from the small sample size of 12 sites per plant. Spruce, among analysed species, is suggested as the most reliable bioindicator, particularly for Cu and Pb (~0.32–0.34). The concentrations of As in needles were too low to observe the 12-point pattern. At the urban-industrial site, ~1.5 km in the SW downwind from the smelter, up to 96 times higher concentration of Cu and 62 times higher concentration of As were observed compared to the background site. Direction and distance of the sites contributed approximately 25–35% of predictive power, while soil parameters account for the remainder, suggesting that bioavailability plays a role, but direct spatial proximity to the smelter remains the dominant factor.
451-03-34/2026-03/200131
Cobiss
195481097
ISBN
978-86-7132-092-4