Assessing environmental drivers and predictive modeling of soil organic carbon using random forest in the Urmia Lake Basin

Details

Supervisors

Bogaert, Patrick

Faculty

Faculté des bioingénieurs

Degree label

Master [120] : bioingénieur en sciences agronomiques, à finalité spécialisée

Abstract

This master's thesis investigates the spatial distribution and drivers of soil organic carbon (SOC) in the semi-arid Urmia Lake Basin in northwestern Iran, a region facing significant environmental challenges due to climate stress and land degradation. The primary aims are to predict SOC at field-sampled locations across four study areas, and to identify the environmental variables and data sources that most significantly affect SOC variability. To achieve these objectives, a Random Forest machine learning framework with external validation was implemented, integrating a wide variety of covariates, including terrain derivatives, climatic variables, remote sensing indices, and, where available, soil properties and land use map. Data were collected from 2014 to 2022 over natural and managed zones, and models were trained and assessed for predictive accuracy and variable importance. Results revealed consistently low SOC concentrations (<1%) and high spatial variability across the region. Topographic variables (e.g., elevation, terrain indices) emerged as the most important predictors of SOC, accounting for about 50% of model importance in most cases, with precipitation as a key climatic driver, especially in natural (non-agricultural) areas. Remote sensing indices provided moderate improvements, largely related to soil texture, while soil properties (phosphorus, potassium, silt, saturation percentage) substantially improved model performance when available. In agricultural and managed zones, prediction was considerably more difficult, as land management practices might disrupt natural SOC-environment relationships. Land use information marginally improved predictive accuracy. The findings highlight the critical influence of local topography and rainfall on SOC distribution in semi-arid environments, the value of including measured soil properties for model improvement, and the need for tailored modelling strategies that account for land use and management practices. Recommendations for future work include testing additional variables, using land use-stratified models, integrating spatial and temporal covariates, and capturing management information to enhance SOC prediction accuracy in complex agroecosystems.