Spatial variability of clay minerals in a semi-arid region of Turkiye

• Published in: Geoderma Regional Vol. 38; p. e00820
• Main Authors: Günal, Hikmet, Acir, Nurullah
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.09.2024
• Subjects: algorithms; Artificial intelligence; calcium; cation exchange capacity; clay; Clay mineral; Covariate; ecosystems; fertilizer application; illite; kaolinite; magnesium; mineral content; Modeling; Prediction; regression analysis; semiarid zones; soil formation; Spatial distribution; Spatial distribution; Covariate; Modeling; Artificial intelligence; Clay mineral; Prediction
• ISSN: 2352-0094; 2352-0094
• DOI: 10.1016/j.geodrs.2024.e00820

Clay minerals are the source of many chemical and physical properties that influence the provision of soil-based ecosystem services. This study aimed to identify the most significant soil characteristics contributing to the spatial variability of clay minerals in a semi-arid region of Turkiye. Additionally, the study assessed the predictive capabilities of Classification and Regression Tree (CART), Random Forest Regression (RF) and eXtreme Gradient Boosting Regression (XGBoost) in estimating soil clay mineral content. Smectite+vermiculite (SMVR) was the most abundant clay mineral in the study area, followed by illite and kaolinite. Hyperparameter tuning significantly improved model accuracy, with root mean square error (RMSE) reductions ranging from 2.53% to 97.3%. The machine learning algorithms demonstrated varying performances in spatial prediction accuracy. The RF model achieved the lowest RMSE (8.587%) and the highest R2 values (0.796) for predicting SMVR. The XGBoost outperformed other models for kaolinite (RMSE: 4.814%, R2:0.713) and illite (RMSE:7.368%, R2:0.613). Exchangeable cations, particularly magnesium (Mg) and calcium (Ca), were identified as crucial factors influencing the spatial distribution of clay minerals. Among these, M concentration had the strongest influence on predicting both SMVR (38.1%) and illite (26.3%). Conversely, for kaolinite prediction, Ca concentration played the most significant role (38.7%), followed by Mg (19.93%). In conclusion, this study demonstrates the effectiveness of machine learning models, particularly XGBoost which achieved the lowest RMSE for all clay minerals investigated. These models offer a valuable tool for predicting clay mineral content in the Kazova Plain. The findings highlight the importance of parent material, weathering processes, and specific soil properties, such as exchangeable cations, in shaping clay mineral distribution. This knowledge not only contributes to a deeper understanding of soil formation in semi-arid environments but also practical applications. For instance, by predicting the abundance of SVMR, known for its high cation exchange capacity land managers can develop targeted strategies for optimizing fertilizer application in the Kazova Plain.