Machine learning-based soil aggregation assessment under four scenarios in northwestern Iran

Abstract

. Soil aggregate stability is crucial for maintaining the arrangement of solid particles and pore space in the soil, even under mechanical stresses. Traditional direct measurements of soil aggregate stability are time-consuming and expensive. This study aimed to spatially predict the soil aggregate stability indices, including the mean weight diameter of aggregates, the geometric mean diameter of aggregates, and the percentage of water stable aggregates, using five machine learning models and environmental covariates in the framework of digital soil mapping. A total of 100 samples were collected from the surface layer (0-15 cm) of soils in the Aji-Chai watershed, northwestern Iran, and their SAS indices were determined by standard laboratory methods. Four scenarios (S) were employed to evaluate the most influencing auxiliary variables, including (S 1 ): topographic attributes, (S 2 ): topographic attributes + remote sensing data, (S 3 ): S 2 + thematic maps (geology, land use/cover maps), and (S 4 ): S 3 + selected soil properties. Among the various machine learning models, the random forest showed exceptional performance and reduced uncertainty for S 4 , compared to the other machine learning models and desired scenarios. The coefficient of deter - mination, concordance correlation coefficient, and normalized root mean squared error values of the random forest model were 0.86, 0.87, and 31.42% for mean weight diameter; 0.80, 0.84, and 31.59% for geometric mean diameter; and 0.54, 0.68, and 20.75% for water stable aggregates, respectively. Additionally, properties such as soil organic matter and clay, followed by remote sensing data, demonstrated the highest relative importance when compared to the other covariates in predicting the soil aggregate stability indices. In conclusion, the random forest ML-based model seems to be able to accurately predict soil aggregate stability indices at the watershed scale. The generated maps can serve as a valuable baseline for land use planning and decision-making. These findings contribute to the scientific understanding of soil physical quality indicators and their application in sustainable land management practices.