Coefficient of linear extensibility of soil can be estimated from hygroscopic water content or clay and organic carbon contents

The determination of the coefficient of linear extensibility (COLE) of soils is crucial for enhancing structural stability in civil engineering applications. Traditional methods for measuring COLE have some practical limitations. In routine soil analyses, clay and soil organic carbon (OC) are often measured, while soil hygroscopic water content (w(h)) is easy to determine on several soil samples simultaneously. The aims of the study were to (1) utilise two data partitioning approaches to develop regression models that estimate the soil COLE from hygroscopic water content, clay and OC contents, and (2) compare the model performance of the developed regression models. We used two data partitioning approaches. First, the calibration models were developed on 53 soil samples from Slovakia and validated with 24 soil samples from the United States (country-wise). Second, the calibration models were built from 67% of the entire dataset and validated with 33% (mixed data). Regression models based on hygroscopic water content accurately estimated COLE regardless of sorption direction or data partitioning approach (RMSE: 0.014-0.023 cm cm(-1)). The inclusion of OC in multiple linear regression models of clay only marginally improved COLE estimation compared to clay alone. For all models, the mixed data partitioning method showed better model validation performance than the country-wise approach. The COLE classes derived from the estimated COLE values compared favourably (72%-94% accurate) to the measured data. Thus, there is a great potential to estimate the COLE from readily available (clay and OC) or easily measurable (hygroscopic water content) soil properties. Highlights Hygroscopic water content (w(h)) is intimately linked to soil properties that determine COLE. Regression models based on wh or clay, and organic carbon content accurately estimated COLE. Data partitioning approach for modelling significantly impacted model performance Samples with very high COLE were better estimated by w(h)-based models than by other variables