Publication details
Journal: Science of the Total Environment, vol. 1047, 182021, July 7th 2026
Doi: doi.org/10.1016/j.scitotenv.2026.182021
Arkiv: hdl.handle.net/11250/5535307
Archive: nva.sikt.no/registration/019f40eaa7e9-a4fb6417-00c6-457e-91b0-6fe8d202dc41
Summary:
Soil salinization, referring to the excessive accumulation of soluble salts in soils, adversely influences nutrient cycling, biodiversity, soil structure, crop production, soil health, and ecosystem functioning. Accurately assessing soil salinity via electrical conductivity (EC) is key to mitigating its impacts. Thus, developing predictive tools for soil EC at regional and continental scales is essential for sustainable soil management. Here, we apply machine learning models to predict soil EC in the European Union (EU) and United Kingdom (UK) soils using different environmental factors like soil, climate, topography, and satellite data as predictors. The model is trained by ≈40,000 soil EC data points from the 2015 and 2018 Land Use/Cover Area Frame Survey data (LUCAS) surveys, complemented by the EC observations from World Soil Information Services (WoSIS) dataset. To improve the model performance, a forward feature selection technique was used resulting in selection of 17 covariates out of initially 34 predictors. The final selected XGBoost model achieved R2 values of 0.68, 0.6, and 0.63 for the training, internal testing, and independent validation datasets, respectively. For the year 2018, we estimate ≈21.7 Mha of EU + UK land exceeds an EC of 0.6 dS/m (at a 1:5 soil to water ratio, the so-called EC1:5). This estimate should be interpreted as elevated predicted EC1:5, rather than a direct estimate of soils meeting protosalic diagnostic criteria. The output of the predictive model consists of a gridded dataset that illustrates the spatial distribution of EC1:5 throughout the study area for the year 2018, along with an associated uncertainty map with a spatial resolution of 1 km.