부산대학교 도서관

Soil salinization is one of the main threats to soils worldwide, which has serious impacts on soil functions. Our objective was to map and assess salt-affectedness on arable land (0.85 km2) in Hungary, with high spatial resolution, using a combination of ensemble machine learning and multivariate geostatistics on three salt-affected soil indicators (i.e., alkalinity, electrical conductivity, and sodium adsorption ratio (n = 85 soil samples)). Ensemble modelling with five base learners (i.e., random forest, extreme gradient boosting, support vector machine, neural network, and generalized linear model) was carried out and the results showed that ensemble modelling outperformed the base learners for alkalinity and sodium adsorption ratio with R2 values of 0.43 and 0.96, respectively, while only the random forest prediction was acceptable for electrical conductivity. Multivariate geostatistics was conducted on the stochastic residuals derived from machine learning modelling, as we could reasonably assume that there is spatial interdependence between the selected salt-affected soil indicators. We used 10-fold cross-validation to check the performance of the spatial predictions and uncertainty quantifications, which provided acceptable results for each selected salt-affected soil indicator (for pH value, electrical conductivity, and sodium adsorption ratio, the root mean square error values were 0.11, 0.86, and 0.22, respectively). Our results showed that the methodology applied in this study is efficient in mapping and assessing salt-affectedness on arable lands with high spatial resolution. A probability map for sodium adsorption ratio represents sodic soils exceeding a threshold value of 13, where they are more likely to have soil structure deterioration and water infiltration problems. This map can help the land user to select the appropriate agrotechnical operation for improving soil quality and yield.