Digital mapping of soil classes in Algeria – A comparison of methods

Abstract

Machine learning models have not yet been tested for digital soil mapping in the Algerian Sahara, where soil information is scarce. In this study, we compared six machine learning algorithms to predict and map soil classes in Zeb El Gherbi, Biskra Province. A total of 331 soil point observations classified according to the World Reference Base and optimal covariates were used as input data in the models. The final map was validated from data of 135 soil points. Our results showed that random forest was the most accurate algorithm. Elevation, multi-scale elevation and grain size index covariates had the most significant influence on the prediction of soil classes. The accuracy of the random forest model increased in the reference soil groups, with principal qualifiers composed of 13 classes compared to the reference soil groups with 5 classes. Haplic Petric Gypsisols, Gleyic Solonchaks and Haplic Gypsisols were the most common soils, constituting 67.46% of the total area mapped. The methodological approach developed could be used to accelerate and improve soil mapping in other Saharan areas of Algeria.