Digital mapping of soil classes with imbalanced class observations: A case study of Lombardy region, Italy

<p>Mapping soil classes to retrieve information for specific soil management strategies according to capabilities and limitations of soil types is an important and very useful application of Digital Soil Mapping (DSM). DSM involves the application of mathematical methods to develop models that explain the spatial distribution of soil characteristics using environmental variables as predictors. Machine learning (ML) algorithms is related to the ability to quantify the high-dimensional and nonlinear relationships between soil characteristics and environmental variables over diverse soil landscapes. However, most common ML algorithms works well on balanced datasets having classes that are approximately represented equally. However, imbalanced soil property class observations might lead to an underestimation or loss of minor classes (small N) and an overestimation of major classes (large N) while modelling with ML algorithms. In this study, we addressed this problem by investigating different resampling strategies such as data resampling techniques and synthetic resampling techniques, which have been developed to counteract imbalanced data problems. We compared the performance of three of the most common ML algorithms (decision tree, random forest, and multinomial logistic regression) with the respective resampling approaches in an agricultural lowland area of Lombardy region, Italy. The study helps to identify: i) the effect of using imbalanced distribution of class observations in digital soil mapping, ii) the resampling strategy that improves and gives better performance, and iii) the best performing predictive model for soil classification for lowland areas combined with an appropriate resampling strategy. The results yield valuable information on how to deal with imbalanced class observations for digital soil mapping in an agricultural lowland area.</p> <p>&#160;</p>