Addressing the issue of digital mapping of soil classes with imbalanced class observations

Considering the nature of soils distribution, an important modeling issue in soil class mapping is imbalanced class observations. Imbalanced number of data in observed soil classes in an area can result in the underestimation or loss of minority classes and an overestimation of the majority classes in predictive modeling. The effect of this phenomenon is that an area of land with comparatively fewer soil profile observations could be unmapped in the digital maps. To address this problem, this paper investigated the usefulness of data pretreatment techniques called over- and under-sampling of data applied on three predictive models including decision trees (DT), random forest (RF), and multinomial logistic regression (MNLR). The study area is situated in the northwest of Iran with 452 profiles observations on a regular grid covering about 12,000 ha. This area has 8 USDA soil great groups with an imbalanced frequency distribution. Results showed that modeling using imbalanced distribution of class observation caused uncertain maps with minority classes being lost and relatively poor accuracies. After data treatment, with over- and under-sampling, all models showed significant improvement in maintaining the minority classes, in both calibration and validation evaluations. Balancing the classes led to a notable decrease in uncertainty of all 3 models by decreasing the confusion index and raising the probability of occurrence for the soil classes in the final maps. Comparing the 3 models, decision trees showed the largest calibration and validation accuracies with and without data treatment. RF has an issue of overestimation of some of the majority classes. Data resampling technique can be a useful solution for dealing with imbalanced class observations to produce more certain digital soil maps.