Coping with imbalanced data problem in digital mapping of soil classes

Abstract An unsolved problem in the digital mapping of categorical soil variables and types is imbalanced number observations, which leads to reduced accuracy loss minority class (the with a significantly lower observations compared other classes) final map. So far, synthetic over‐ under‐sampling techniques have been explored science; however, more efficient approaches that do not drawbacks these guarantee retention classes produced map are essentially required. Such suggested present study for include machine learning models ensemble gradient boosting, cost‐sensitive one‐class classification (OCC) combined multi‐class classification. In this regard, extreme boosting (XGB) as an learner, decision tree (CSDT) within C5.0 algorithm, support vector (OCCM) were investigated eight great groups naturally frequency northwest Iran. A total 453 profile data points used area. split was done manually each separately, resulted overall 70% calibration 30% validation. The bootstrapping approach (with 10 runs) performed produce multiple maps model. bootstraps evaluated against hold‐out validation dataset. average values measures, including Kappa (K), (OA), producer's (PA) user's (UA), explored. addition, results previous same area, resampling deal mapping. findings show all three methods can well problem, OCCM showing highest K (= 0.76) OA 82) stage. Also, model Comparing demonstrates newly remarkably increase both individual