Comparison of outlier detection methods and their impact on rangeland measurement and Assessment studies

This study compared of univariate outlier detection methods among vegetation data in a study of the effect of grazing intensity in the rangelands of arid regions. For this purpose, after measuring the vegetation cover in the rangeland and before the statistical analysis, the presence of outlier data was examined as the assumption of parametric comparison tests. In this study, eight methods including the boxplot and IQR (Tukey method), standard deviation of the mean (three-sigma rule), median absolute deviation (Hampel method), trimmed mean, 1st percentile and 99th percentile, The Chi Square test (χ²), the Grubbs Test (ESD) and the Rosner test (generalized ESD) were used. The results showed that the vegetation cover of rangelands with light and moderate grazing intensity was not normally distributed (Shapiro-Wilk test: p≤0.05). Even deletion of outliers did not lead to a normal distribution, but it resulted in the homogeneity of variances (Levene's test: p≥0.05). The modified Z-score and the Grubbs and Rosner tests (p≥0.05) did not identify outliers from the vegetation cover data. Among the methods evaluated, the boxplot and MAD method, which are not dependent on the mean, are more suitable for the vegetation cover. Therefore, before performing any comparison test, a combination of visual and statistical methods is recommended to evaluate the presence of outliers.