Influences of outliers on performance of geographically weighted random forest for modelling cadmium concentrations in topsoil of the northern part of Ireland

Cadmium (Cd) is a toxic element ubiquitously distributed in the environment. Numerous models have been employed to predict soil Cd concentrations, among which local models can better capture spatial patterns and yield more accurate predictions than global models. However, their sensitivity to outliers could lead to substantial local errors. In this study, we aim to assess and reduce the outlier effect in local modelling based on topsoil Cd in Ireland from Tellus project and 12 influential factors. Geographically weighted random forest (GWRF) was integrated with outlier detection tools Local Moran's I (GWRF-LISA) and Z-score normalization (GWRF-Z). The local models were compared against traditional global random forest (RF). Results showed that outliers could cause radial clusters, leading to spatially autocorrelated residuals. This effect strengthens with increasing bandwidth in local models. Z-score can effectively reduce outlier effect by adaptive removal of outliers. Among the four models, GWRF-Z produced the most accurate predictions, but its interpretability was limited by small bandwidths. SHAP values of RF revealed that precipitation, pH, and soil type were dominant factors in about 60 % of the study area, indicating the significant role of pedoclimatic processes in Cd distribution in Ireland. This study has clarified the influence of outliers in local modelling and highlighted the effectiveness of Z-score in reducing outlier effect. The proposed approach showed potential applications in broader regions. These findings provide a scientific basis for spatially targeted interventions and support local decision-making.