Answers are grounded in this paper and cite their sources.
Evaluates four bias-correction methods for temperature in South America, showing the multilayer perceptron method as most effective.
Near‐surface air temperature from the ERA5 reanalysis is widely used for climate analysis, impact studies and environmental modelling over South America, but it exhibits systematic and regionally dependent biases that limit its direct applicability. This study aims to improve ERA5 daily maximum (TMAX) and minimum (TMIN) temperatures through physically informed regionalisation and bias correction. South America was clustered into homogeneous topoclimatic regions using k ‐means, spectral clustering and Gaussian Mixture Models, based on ERA5 thermodynamic variables, geographic coordinates and elevation, with the latter providing the most physically consistent partition into five clusters. Within each cluster, four bias‐correction approaches were evaluated: Mean Bias (MB) and Model Output Statistics (MOS), applied at the cluster level, and Spatial Bias Correction (SBC) and a Multi‐Layer Perceptron (MLP), which account for spatial variability across the broader domain. All methods were calibrated and assessed using independent subsets of in situ observations. ERA5 shows a pervasive cold bias in TMAX and a warm bias in TMIN, with the largest errors occurring in mountainous and high‐latitude regions. All methods reduce these errors, with the MLP providing the most consistent improvements across regions and seasons. However, its performance is interpreted in light of its greater flexibility and ability to represent non‐linear spatial–temporal relationships using physically meaningful predictors, rather than as evidence of universal superiority. The comparison reflects differences in modelling approaches, from linear and spatial interpolation to non‐linear learning, within a consistent regionalised framework. Model performance is influenced by the spatial distribution of observations. In data‐sparse regions such as the Amazon, MB, MOS and SBC are constrained by limited station coverage, whilst the MLP can partially compensate through cross‐regional generalisation, which may reflect the transfer of relationships learned in data‐rich regions. In mountainous and high‐latitude regions, all methods are affected by observational limitations and unresolved physical processes, resulting in higher uncertainty. Overall, the MLP demonstrates strong effectiveness for bias correction and shows potential for predictive applications.
No takes yet. Share an insight, caveat, or question.
Rozante et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: