Accurate glacier mapping is crucial for assessing future water security in Andean ecosystems. Traditional accuracy assessment may be biased due to overlooking spatial autocorrelation during map validation. In recent years, spatial cross-validation (CV) strategies have been proposed in environmental and ecological modeling to reduce bias in predictive accuracy. In this study, we demonstrate the influence of spatial autocorrelation on the accuracy assessment of glacier surface predictive models. This is achieved by comparing the performance of several widely used machine learning algorithms including the gradient-boosting machines (GBM), k-nearest neighbors (KNN), random forest (RF), and logistic regression (LR) for mapping nine main Peruvian glacier regions. Spatial and non-spatial cross-validation methods were used to evaluate the model’s classification errors in terms of the Matthews correlation coefficient. Performance differences of up to 18% were found between bias-reduced (spatial) and overoptimistic (non-spatial) cross-validation results. Regarding only spatial CV, the k-nearest neighbors were the overall best model across Huallanca (0.90), Huayhuasha (0.78), Huaytapallana (0.96), Raura (0.93), Urubamba (0.96), Vilcabamba (0.93), and Vilcanota (0.92) regions, consistently demonstrating the highest performance followed by logistic regression at Blanca (0.95) and Central (0.97) regions. Our validation approach, accounting for spatial characteristics, provides valuable insights for glacier mapping studies and future efforts on glacier retreat monitoring. Incorporating this approach improves the reliability of glacier mapping, guiding future national-level initiatives.

Accurately glacier mapping is crucial for understanding climate change impacts, but existing efforts may be biased due to overlooking spatial autocorrelation during map validation. To address this, we compared several widely used machine learning algorithms as gradient boosting machines (GBM), k-nearest neighbor (KNN) and random forest (RF) with parametric logistic regression (GLM) and an unsupervised remote sensing-based method (NDSI) for mapping Peru's glacier regions in a thoughtful experimental setup. Spatial and non-spatial cross-validation methods were used to evaluate model’s performance and compared with a fully independent test set. Performance differences of up to 18% were found between bias-reduced (spatial) and overoptimistic (non-spatial) cross-validation results when compared to independent test set, emphasizing the need to consider spatial autocorrelation when using machine learning for glacier mapping. K-nearest neighbors (KNN) was the overall best model across regions consistently demonstrating the highest performance followed by logistic regression (LR) and gradient boosting machines (GBM). Our novel validation approach, accounting for spatial characteristics, provides valuable insights for glacier mapping studies and future efforts on glacier retreat monitoring. Incorporating this approach improves the reliability of glacier mapping, guiding future national-level initiatives.