Integrating local spatial outliers into context-aware spatial prediction models

Improving spatial prediction remains a key challenge in geographical and environmental modeling, particularly in regions with strong spatial heterogeneity. This study proposes a local-outlier-based context-aware model, named LOCA, for precipitation spatial prediction. The novelty of LOCA lies not in spatial stratification or local modeling alone, but in formalizing local outliers of explanatory variables as continuous spatial contextual predictors and integrating them with spatial stratification and stratum-specific machine-learning models. Unlike conventional approaches that mainly treat outliers as errors or noise, LOCA uses positive and negative departures from the surrounding spatial background to represent local anomaly information. The model was evaluated through a case study of precipitation prediction in Qinghai Province, China, using CRU TS precipitation data, ERA5 meteorological variables, and topographic data for 2022. Because the target precipitation product was derived from CRU TS at 0.5° resolution, this case study should be interpreted as a statistical downscaling and fine-resolution spatial prediction experiment rather than direct validation against independent 1 km precipitation observations. K-means, Gaussian mixture model, and spectral clustering were used for spatial stratification, and multiple machine-learning models were compared within different strata. Model performance was assessed using spatial block cross-validation and compared with global machine-learning models, stratified context-aware models without local outlier features, and additional spatial reference models. The best-performing configuration, GMM-based LOCA, achieved an R2 of 0.858 ± 0.125, RMSE of 77.67 ± 15.43, and MAE of 57.12 ± 14.44. The improvement was associated with two complementary mechanisms: spatial stratification enhanced regional adaptability, while local outlier-derived features provided supplementary information on local spatial contrasts. These findings suggest that explanatory-variable local outliers can provide useful contextual information for statistical downscaling and spatial prediction in heterogeneous environments.