Comparison of linear, generalized additive models and machine learning algorithms for spatial climate interpolation

Abstract

Geospatial atmospheric data is the input variable of a wide range of hydrological and ecological spatial models, many of which are oriented towards improving the socioeconomic and environmental sustainability. Here, we provide an evaluation of machine learning (ML) methods for the spatial interpolation of annual precipitation, minimum and maximum temperatures for a mountain range, in this case, the Pyrenees. To this end, this work compares the performance and accuracy of multiple linear regressions (MLR) and generalized additive models (GAM) against five ML methods (K-Nearest Neighbors, Supported Vector Machines, Neural Networks, Stochastic Gradient Boosting and Random Forest). The ML algorithms outperformed the MLR and GAM independently of the predictor variables used, the geographical sector analyzed or the elevation range. Overall, the differences between ML algorithms are negligible. Random Forest shows a slightly higher than average accuracy for the spatial interpolation of precipitation (R2 = 0.93; MAE = 70.44 mm), whereas Stochastic Gradient Boosting is the best ML method for the spatial interpolation of the mean maximum annual temperature (R2 = 0.96, MAE = 0.43 ºC). Stochastic Gradient Boosting, Neural Networks and Random Forest have similar performances for the spatial interpolation of the mean minimum annual temperature (R2 = 0.98, MAE = 0.19 ºC). Results presented here can be valuable for the past and future climate spatial analysis, environmental niche modelling, hydrological projections, and water management.