Krigagem com regressão espaço-temporal com modelos GAMLSS

Abstract

The spatio-temporal variability of a phenomenon can be decomposed by the trend component and the stochastic residue. This thesis aims to consider different probabilistic models in the adjustment of the trend component, since the precise estimation of the spatial-temporal distribution of the response variable requires an adequate probability distribution. Gaussian regression models and the generalized additive models for position, scale and shape (GAMLSS) were considered for the adjustment of this component. The residuals produced by this regression were modeled by spatio-temporal covariance functions, which took into account spatial, temporal and spatial-temporal dependence. The Spatiotemporal regression Kriging that combined kriging on the residues with the GAMLSS regression showed good results, as evidenced by the statistics obtained after cross-validation. The methodology presented in this thesis allowed the creation of future scenarios for the study area, interpolating unobserved locations and predicting estimates in future years.