Geoestatística e modelos não lineares e de efeito misto para modelagem da capacidade produtiva do sítio em florestas de Pinus taeda L.

Abstract

The research explored the modeling of dependence on autocorrelated data using covariates of low cost and high sampling intensity, originating from remote sensing, by determining of the site capacity production of Pinus taeda forests. For this end, generalized nonlinear and mixed effect models and geostatistics techniques used. Geostatistics was applied with external drift kriging, using the dominant height with predictor variable of 255 sample plots processing, collected in the municipality of Coronel Domingo Soares, Paraná, Brazil and eight remote sensing covariates originating from LANDSAT-8 image processing (SRI, NDVI, NDWI, NBR and SAVI) and SRTM (altitude, slope and relief) in addition to a multiple model chosen by Best Subset Selection. The semivariograms tested were spherical, exponential and Gaussian and the goodness of fit was given by the leave-one-out cross validation, considering statistical criteria and graphical analysis of residuals. Generalized nonlinear (gnls function) and mixed effect (nlme) modeling used the variables dominant height and age of 1,277 plots, remeasured at least 3 times, collected in the south-central region of Paraná, Brazil, in addition to the slope covariate originated from the SRTM image and classified in four classes. The traditional model (nls) was compared with techniques that used the slope class with a covariate in the parameters models and with a random effect; in all possible scenarios of inclusion or not in the model parameters. For those selected, data dependence was modeled by the variance and covariance matrix. The quality of the adjustment was based on statistical criteria and graphical analysis of the residuals. The methodologies tested, proved to be efficient to achieve the proposed objectives, because in addition to correcting problems of heteroscedasticity and error correlation, they presented better statistics than the traditional model, consequently producing more reliable estimates.