Imputação múltipla e funções de pedotransferência para estimativa da densidade de solos da Amazônia oriental.

Abstract

Soil density (Ds) is a basic physical attribute in soil science, very important because of its relationship to various soil properties and processes such as hydraulic properties, soil compaction and erosion. It is a dynamic attribute that varies according to soil type, management, vegetation cover, etc. Despite the importance of Ds, it is common to find databases around the world that do not have measurements of this attribute for all or some records, because the measurements are laborious because it requires undisturbed soil sample collection (difficult to obtain). To solve this problem, pedotransfer functions are used to estimate soil attributes from available ground survey data. Another very common problem is the occurrence of incomplete sets of variables, due to several factors such as: typo, not measuring a certain attribute in a certain place (samples from different places). The general objective of this work was to evaluate the efficiency of the Multivariate Imputation by Chained Equations (MICE) method to fill the soil physical-water attribute database with missing data and to develop pedotransfer functions for soil density estimation. The data used were compiled from several sources, from 1997 to 2014 to compose the database, which contained two soil classes (argisols and latosols) at a depth of 0 to 60 cm, consisting of 631 samples and eleven variables. , of which five contained missing data. These were treated by imputing missing data using linear regression models. With imputation the size of the database and the overall distribution of the data were preserved. Soil density (Ds) was obtained by pedotransfer functions (FPT) using two database compositions: with imputation and without imputation of data. Predictor variables were selected by the stepwise method. The best performance was obtained for FPT1 (without imputation), which used the clay content and macroporosity as covariates, and for FPT4 (with imputation) that used as soil organic matter and silt as covariates. FPT4 is best suited for using covariates most easily found in a soil database. Of the FPTs available in the literature by Tomasella and Hodnett (1998) presented the best performance. Estimation of soil attributes using imputation database has the potential to solve problems of lack of soil data in the Amazon region.