Funções de pedotransferência para estimativa da densidade dos solos da Bacia do Solimões, sob Floresta Tropical Amazônica

Abstract

Soil bulk density (BD) data, commonly, are not available in soil survey reports making difficult the generation of soil attributes maps such as availability of water and carbon stock (CS). Lack of BD data is even more limiting in regions such as the Brazilian Amazon. In these regions, the dense forest and the limited access roads increase costs of soil mapping and sampling. In the absence of measured data, researchers have developed pedotransfer functions (PTFs) to estimate this attribute. Although these functions allow overcome the limitation of measured data, few studies in the literature get deeper on measuring the accuracy of these functions based on a set of external data. Furthermore, few studies quantify the effects of different PTFs on the final quality of maps of spatial variability of attributes such as carbon storage. Considering this explanation, this study had as objectives: a) to develop PTFs for estimating BD in soils of the Solimões Basin; b) to compare the performance of the models generated against other models found in the literature; and, c) to quantify the effect of using PTFs to estimate the BD, on the quality of spatial variability of CS. For the generation of PTFs to estimate BD, physical and chemical data measures in 96 soil profiles were used. The soils were sampled during a detailed soil survey carried out in the Base Operacional Geólogo Pedro de Moura, located in Coari, Amazonas State. Multiple linear regression models were developed and their performance was compared with those existing in the literature for the Brazilian Amazon region (Bernoux et al., 1998; Tomasella & Hodnett, 1998; and Benites et al., 2007). The following statistical indicators were applied: coefficient of determination (R2), the mean prediction error (MPE), the Root mean Square error of prediction (RMSPE), and the mean absolute error (MAE). From the different BD data (measured and estimated by PTFs), maps of spatial variability of CS were generated by using ordinary kriging. Also, to compare the maps obtained from the distinct BD data, spatial analysis was performed using map algebra, which consisted on subtracting the generated map with measured data from the maps generated with the PTFs. It was observed that the MG1, MG2 and MG3 models showed the best levels of MEP, MAE, and R2 RMSPE, in the validation phase. However, the MG3 selected the lowest number of variables, thus it was considered the best model. The maps of spatial variability generated for CS estimated by using BD from the models MG1, MG3, and Tomasella & Hodnett (1998) showed a range of values very close to the values found for the map of CS measured. On the other hand, the maps obtained by the models of Benites et al. (2007) and Bernoux et al. (1998) tended to overestimate the CS values. The MG3 was more efficient in predicting the BD for soils of Solimões Basin; which was expected, since the function was developed with the database collected in the same environment condition. As for the CS spatial variability, the map generated using the values calculated with the BD estimated by the MG3 model, presented the closest similarity to the map from CS measured values. The results of spatial analysis showed that the CS estimated by the MG1 and MG3 models had the lowest error, with values of 0.09 ± 0.5 kg C m-2 and 0.05 ± 0.5 kg C m-2, respectively. Besides, the maps showing the greatest errors were those generated using PTFs published by Benites et al. (2007) and Bernoux et al. (1998), with values of 1.56 ± 0.5 kg C m-2 and 1.44 ± 0.5 kg C m-2, respectively; when compared to the map of CS with measured values.