Mapeamento de solos com ênfase em análises multiescalares para aplicação na agricultura familiar

Abstract

Soil information is essential for the sustainable intensification of family production systems, and it is recommended that soil classes be mapped down to the subgroup level, in addition to the mapping of textural classes. In the Amazon context, soil information needs to be on a scale compatible with small rural properties, despite the large extension of the region, so that it can be useful for family farmers and technical assistance. To this end, it is necessary to adopt multiscale strategies for digital soil mapping. In the present study, the mapping of textural classes was performed on a scale of 1:25,000 and the mapping of soil classes on a scale of 1:100,000. On the other hand, the multiscale influence of topography on soil distribution has a complex pattern that is related to the overlapping of pedological processes that occurred at different times and driving forces that are correlated with different scales. In this sense, we tested the hypothesis that generalized geomorphometric covariates at multiple scales can improve the accuracy of pedometric modeling. In the mapping of textural classes, the Random Forest algorithm was applied in a multiscale geomorphometric database to predict the granulometric percentages on the soil surface. The generalized covariates improved the accuracy of the textural classification, with the Kappa Index increasing from 0.43 to 0.62. In the mapping of soil classes, fuzzy set models were applied using generalized covariates to define preferential environments for the occurrence of the pedogenetic processes of gleization, elutriation and clay translocation, as opposed to the wide spatio-temporal distribution of the ferralization process. Fuzzy membership rules and covariate scales were defined for each of the processes, considering the soil knowledge obtained from soil surveys in transects. The generalized covariates improved the accuracy of the mapping, with the Kappa Index going from 0.4 to 0.69. The results demonstrate that knowledge-based models that observe the multiscale influence of topography on soil formation processes can improve the prediction of soil classes in intermediate-level soil surveys. Therefore, it is concluded that the use of generalized geomorphometric covariates in multiple scales results in greater accuracy of the models, being a flexible method for use both in machine learning approaches and in modeling based on the pedologist's knowledge. Still, from the geomorphometric point of view, it is necessary to test generalization methods, focusing on the preservation of features relevant to the soil-landscape relationship.