Estratégias para predição de classes de solo

Abstract

In order to provide information with greater agility and adequate spatial resolution to supply the demand for soil information, digital soil mapping (DSM) is an alternative to map classes and soil properties, taking advantage of the increasing availability of techniques processing and data mining. In this scenario, data that allow a clear understanding of scientific performance and relate them to the patterns of global scientific production can help in the paths to be followed by the research, and may even contribute with new public policies. The possibility of making use of previously generated information on the ground, called legacy data, can help as input information to the DSM at a reduced cost, since there is no need for new collections. As the DSM products make it possible to estimate uncertainty, a comprehensive analysis can contribute to map quality. If uncertainty is quantified and spatialized, this information can be used to improve sampling and optimize information generation. Thus, the objectives of this work were (1) to characterize the scientific production in digital mapping of soils in Brazil and in the world, from 1996 to 2017, in the Scopus and Web of Science databases; and (2) evaluate additional data collection techniques to improve soil class predictions using legacy data. For this, two studies were carried out. In the first, we searched for terms related to MDS in the databases, including searches for terms in the titles, abstracts, and keywords of articles. From this, a set of bibliometric indexes of the results were generated using the Bibliometrix package in the R environment. In the second study, a soil class map was generated based on environmental covariates using legacy data in an area of 13000 km² of the central region of Rio Grande do Sul State, which is among the priority areas of PronaSolos. The maps were evaluated by cross-validation and external validation, in addition to uncertainty maps expressing the areas with greater confusion of the model. In addition, strategies were tested to obtain additional points to the calibration set based on legacy maps and guided-uncertainty resampling. Study 1 demonstrated that, in the general context, the increasing number of articles in DSM was published for the most part in the Geoderma journal. Among the 10 journals most published articles, the Revista Brasileira de Ciência do Solo is the only open access journal. Although there are countries at the forefront of DSM, such as the United States and Australia, Brazil's position in the number of articles and authors cannot be overlooked, showing the importance of the country's participation in DSM research. Study 2 resulted in a map of soil classes, generated only legacy data, with an accuracy of 0.49 in external validation and a general uncertainty of 0.84. A hybrid set using legacy data from different sources was able to improve accuracy to 0.55 and reduce uncertainty to 0.77. However, while legacy map data has brought benefits to the model, they have shown inconsistencies due to its resolution. The uncertainty-guided resampling, by the improvement brought to the model using a small amount of data, was the strategy that demonstrated the greatest potential. Our data demonstrate that DSM is a promising technique and can be used as a methodology in the Programa Nacional de Solos (PronaSolos).