Anais do XXXV Congresso Brasileiro de Ciência do Solo

MODELING AND REFINING SOIL MAPS FROM LEGACY DATA COMPARING KNOWLEDGEMINER WITH DECISION TREES

SÉRGIO HENRIQUE GODINHO SILVA⁽¹⁾; MICHELE DUARTE DE MENEZES⁽¹⁾; PHILLIP RAY OWENS⁽²⁾; NILTON CURI⁽¹⁾;
1 - UNIVERSIDADE FEDERAL DE LAVRAS; 2 - PURDUE UNIVERSITY;

Diverse projects are being carried out worldwide focusing on the development of more accurate soil maps. This work aimed to compare two data mining tools, KnowledgeMiner and decision trees, to retrieve the soil legacy data from a detailed soil map of a watershed in Minas Gerais, Brazil, and then to create and validate the soil maps in the field to identify the best method for future refining of soil maps. The study area is a watershed located in Nazareno county, state of Minas Gerais, Brazil. From the existing detailed soil map, terrain attributes information was retrieved by each polygon of the map. KnowledgeMiner and decision trees were employed to identify the pattern of each soil class according to 12 terrain attributes and to create a new soil map by method. Validation was performed in the field at 20 places chosen by cost-constrained conditioned Latin hypercube scheme. KnowledgeMiner maps had an accuracy of 80% and 0.6524 kappa index against 55% and 0.0674 for decision trees. Digital mapping tools are contributing to improve existing soil maps using legacy data. KnowledgeMiner had a better performance than decision trees to retrieve knowledge and map the soils of the study area.

[baixar arquivo pdf]