Mapeamento de culturas agrícolas a partir de classificação orientada a objeto subsidiada por técnica de mineração de dados

Abstract

Previous studies revealed difficulties in agricultural crop mapping by conventional classification techniques. Crops show diverse spectral patterns in satellite images due to its stage of phenological development, type of agricultural management among other characteristics, demanding therefore imagery with a good temporal resolution and low cloud cover, in addition to adopted classification methods. The objective of this study was to develop a methodology to map the sequences of agricultural crops in a study area inside the Cuiabá and São Lourenço watersheds, in the State of Mato Grosso, using multi-sensor imagery and complementary spatial data through a object oriented classification approach supported by a data mining algorithm. The EVI (Enhanced Vegetation Index) and Pixel Reliability of 23 MODIS images from the agricultural cycle 2010/2011, MOD13Q1 product, as well as previously existing land use and cover and vegetation map of the Upper Paraguay basin for the year 2008, conducted by the WWF (World Wide Fund For Nature), which had its segments actualized by Landsat TM images, were used for classification. Class separability according to pixels degradation levels was evaluated using the Jeffries Matusita (JM) distance. Imagery and spatial data layers were combined through the implementation of two classification models based on decision trees generated by the J48 data mining algorithm of the WEKA software (Waikato Environment Knowledge Analysis). Classifications were implemented in the InterIMAGE software, a platform developed for Object Oriented (OO) image classification and compared with two Maximum Likelihood classifications. Maximum Likelihood and OO classifications resulted in a Kappa Index (IK) of 0.70 and 0.53 respectively, indicating better overall performance of the conventional classifier. It could be observed however that with the exception of two classes, OO classifications resulted in User and Producer Accuracies above 63%, reaching levels of 90%, and that more satisfactory results were obtained by OO for fields with areas less than 100 ha.