Avaliação do uso de variáveis temporais na classificação da cobertura da terra

Abstract

The monitoring of land surface has been performed by remote sensors aboard satellites, which allows us to obtain information on changes of landscape over time. The use of multitemporal data aids in the classification of land cover and enhances its characterization considering phenological aspects, as well as continuous or sudden variations in the landscape. However, the implications of different temporal frequency of remote sensing images in a classification over time are unknown. Considering the difficulty of acquiring quality images for comprising a time series, it is highly important to evaluate which frequency is necessary to obtain a precise multitemporal classification. Therefore, this work had the objective of extracting information of time series remote sensing data to improve the classification precision of land cover, thus contributing with the understanding of temporal patterns of various land cover classes and indicating appropriate approaches for the mapping that will be performed during the implementation of the CAR. The study area is located in the Center-West of the state of Minas Gerais, Brazil, and the analyses were confined to an area of 625 km2 referent to a Rapideye scene (tile 2330320). All images of the Landsat-5 TM sensor (orbit/point 219/073) available were acquired between 2000 and 2011. However, as tropical regions present large incidence of clouds and eventual unavailability of TM images, synthetic images derived from the fusion of TM and MODIS images were used. In order to do this, a set of 273 images of the MODIS sensor (product MOD13Q1) were acquired. The fusion of TM and MODIS was performed by the STARFM algorithm. The original and synthetic TM images comprised Stational (4 images/year), Bimonthly (6 images/per year), Monthly (12 images/year) and Complete (23 images/year) time series. The time series were analyzed by the greenbrown package and the generated results were incorporated as attributes in the machine learning algorithm. The results showed that the use of parameters generated from the greenbrown package, even when using synthetic TM images, provided an improvement in the performance of the classifiers, highlighting the SVM and RF algorithms. In addition, no significant differences were found between the classifications generated using different temporal frequency attributes. Thus, the multitemporal classification of the Rapideye image using parameters extracted from a TM time series, was shown to be applicable and promising, since the Rapideye data did not dispose of sufficient multidata images.