Brazilian land use data currently come from the Brazilian national agricultural census, while land cover data come from global data sets with sparse temporal coverage. These sources are no longer adequate for the needs of the Earth system modeling community. The goal of this research is to develop a methodology to improve information about land use and land cover trajectories in Brazil through the use of long-term satellite image datasets with high temporal frequency. These datasets, which yield a sequence of data points in a time series that reflects vegetation phenological cycles (Figure 1), can be used to detect land use and land cover changes.
Figure 1. Phenological vegetation cycles reflected in vegetation index time series.
A 2-band Enhanced Vegetation Index (EVI2) provided by LAF/INPE [1] was used in this work. These data are based on the MODIS MOD13 Q1 product, 16-day composite images with 250 m spatial resolution. The classical data mining method of Dynamic Time Warping (DTW) [2] was used to match typical vegetation patterns in long-term EVI2 time series. A test area with 8.325 ha, located in Mato Grosso Brazil, was compared to official TerraClass maps [3] from 2008 and 2010 for the following classes: forest, pasture, and crop.
In preliminary tests the DTW classification shows satisfactory agreement with official TerraClass maps from 2008 and 2010 with global accuracy of 78.2% and 85.0%, and Kappa coefficients of 0.62 and 0.72, respectively, for each year. The main classification errors occur in boundaries between different land covers (Figure 2), where MODIS pixels contain a mixture of surface reflectance. Furthermore, the maps yielded by DTW display impossible land use/cover transitions, such as the following three years sequence: forest→clear cut→ forest.
Exploratory DTW results show significant potential to detect land use and cover changes, which is useful not only to provide land use and cover maps but also to understand land cover changes. To improve the results of the DTW classifications, it is essential to include post-processing steps with rules for land use and cover transitions and spatial filtering.
References
[1] Freitas, R. M. de; Arai, E.; Adami, M. et al. Virtual laboratory of remote sensing time series: visualization of modis EVI2 data set over south America. Journal of Computational Interdisciplinary Sciences, v. 2, n. 1, p. 57–68, 2011.
[2] CRA/INPE. Projeto TerraClass. 2012. Valuable: Access: July 2013.
[3] Rabiner, L.; Juang, B. Fundamentals of speech recognition. Prentice-Hall International, Inc., 1993.
Note
Victor Maus, of the National Institute for Space Research/Federal University of Pampa, Brazil, is a Brazilian citizen. He was funded by IIASA's Brazilian National Member Organization and worked in the Ecosystems Services and Management (ESM) Program during YSSP.
Please note these Proceedings have received limited or no review from supervisors and IIASA program directors, and the views and results expressed therein do not necessarily represent IIASA, its National Member Organizations, or other organizations supporting the work.
Further information
International Institute for Applied Systems Analysis (IIASA)
Schlossplatz 1, A-2361 Laxenburg, Austria
Phone: (+43 2236) 807 0 Fax:(+43 2236) 71 313