High deforestation rates in Amazonia have motivated considerable efforts to monitor land-cover changes based on satellite images and image processing techniques. Most commonly, MODIS images are used to provide low-cost region-wide coverage at nearly monthly frequencies, but they offer only coarse resolution; Landsat TM has been used in a majority of studies for nearly two decades, but these data are expensive, and provide, at best, yearly coverage because of clouds. Here, a new approach to estimate forest change is proposed based on the integration of TM and MODIS images. TM images are processed using a hybrid approach including spectral mixture, expert rules, and unsupervised classification, to generate a reference forest image. Three fraction images are derived from MODIS surface reflectance data; expert rules are used to generate a refined vegetation image; and a regression model is then developed between the TM-derived forest and MODIS-derived vegetation data to assess the fractional forest area. This approach was initially applied to 2004 MODIS and TM images from Rondônia, and the regression model was transferred to 2000 and 2006 MODIS images. A similar exercise was made in Pará state for the estimation of forest area in 2005. Compared to TM-derived reference data in Rondônia, the system error for the MODIS-derived forest areas was 1.56% and 4.19% for 2004 and 2000 images, respectively. Compared to INPE Prodes data, the error for total forest area in Rondônia in 2004 and 2000 are -0.97% and 0.81%, respectively. The major advantage of this approach is that coarse spatial resolution images from MODIS and AVHRR can be used to estimate fractional forest cover for large areas in a short time, requiring limited work, but yielding accuracies comparable to Landsat TM-derived results.
Science Theme: LC (Land Use and Land Cover Change)