Paraguay Forest Change Product


The wall-to-wall Forest Cover (FC) change map was developed for the entire Paraguay using the 1990s (T1) and 2000s (T2) Landsat observations. The T1 and T2 images were stacked to make an image pair and analyzed using an iterative clustering-supervised labeling method to produce a FC change map.


The iterative clustering-supervised labeling method consisted of two major processes: unsupervised isodata clustering and supervised labeling of clusters based on defined training pixels. During the initial clustering, an image-pair was classified using up to 250 clusters. The visual interpretation of Landsat images was then performed to identify training pixels for all clusters with the assistance of local experts. An in-house program called "supervised labeling" was used to count the training pixels within each cluster. If a cluster contains training data of only one class, the program automatically assigned that class to the concerned cluster. Clusters containing training data of multiple classes were called "cross-over" clusters. The pixels within the "cross-over" clusters were subjected to another isodata clustering- supervised labeling process. This clustering-labeling process was repeated until there were no more "cross-over" clusters and all pixels were classified appropriately. Because the "cross-over" clusters became finer through iterations until they contained training data of one class, this iterative procedure yielded highly reliable FC change product. Further, this study demonstrated that computer- processing time was considerably reduced, since further unsupervised classifications were carried out only either on parts of an image or on subsets of the clusters which belonged to more than one class or which were misclassified.

Accuracy assessment

Because all reference data were acquired around 2000, the accuracy assessment was performed on the FC map in the 2000s which was derived from the wall-to-wall map of FC change during 1990-2000 developed in this study. The 2000s FC map was then assessed at five locations using high resolution satellite images and two groups of locations using aerial photos. Approximately 300 samples were randomly selected over each location or location groups. The results show the minimum overall accuracy being 92% and the maximum being 97.5%, indicating that the 2000s FC map is considerably accurate over the area extent where reference data are available. The class specific user's and producer's accuracy values are mostly above or near 90%, demonstrating that the map is accurate for both classes, forest and non-forest. The accuracy of 1990s FC map should be expected as high as that of the 2000s FC map for the following reasons: 1) the 1990s Landsat images and the 2000s images had almost identical spatial and spectral characteristics; and 2) the two maps were derived simultaneously using a single method. Therefore, this study concludes that the map of FC change during 1990-2000 is highly reliable.

Reference data

Reference data for validating the product are high resolution satellite images and aerial photos. High resolution satellite images include 3 IKONOS images with a spatial resolution of 4 meters and 2 QuickBird images with a spatial resolution of 2.8 meters. A total of 211 aerial photos were acquired between 1999 and 2003 with 75 and 136 near/within the Cerrado region and the Atlantic forest region, respectively. Spatial resolutions for these photos range from 0.5- to 2-meter.


This product was funded under the NASA Land Use Land Cover Change Program, under award NAG59337. Local assistance and verification in Paraguay was provided by Guyra Paraguay. Product generation was conducted by the Deforestation Mapping Group at the Global Land Cover Facility, with Alice Altstatt and Sunghee Kim as primary analysts.

Paraguay Forest
Copyright © 1997-2016 University of Maryland. All rights reserved
University of Maryland National Aeronautics and Space Administration GOFC_GOLD