We have classified the land use and land cover (LULC) for an extensive Landsat image time series throughout our study area. As stated on the Imagery Products page, our study area spans four different Landsat scenes. Our main focus for ground data collection to support the classifications has been in the area covered by the Path 9 Row 60 images, as that scene contains the vast majority of our study sites for all projects. However, the lack of ground data in the other three Landsat scenes has led to the implementation of two different classification approaches.
Path 9 Row 60 – Detailed Classification
We have produced a detailed 16-class classification for the P9R60 Landsat images. This classification scheme captures different stages of forest growth and regrowth, many important cash and subsistence crops, areas of new urban growth (seen as barren), and existing towns in the region. The actual classes are:
- Primary Forest
- Swamp Forest
- Secondary Forest
- Rastrojo (Early Secondary Succession)
- Pasture with No Trees
- Pasture with Few Trees
- Pasture with Many Trees
- African Palm
The classifications have been completed for seven dates from 1986 to 2002.
The classification process is a hybrid unsupervised-supervised approach that maximizes the classification potential with a minimum of ground-truth data available.
For more information on the methodology, please contact us.
Other Scenes – More Generalized Classification
The lack of sufficient ground data to support the classifications of the Path 8 Row 60, Path 8 Row 61, and Path 9 Row 61 scenes led to the development of an alternative classification approach. This approach also utilizes a hybrid unsupervised-supervised methodology, but is less rigorous than the method used on the P9R60 scenes.
Due to the lack of ground data in these areas, the resultant classification scheme is much more general in nature. The classes are:
For those images that contain clouds and cloud shadows (which unfortunately is nearly all of them), the clouds and shadows are coded as Unclassified so that they are seen as background. In effect, there are only four classes of any use, as the Unclassified pixels are omitted from analyses.
An additional step has been added to the classification of these scenes to extract additional information. The generation of fractional cover layers provided us with the ability to add additional detail to the four general classes. For example, we can now separate Forest into areas of varying degrees of density, and we can effectively separate out pasture from agriculture in the Agriculture/Pasture class.