Amazon and Central Africa Forest Change Products

Description

The NASA Landsat Pathfinder Humid Tropical Deforestation Project mapped global deforestation for the humid tropics. Data sets from both the TM (Thematic Mapper) and MSS (Multispectral Scanner System) of Landsat were used for three time periods in the 1970s, 1980s, and 1990s. The project focused on the three regions where most of the tropical deforestation in the world has occurred - the Amazon Basin, Central Africa, and Southeast Asia. Mapping deforestation in these three regions accounts for the majority of deforestation activities in closed tropical forests worldwide.

Methods

Digital image processing in conjunction with spatial analysis in a Geographic Information System are effective means for quantifying deforestation We use high resolution Landsat because it yields much better precision than AVHRR-based analyses. Automated classification and manual editing has been found to provide significantly faster and more accurate than hand digitizing alone. Further, it would be very difficult, if not impossible, to reproduce the automated classification level of detail by hand digitizing alone. The approach adopted in processing Landsat data is to exploit automated methods to the fullest extent possible but then to use the skills of the human interpreter to improve the classification.

Details of this methodology as carried out by the University of Maryland are described in the following paragraphs:

Processing the raw satellite data into our vegetation classes:
The University of Maryland Landsat Pathfinder Project uses mulit-spectral/multi-temporal data sets to produce accurate, consistent and rapid classification. Under our approach we coregister images from different dates for the same WRS path/row tile. We then use the spectral bands for both image dates as input for unsupervised clustering. The resultant clusters represent both cover types which remain unchanged between the dates and areas which have changed. This procedure has been found reliable in distinguishing between changes due to phenological change from those due to more permanent changes associated, for example, with deforestation or regrowth.

We use PCI's EASI/PACE software for image processing and coregistration. GCPWorks, a module of EASI/PACE, is used to coregister the images. We currently use analyst identified control points for coregistration. While we are testing automatic procedures, these methods have not yet yielded consistent sub-pixel registration we obtainable from our control points identified by image analysts.

Our unsupervised clustering algorithm uses the EASI/PACE histogram clustering process Isodata clustering (ISOCLUS). We use ImageWorks, a module of EASI/PACE, to display and compare the output of the ISOCLUS with the raw data bands. The output clusters are color coded using a pseudocolor table(PCT). Usually the initial clustering will not be enough to completely distinguish between classes. A bitmap mask is created of the output clusters that are confused between two or more classes and an additional ISOCLUS is run on the data under this mask. This process is repeated until a satisfactory discrimination is achieved. The ISOCLUS output clusters are then assigned to our classes. (Panamazon; forest, deforestation, revegetation, non-forest vegetation, undifferentiated unforested, cloud/cloud shadow, and water. Central Africa; forest, degraded forest, non-forest, cloud/cloud shadow, and water.) These steps produce a classification where almost all the polygons of the desired land cover types are identified.

Editing/Quality control of classified image:
While the iterative ISOCLUS procedure produces a much more accurate product then previous procedures using supervised classification and hand editing of a vector product, there are usually still small corrections that need to be made by hand.

Currently, edits are done directly on the raster product. The raster product is vectorized and overlaid directly onto the raw image data in the ImageWorks image handler. This allows the image processor to use all the image enhancement tools needed to appropriately interpret the image as well as compare the output product to the raw image data from both dates being considered in the classification. If the image analysts finds errors in the classification they draw bitmaps over that area and either edit the output product using a modeling statement to reassign the classes or run an additional ISOCLUS on the area. This decision is dependent on the complexity and size of a regio. More difficult areas have the clustering procedure run on them again.

Common corrections include: aggregating clouds and heavy haze into the cloud class; correcting computerized misclassifications between water, cloud shadow, and burn scars in non-forest, all of which have very similar spectral signatures; correcting for misclassification between deforestation and non-forest, as well as topographic effects.

A final assessment is carried out by the laboratory manager. This helps ensures consistency of results. Once the lab manager feels a coverage is complete, the project PI's and personnel who have visited the field review the finished coverage based on their field experience. Any questions about interpretation that cannot be answered during this process are recorded in the IMS, and in-country experts are contacted for advice. A mechanism is firmly in place where, as auxiliary information is made available, the coverage can be improved.

The current system is producing a consistent and accurate product as is demonstrated by the fact that little, or no, thematic corrections are necessary when adjacent coverages are joined together.

Country product creation:
Once the images are finished, they need to be converted to the ARC/Info Grid format and registered to adjacent tiles and to their correct location on Earth. Until this point, they are in the coordinate system provided by the satellite meta data, and while this is close, it is not correct. We are using the Digital Chart of the World's country boundaries for this registration. The DCW is a 1:1,000,000 scale vector basemap of the world. It was originally created by the United States Defense Mapping Agency (DMA) and was adapted for use with Arc/Info software. The primary source for the DCW is the DMA Operation Navigation Chart (ONC) series.

Each country border coverage is moved to the corresponding border in DCW. This can be performed accurately in most locations because the border follows rivers that are easily distinguished in the images. The scenes are also moved so that features match up in adjacent scenes. The registration process involves affine moves and rotations of the complete scenes. No rubber sheeting is being performed on the scenes. Since all images for a WRS path/row tile will be coregistered to the same image this georegistration step needs to be done only once. All subsequent images can be transformed to the location of the 'base' georegistered grid.

In some locations country boundaries may be locally systematically displaced from topographic features such as rivers apparently as a result of local errors in the DCW. In such cases the topographic features are used as the boundary.

Once the grids are georegistered, they are merged together in ARC/Info. The grids are merged so as to maximize the usage of clear, cloud-free imagery.

History

The NASA Landsat Pathfinder Humid Tropical Deforestation Project is funded through a collaborative effort between the University of Maryland at College Park's Geography Department, The University of New Hampshire's Complex Systems Research Group and NASA Goddard Space Flight Center's GIMMS Group. The goal of this work is to map global deforestation for the humid tropics. Data sets from both the TM (Thematic Mapper) and MSS (Multispectral Scanner System) of Landsat are being used for three time periods in the 1970s, 1980s, and 1990s.

The project is focusing on the three regions where most of the tropical deforestation in the world has occurred - the Amazon Basin, Central Africa, and Southeast Asia. Mapping deforestation in these three regions will account for the majority of deforestation activities in closed tropical forests worldwide. As currently configured, the Pathfinder will not compile an inventory of the entire tropics. Some parts of South Asia, West Africa, Madagascar, and Central America will not be covered. About 75% of the tropical rain forest areas are included in the Pathfinder data plan. The University of Maryland has responsibility for the non-Brazilian Amazon Basin (also known as the Pan-Amazon) and central Africa. The University of New Hampshire has responsibility for the Brazilian Amazon Basin and Southeast Asia.

Amazon & Africa
Copyright © 1997-2016 University of Maryland. All rights reserved
University of Maryland National Aeronautics and Space Administration GOFC_GOLD