Monitoring selective logging in tropical forests, a heuristic approach aided by automated classification. The techniques in Brown, et al. (2005), described on the Multi-temporal Surveys and Biomass Estimates page, used automated image processing to make the visual inspection and measurement of individual trees in Stereo Analyst as efficient as possible. We have continued to refine this approach to monitor and measure forests around the world for biomass production, natural mortality, and illegal clearing or selective logging.
The ability to collect elevation data and tightly register multi-temporal coverages makes it possible to track and measure changes in the canopy with an accuracy and efficiency that cannot be matched by object-based image processing of one-time coverage or high resolution satellite data.
In this mosaic section from a series of surveys to monitor illegal logging on the Los Amigos Conservation Concession in southeastern Peru , two 1-m DEMs were compiled from sets of imagery flown three years apart and then registered to each other, subtracting the second DEM from the first to create a difference image of changes in the surface of the canopy since the first set of photographs.
That difference DEM was then converted to a vector point file and automatically edited to remove all data except new gaps in the canopy more than four meters across and eight meters in depth, indicating sites to review.
There are two basic approaches to identifying selective logging from aerial imagery. One is an attempt to locate and classify gaps by brute force image processing. Our approach is different. We believe visual inspection is necessary to identify the subtle ancillary signs that distinguish harvesting from natural mortality or simple shadow effects. A photointerpreter is better and faster at this than software, so we use the preliminary image processing to generate potential sites, making the photointerpreter's job more efficient.
In this case, each potential site was inspected with the two mosaic strips in geographically linked windows, allowing the interpreter to assess the size of missing trees and determine if the canopy showed signs of removal by logging rather than loss by natural mortality, the latter being more likely in this example. That inspection process sounds labor-intensive, but once the initial automated selection of sites has been derived from the difference image, it is actually faster, more efficient, and more accurate than methods using software-driven decision making routines.
Image at right requires use of red/blue glasses to view in stereo.
Logging sites identified in the second set of images are outlined and their ground elevations measured in Stereo Analyst. These points and polygons are then superimposed on the first set of images (in Stereo Analyst) to measure the crown areas of missing trees and to determine their heights by subtracting the ground point elevations at the gaps from the elevations at their crowns. Dr. Brown’s allometric regressions can then be applied to these measurements to calculate an estimate of the biomass or board feet removed by logging.
Monitoring logging with non-multi-temporal coverage at smaller scales. This approach can also work to simply detect selective logging with one-time regional coverage at smaller scales if a corresponding high resolution DEM of the canopy is still generated from the imagery. In this example of 50 cm-per-pixel color coverage (lower right), larger abrupt gaps in the canopy due to selective logging are clearly visible, but older areas of disturbance that have partially grown back or been covered with brush are more difficult to tag for inspection as potential logging sites because they lack sharp shadows and have a similar color and texture to the surrounding canopy. A 2-meter DEM can be generated from this imagery (lower left), and processed to identify distinct neighborhood variations in elevation that make these areas easier to identify as potential sites in a preliminary classification, because the contrast between the re-grown areas and their immediate neighbors becomes more distinct and visually apparent (example in red outline). In a multi-tiered tropical forest canopy, the same approach can be used as a first cut to classify large emergents that rise above surrounding trees.