Detecting Forest Degradation in the Congo Basin by Optical Remote Sensing
Rahm, Mathieu1; Cayet, Lauriane2; Anton, Vrieling3; Mertens, Benoit4
1Eurosense, BELGIUM; 2Walphot, BELGIUM; 3ITC, NETHERLANDS; 4IRD, FRANCE
Reducing carbon emissions from deforestation and forest degradation is a major objective in international climate change mitigation efforts, in particular for the REDD+ initiative. Carbon emissions associated with forest degradation are estimated to be three times greater than those from deforestation in sub-Saharan Africa. Nevertheless, while deforestation can be assessed accurately through remote sensing, detection and mapping of forest degradation is more challenging because the features of interest are generally of small spatial size, and vegetation regrowth quickly affects (within a year) the spectral reflectance of the degradation features.
To date, there are no widely accepted and operational approaches for monitoring forest degradation by remote sensing. In most studies, medium resolution sensors such as Landsat (30m spatial resolution) have been used so far to address this issue but many small-size degradation features (such as small-scale non-mechanized logging) cannot be detected at that spatial resolution.
For REDD+, a simple, repeatable and robust method using higher spatial resolution data with high temporal resolution seems the best way forward for a detailed mapping of forest degradation.
In this paper, we present a methodology to detect, map and quantify forest degradation processes and area of changes (e.g. forest canopy gaps, small clearings, logging roads) from optical satellite imagery at high spatial resolution (<10m). The method is based on semi-automated, object-oriented image classification. Multi-temporal analysis is carried out (i) to detect and quantify areas where changes occurred, and (ii) to identify different intensities of degradation.
The approach is based on multi-date segmentation, which creates multi-date objects grouping pixels that are spatially, spectrally and temporally similar (Figure 1). Step 1 consists of delineating "large objects", using spectral, scale and compactness parameters. In step 2, for each image individually, small patches of bare soil (canopy gaps, narrow logging roads) are delineated as "small objects" within the large objects of step 1. Figure 1c and 1d illustrate the two levels of image segmentation for t1 and t2 respectively. Step 3 consists of tabulating the amount of bare soil at each time step within the large objects. Five levels of forest degradation are defined for each object according to the percentage difference of bare soil between t1 and t2 (Figure 1e). Large objects whose percentage of bare soil is higher in t1 as compared to t2 are classified as regeneration. Large t2 objects for which the threshold of tree cover falls below the level defined for forest (10 or 30 percent) should be considered as deforested areas even if in the legend they are integrated in degradation level 5. The approach was applied on two case studies in Central Africa, through two related projects: (1) REDDiness which aimed at evaluating the potential of satellite imagery to assess forest degradation in two small study sites, one in Gabon and one in the Republic of Congo; (2) EO4REDD which focuses on the development of a robust, affordable and reliable forest degradation monitoring method for the Mai Ndombe region in the Democratic Republic of Congo. Based on results of the two projects, this paper discusses the potential of high and very high resolution imagery to detect forest degradation as a requirement to REDD+. We argue which current and future (such as Sentinel-2) optical satellite platforms are able to detect forest degradation in the Congo Basin at what accuracy and cost. We discuss which spatial resolution and revisiting times are required for effective monitoring of forest degradation, and how national estimates of forest degradation could be made within the framework of REDD+.
Figure 1: Illustration of the degradation mapping approach for the Gabon study area of the REDDiness project, with a) Quickbird image of December 2010, b) Quickbird image of March 2012, c) the two levels of segmentation from Quickbird 2010, d) the two levels of segmentation from Quickbird 2012, e) the degradation map showing the percentage difference of small patches of bare soil per level 1 object between 2010 and 2012, f) an overview of the 20x10km study area with the black box indicating the location of the other figures.
Based on results of the two projects, this paper discusses the potential of high and very high resolution imagery to detect forest degradation as a requirement to REDD+. We argue which current and future (such as Sentinel-2) optical satellite platforms are able to detect forest degradation in the Congo Basin at what accuracy and cost. We discuss which spatial resolution and revisiting times are required for effective monitoring of forest degradation, and how national estimates of forest degradation could be made within the framework of REDD+.