Step 2. Segmenting NBR

The purpose of this section is to briefly instruct users who are given an NBR image how to both produce a dNBR image and to then segment the image into low, moderate, and high region.                                       

<<<< Step 1. Calculating NBR                                    

Once you are able to produce a NBR image is is an easy step to produce the so-called dNBR image. Simply select a Landsat image acquired as close as possible to immediately after the fire occurred. Next select two more Landsat images: one acquired one year before when the fire occurred and one acquired one year after the fire occurred.

To calculate dNBR we commonly you calculate NBR for the one year pre-fire image and for on the post-fire images. Recent literature has shown that using a 1-yr post-fire image works better than an immediate at describing the dominant effects that fire has on the vegetation. This presumably because in the immediate post-fire imagery tree mortality by thermally induced girdling would not yet be apparent. 

Once you have calculated NBR for both a pre and post-fire image. DNBR is simply calculated by subtracting the post-fire image values from the pre-fire image values: NBR_pre-NBR_{1-yr post fire}

_{This will produce an output image with values ranging between -2 and +2. It is common for people to then multiple these values by 1000 to produce integer values between -2000 and + 2000.}

 At this point most people then select arbitrary thresholds to determine whether these dNBR values represent low, moderate, or high "severity". Application of such thresholds are at best controversial.

For example, the NBR approach was first developed in Glacier National Park and therefore it does not make sense that any thresholds developed there should work in environments not even similar to GNP. This had led many researchers to develop their own thresholds based on their field-sites. The best approach is to always keep the raw dNBR data (i.e. with no thresholds) to ensure that if regional standards are adopted you can always go back to the original data. If re-scaling data (e.g., to 256 bit), make sure you include ALL the possible values (even if there are none present). i.e. scale from -2000 to +2000 as if you only scale where values are present then its is impossible to re-create the original data.  For example, if in one dNBR image we have values from -256 to +600 and in another we have +100 to +700. Then if we scale each image to 0-255 we would not know what we had originally unless that data was also provided.