Uncertainty assessment
In this merge request, I attempt to derive an uncertainty metric for the emissions. The uncertainty metric consists of:
- the emissions and sinks in the areas whose accuracy is below the accuracy threshold
- the uncertainty of the emission factors.
To address Point 1, I obtain the LULC classification twice: Once with the user-defined threshold and once with a threshold of 0, forcing every pixel to get a LULC value. Then I calculate the emissions and sinks in the areas whose accuracy is below the accuracy threshold and add them as uncertainty margins to the emission values.
To address Point 2, I calculate the emissions for three different scenarios: a high- low- and mean emission scenario. In my sources I found two carbon stocks for most LULC classes (I have two main sources, therefore two different values for most classes). In the high emission scenario, I use a combination of emission factors that leads to the highest possible emissions and in the low emission scenario, I use a combination of emissions factors that leads to the lowest possible emissions. For the mean emission scenario, I use the mean between the higher and the lower carbon stock for each LULC class. The emissions are calculated for the mean emission scenario and then the difference of the emissions to the high- and low emission scenarios is added as uncertainty margin.
The final uncertainty margins are calculated for the total gross emissions, sinks, and net emissions and are the sum of the uncertainty margins from point 1 and point 2. They are given in the summary table. I looked into the calculated uncertainties with different configurations of the LULC utility. Actually, only the fusion_modes "only_model" and "mean_mixin" worked for me. The other modes returned a tiff with a much larger file size, but it was not readable somehow. Mean_mixin achieved lower uncertainties than only_model. With a threshold of 75%, the uncertainties were ridiculously high. With a threshold of 50% the uncertainties were lower, but still around +- 75 to 100 % (which is to be expected I guess).
Probably makes sense to discuss this uncertainty method again in detail with Sven, as I'm not sure that it entirely makes sense.