I am running the InVEST NDR model (3.13.0) on a global scale. I want to understand why there are many “empty pixels” in my output file (n_surface_export.tif), especially in island regions. “empty pixel” basically means that there is no data at the edges of land masses, specifically in islands. Compared to Chaplin and Kramer et al 2019 that also used the InVEST NDR model on a global scale, it has less “empty pixels”. My hope is to know what ate the levers that I can pull to reduce the number of “empty pixels” in my output file.
I assume it has something to do with my choice of watershed. The watershed that I am using was not deliniated from the DEM file but based upon it. Nevertheless, there are different Pfafstetter watershed levels that indicate its scale. My guess is, I need to choose a higher level (to give a more detailed watershed shape) to reduce these empty pixels. Nevertheless, re-running the InVEST NDR model on a global scale would take time so I wanted to ask in the forum first to check if my idea makes sense or not.
My other guess is based on the idea that empty pixels basically mean that the model thinks that the sediment or nutrient has already reached a water body (i.e., ocean) and thus it stops calculating. However, based on my LULC map, the area where there are no pixels is not the ocean. Therefore, its a bit mind bobbling here.
Supporting files or information that elaborate upon what was previously mentioned?
Example of Pfafstetter Watershed Level 04 (currently used)
This happens a lot along the coast. And it’s often explained by how the Threshold Flow Accumulation value affects stream creation. Check out the User Guide Defined Area of Inputs section for NDR (and also read the additional information linked to in the SDR chapter) and see if that helps.
After reading the documentation in more detail, and conducting some tests, I have come to the following conclusion:
The granular your stream network is, the less no data pixels that you have. InVEST would not calculate the SDR because the stream is the “end-goal” of the sediment flow, and without a stream there is no “end-goal”. Hence, there are no pixels.
This means, to reduce the number of no pixels, which especially is relevant in island countries such as Indonesia, Iceland, or Sri Lanka, one would need to reduce the TFA number.
In my study, I reduced the TFA from 5000 to 500 and this improved the output data for island countries. This however comes as a trade-off in other parts of the globe where one has relatively low elevation, for example in the Australian outback or Xinjiang province in China. By reducing the TFA, one obtains streams that are wider then they are supposed to be.
Additionally, I tried to change the DEM with something that had a better spatial resolution but this did not help as well. One still saw these widened “streams”, especially in desert areas.
In this regard, I see two ways of going forward:
First Option
Perform a “join” of two raster files. This means, run the NDR model on two TFAs (5000 and 500). Identify where on the globe TFA 500 does not overlap TFA 5000 then perform a join of the two resulting in a raster map that is based on TFA 5000 and 500.
The advantage here is that you have significantly less “empty” pixels globally. However, the scientific logic behind the concept might be lacking, because in principle, this means you assume two different river maps and they might not perfectly join with one another.
Second Option
For island countries, use the results obtained from an SDR run with the TFA500. Meanwhile, for non island countries, use the results obtained from an SDR run with the TFA5000. This approach is in principle similar to a mosaic raster, in which one “mosaics” two raster from TFA500 and TFA5000.
The advantage here is that it has more scientific basis, because like any other spatially explicit study, you can “combine” different regions of the globe due to unequal data quality in the regions. However, it remains to be answered, how exactly should we devide the globe. By country? by continent? This is open.
What are your thoughts on my ideas? Have you seen perhaps a similar case where people try to combine the results of two different InVEST SDR runs? I attached some pictures to give a better understanding of the points that I mentioned here. TFA Issue.zip (583.8 KB)
