Investigate Differences in On-the-Fly Unit Conversion and Raster Calculations

[siddharth0248]

Summary

While testing the new TiTiler-CMR capabilities, we observed discrepancies between dynamically generated layers and the reference COG products. We'd like to better understand the source of these differences before enabling these features more broadly.

1. On-the-Fly Unit Conversion

TiTiler-CMR can now perform unit conversions directly from NASA CMR. We tested converting MiCASA NPP from kg C/m²/s to g C/m²/day on the fly.

Preview:
https://deploy-preview-858--ghg-demo.netlify.app/exploration?search=&datasets=%5B%7B%22id%22%3A%22micasa-co2-flux-npp-cmr%22%2C%22settings%22%3A%7B%22isVisible%22%3Atrue%2C%22opacity%22%3A100%2C%22analysisMetrics%22%3A%5B%7B%22id%22%3A%22mean%22%2C%22label%22%3A%22Average%22%2C%22chartLabel%22%3A%22Average%22%2C%22themeColor%22%3A%22infographicB%22%7D%5D%2C%22colorMap%22%3A%22purd%22%2C%22reScale%22%3A%7B%22min%22%3A0%2C%22max%22%3A8%7D%2C%22scale%22%3A%7B%22min%22%3A0%2C%22max%22%3A8%7D%7D%7D%2C%7B%22id%22%3A%22micasa-co2-flux-npp-cmr-raw%22%2C%22settings%22%3A%7B%22isVisible%22%3Atrue%2C%22opacity%22%3A100%2C%22analysisMetrics%22%3A%5B%7B%22id%22%3A%22mean%22%2C%22label%22%3A%22Average%22%2C%22chartLabel%22%3A%22Average%22%2C%22themeColor%22%3A%22infographicB%22%7D%5D%2C%22colorMap%22%3A%22purd%22%2C%22reScale%22%3A%7B%22min%22%3A0%2C%22max%22%3A9.25925926e-8%7D%2C%22scale%22%3A%7B%22min%22%3A0%2C%22max%22%3A9.25925926e-8%7D%7D%7D%5D&taxonomy=%7B%7D&date=2024-12-31T06%3A00%3A00.000Z&aois=%5B%5D&dateRange=&dateCompare=

The preview includes:

• (Monthly-On fly) Net Primary Production (NPP) – on-the-fly converted from the CMR source.
• (No conversion - Monthly) Net Primary Production (NPP) – raw CMR values for comparison.

You can also compare these with the existing (Monthly Mean) Net Primary Production (NPP) COG layer currently available in the GHG Center.

Observation

We noticed small differences between the converted layer and the reference data. The differences appear to vary spatially, which is unexpected since the unit conversion is simply multiplication by a constant.

2. Raster Calculations

TiTiler-CMR can also generate derived layers through raster expressions.

Preview:
https://deploy-preview-858--ghg-demo.netlify.app/exploration?search=&datasets=%5B%7B%22id%22%3A%22micasa-co2-flux-nbe-m%22%2C%22settings%22%3A%7B%22isVisible%22%3Atrue%2C%22opacity%22%3A100%2C%22analysisMetrics%22%3A%5B%7B%22id%22%3A%22mean%22%2C%22label%22%3A%22Average%22%2C%22chartLabel%22%3A%22Average%22%2C%22themeColor%22%3A%22infographicB%22%7D%5D%2C%22colorMap%22%3A%22coolwarm%22%2C%22reScale%22%3A%7B%22min%22%3A-4%2C%22max%22%3A4%7D%2C%22scale%22%3A%7B%22min%22%3A-4%2C%22max%22%3A4%7D%2C%22analysisVariableOptions%22%3A%5B%22b1%22%5D%2C%22analysisVariable%22%3A%22b1%22%7D%7D%2C%7B%22id%22%3A%22micasa-co2-flux-nbe1-cmr-raster%22%2C%22settings%22%3A%7B%22isVisible%22%3Atrue%2C%22opacity%22%3A100%2C%22analysisMetrics%22%3A%5B%7B%22id%22%3A%22mean%22%2C%22label%22%3A%22Average%22%2C%22chartLabel%22%3A%22Average%22%2C%22themeColor%22%3A%22infographicB%22%7D%5D%2C%22colorMap%22%3A%22coolwarm%22%2C%22reScale%22%3A%7B%22min%22%3A-4%2C%22max%22%3A4%7D%2C%22scale%22%3A%7B%22min%22%3A-4%2C%22max%22%3A4%7D%2C%22analysisVariableOptions%22%3A%5B%22b1%22%5D%2C%22analysisVariable%22%3A%22b1%22%7D%7D%2C%7B%22id%22%3A%22micasa-co2-flux-nbe-cmr-raster%22%2C%22settings%22%3A%7B%22isVisible%22%3Atrue%2C%22opacity%22%3A100%2C%22analysisMetrics%22%3A%5B%7B%22id%22%3A%22mean%22%2C%22label%22%3A%22Average%22%2C%22chartLabel%22%3A%22Average%22%2C%22themeColor%22%3A%22infographicB%22%7D%5D%2C%22colorMap%22%3A%22coolwarm%22%2C%22reScale%22%3A%7B%22min%22%3A-4%2C%22max%22%3A4%7D%2C%22scale%22%3A%7B%22min%22%3A-4%2C%22max%22%3A4%7D%2C%22analysisVariableOptions%22%3A%5B%22b1%22%5D%2C%22analysisVariable%22%3A%22b1%22%7D%7D%5D&taxonomy=%7B%7D&date=2019-05-28T05%3A00%3A00.000Z&aois=%5B%5D&dateRange=

The preview contains:

• Original NBE COG
• NBE calculated using Rh + FIRE + FUEL − NPP
• NBE calculated using NEE + FIRE + FUEL

The documentation references both equations in different places, so we tested both. The NEE + FIRE + FUEL result appears to more closely match the reference COG, although differences remain (particularly over Central Africa).

Expected behavior

Since:

NEE = Rh − NPP

the following expressions should be mathematically equivalent:

NBE = Rh + FIRE + FUEL − NPP

and

NBE = NEE + FIRE + FUEL

Therefore, we would expect both raster calculations to produce identical outputs. The observed differences suggest either:

• one of the raster expressions may not be evaluated as intended,
• there is an implementation issue, or
• numerical precision is affecting the results.

Questions

• Can we visualize the difference (or residual) between the generated layers and the reference COGs?
• For the unit conversion, why would multiplying by a constant produce spatially varying differences?
• For the raster calculations, are both expressions being evaluated correctly?
• Could these discrepancies simply be due to single- vs. double-precision floating point arithmetic?

Understanding these differences will help determine whether the behavior is expected or if there is an issue in the implementation.

[hrodmn]

While testing the new TiTiler-CMR capabilities, we observed discrepancies between dynamically generated layers and the reference COG products. We'd like to better understand the source of these differences before enabling these features more broadly.

Thanks for writing all of this up @siddharth0248! I have been playing around with the various layer configurations and I am not able to clearly reproduce all of these issues. I'm just leaving some notes here from my initial explorations.

1. On-the-Fly Unit Conversion

I do not any differences between the layer with the constant unit conversion and the raw values! Here are two requests that yield identical images:

No unit conversion (rescale=0,9.25925926e-8: request url for a tile in Central Africa

With unit conversion (rescale=0,8 and expression=86400000*NPP): request url for the same tile as above

2. Raster Calculations

The preview contains:

• Original NBE COG
• NBE calculated using Rh + FIRE + FUEL − NPP
• NBE calculated using NEE + FIRE + FUEL

The documentation references both equations in different places, so we tested both. The NEE + FIRE + FUEL result appears to more closely match the reference COG, although differences remain (particularly over Central Africa).

Expected behavior

Since:

NEE = Rh − NPP

the following expressions should be mathematically equivalent:

NBE = Rh + FIRE + FUEL − NPP

and

NBE = NEE + FIRE + FUEL

Therefore, we would expect both raster calculations to produce identical outputs. The observed differences suggest either:

• one of the raster expressions may not be evaluated as intended,
• there is an implementation issue, or
• numerical precision is affecting the results.

I did a similar test where I requested the same x/y/z tile for all of these layers and from what I can tell the NEE + FIRE + FUEL expression is yielding identical results to the COG while the Rh + FIRE + FUEL − NPP is different.

COG tile (example tile request url)

NEE + FIRE + FUEL tile (example tile request url)

Rh + FIRE + FUEL - NPP tile (example tile request url)

The Rh + FIRE + FUEL - NPP expression does not appear to add up to the same thing as NEE + FIRE + FUEL (based on those tile requests).

Right now I do not suspect any issues in the raster math in titiler-cmr, maybe you are seeing some rendering artifacts when you are browsing in the portal? I have noticed this type of thing when I have multiple layers loaded in your preview links. Some of the layers remain partially loaded at a different resolution which makes them appear slightly different. If it would be easier to go through this together on a call I'm happy to do that sometime!

[siddharth0248]

@hrodmn can we try to pull a tile from this region (WebMercatorQuad/5/16/17) to test NBE calculations:

[hrodmn]

@siddharth0248 I am trying to access the COGs in S3 but I don't have credentials and my titiler request method of downloading a tif is hitting an error. Can you download this file and share it with me in slack? s3://ghgc-data-store/micasa-carbonflux-monthgrid-v1/MiCASA_v1_NBE_x3600_y1800_monthly_201905.tif

[hrodmn]

I worked up a comparison of three NBE sources: an original granule NetCDF file, a COG rendered via titiler from the GHG STAC collection, and TiTiler-CMR. The analysis showed that all three of the sources are very slightly different from one another in some areas, probably due to resampling noise. Based on this analysis I don't think there is anything to be concerned about when rendering MiCASA FLUX data via TiTiler-CMR!

https://gist.github.com/hrodmn/03941d5e94b5c0b0ba084e51333cce2b