Landsat Temporal Composite overlapping paths very different pixel values Google Earth Engine

Question

I use GEE to create temporal composites from Landsat imagery (right now using Landsat 8 Collection 2, but same issue with Collection 1 or Landsat 7) for a particular island. I noticed that parts of my composite were getting weird pixel values. When I investigated further, I found that one of my islands of interest is at the intersection of overlapping Landsat paths (WRS Path 83 & 84) where one path (84) only overlaps at the bottom corner of the path/row and only covers about half the island. Now path 83 covers the entire island. If I create a median temporal composite (find all images over a 4 year period, remove all clouds, then use a reducer (50th percentile) on the image collection) on images from both paths (aka all available images that might overlap my area of interest), I get compositeAll which looks a bit weird in the north western corner of the island. If I create a temporal composite ONLY using path 83 or path 84 (composite83 or composite84) I get a very different image. Here is the composite 84:

while for composite83:

and for all images compositeAll . All the above images are visualized with the same parameters to highlight the different pixel values present in the two paths. Any idea how to correct this other than manually only using one path? I would rather be able to use all available images for a given location...but clearly my code is not correcting the images fully?

Here is a chart showing the mean pixel values in an area of the north western side of the lagoon of the island for all three composites for bands 1-8. you can see that the pixel values for the composite from Path 84 is much higher than composite 83 and is not the same across the different bands. This means that the composite using both path 83 & 84 gets weird looking pixel values. I'm attaching the code. I've

https://code.earthengine.google.com/?scriptPath=users%2Facortiz%2FGE254%3Aload_singleimages

Answer 1

There could be several reasons that the intersection of the paths are spectrally offset.

1. Bidirectional reflection distribution function (BRDF). The illumination angle difference is relatively great at path intersection (lower angle for east scene, higher angle for west scene).
2. Differences in atmosphere (are you using surface reflectance?). Minor haze differences between observation dates, atmosphere path length differences (longer distance for east scene, shorter for west scene).
3. Observation timing. How far apart are the east and west scene observation dates? Could phenology differences influence the signal? Sun angle changes throughout the year which affects reflectance.

You could try to normalize one image to another. If they are close in date, simple linear regression would probably suffice - here is an example.