![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F349155%2Fb10c5ad284a86b86546bda40e6e41adc%2FSakvaChart1.png?generation=1753433849378630&alt=media\")
Hi everyone,
\nI came across planet 2486733311, which shows an unusual discrepancy between the FGS and AIRS data. In the plot, you can see the time-averaged FGS channel (in orange) and the mean of the AIRS channel (in blue), both normalized by the maximum value of their respective time series - so both curves peak at 1.
\nTypically, the difference between these two curves is quite small, but that’s not the case here.
\nMy model predicts 0.00123 vs the true value of 0.00897 for FGS, and 0.00951 vs 0.00903 for the mean AIRS value - so the error is much smaller for AIRS.
\nWhat do you think might be going on here?
Hi everyone,
\nI came across planet 2486733311, which shows an unusual discrepancy between the FGS and AIRS data. In the plot, you can see the time-averaged FGS channel (in orange) and the mean of the AIRS channel (in blue), both normalized by the maximum value of their respective time series - so both curves peak at 1.
\nTypically, the difference between these two curves is quite small, but that’s not the case here.
\nMy model predicts 0.00123 vs the true value of 0.00897 for FGS, and 0.00951 vs 0.00903 for the mean AIRS value - so the error is much smaller for AIRS.
\nWhat do you think might be going on here?
Hi @ilu000 and @sakvaua , sorry for the late reply. Yes we can confirm the FGS 1 channel here is faulty. We have checked the rest of the dataset and this is the only case. We will suggest two actions : 1. completely discard this particular sample or 2. use AIRS information to recover / approximate FGS1. Hope this is helpful!
", "rawMarkdown": "Hi @ilu000 and @sakvaua , sorry for the late reply. Yes we can confirm the FGS 1 channel here is faulty. We have checked the rest of the dataset and this is the only case. We will suggest two actions : 1. completely discard this particular sample or 2. use AIRS information to recover / approximate FGS1. Hope this is helpful! ", "votes": 5, "replies": [ { "id": 3269328, "postDate": "2025-08-14T09:46:49.917Z", "content": "Thank you @gordonyip. Confirming that this is the only sample that has this issue (I assume including the test set) is exactly what we need to know.
\nExceptional great communication from your side. Kudos!
Yes including the test set :)
", "rawMarkdown": "Yes including the test set :)", "votes": 2 } ] }, { "id": 3270713, "postDate": "2025-08-17T07:16:44.753Z", "content": "May I ask what was causing this problem? Just out of curiosity.
", "rawMarkdown": "May I ask what was causing this problem? Just out of curiosity.", "replies": [ { "id": 3273148, "postDate": "2025-08-22T08:23:43.723Z", "content": "I wish i could give you a full answer on this - but this is a very odd and rare case we (well you guys) discovered and it didnt happen anywhere else. If I have to make a guess - we believe it could be an copy-and-paste issue.
", "rawMarkdown": "I wish i could give you a full answer on this - but this is a very odd and rare case we (well you guys) discovered and it didnt happen anywhere else. If I have to make a guess - we believe it could be an copy-and-paste issue.", "votes": 1 } ] } ] }, { "id": 3266879, "postDate": "2025-08-09T21:03:25.800Z", "content": "It honestly looks like a bug to me. I can't think of a valid explanation why the target would be that large for the FGS wavelength.
\nIf indeed this is an invalid data sample in train, we can only hope such samples don't happen in test.
\n@gordonyip
Hi @ilu000 and @sakvaua , sorry for the late reply. Yes we can confirm the FGS 1 channel here is faulty. We have checked the rest of the dataset and this is the only case. We will suggest two actions : 1. completely discard this particular sample or 2. use AIRS information to recover / approximate FGS1. Hope this is helpful!
", "votes": 5, "replies": [ { "id": 3269328, "author_name": "Pascal Pfeiffer", "author_url": "", "post_date": "2025-08-14T09:46:49.917000", "content": "Thank you @gordonyip. Confirming that this is the only sample that has this issue (I assume including the test set) is exactly what we need to know.
\nExceptional great communication from your side. Kudos!
Yes including the test set :)
", "votes": 2, "replies": [] } ] }, { "id": 3270713, "author_name": "c-number", "author_url": "", "post_date": "2025-08-17T07:16:44.753000", "content": "May I ask what was causing this problem? Just out of curiosity.
", "votes": 0, "replies": [ { "id": 3273148, "author_name": "Gordon Yip", "author_url": "", "post_date": "2025-08-22T08:23:43.723000", "content": "I wish i could give you a full answer on this - but this is a very odd and rare case we (well you guys) discovered and it didnt happen anywhere else. If I have to make a guess - we believe it could be an copy-and-paste issue.
", "votes": 1, "replies": [] } ] } ] }, { "id": 3266879, "author_name": "Pascal Pfeiffer", "author_url": "", "post_date": "2025-08-09T21:03:25.800000", "content": "It honestly looks like a bug to me. I can't think of a valid explanation why the target would be that large for the FGS wavelength.
\nIf indeed this is an invalid data sample in train, we can only hope such samples don't happen in test.
\n@gordonyip