![\"train/ID230058600.jpg\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1002196%2F4706dce311cf971bc3e1724c482a65ea%2FID230058600.jpg?generation=1763503032471996&alt=media\")
I was spot checking some training data and found some unexpected, potentially wrong annotations. For example, this image has Dry_Clover_g of 46.83 and Dry_Green_g of 0. However, clover is barely visible in the image and there is clearly other green mass in there.
\n
I was spot checking some training data and found some unexpected, potentially wrong annotations. For example, this image has Dry_Clover_g of 46.83 and Dry_Green_g of 0. However, clover is barely visible in the image and there is clearly other green mass in there.
\n
there are also 27 rows that have clover specified in species but Dry_Clover_g set to 0.
\n![\"clover\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1002196%2F204d08c518d098a7e957201a84933ab3%2FScreenshot%202025-11-20%20at%204.53.22PM.png?generation=1763686424389415&alt=media\")
It is interesting. we should drop these sampels
", "rawMarkdown": "It is interesting. we should drop these sampels", "replies": [ { "id": 3360650, "postDate": "2025-12-02T12:18:59.617Z", "content": "And how exactly are you planning to drop them in the test set?\nDon't forget, similar samples are present in the test dataset as well.
", "rawMarkdown": "And how exactly are you planning to drop them in the test set?\nDon't forget, similar samples are present in the test dataset as well." }, { "id": 3360663, "postDate": "2025-12-02T12:36:22.203Z", "content": "In the training set, the \"species\" field is labeled \"clover\". Normally, the \"Dry_Clover_g\" field should have a value, but it doesn't.There are a few explanations. First, the pasture might contain clover, but it's not visible in the images. Second, there could be prediction errors, or the amount of clover is so small that it's approximated as 0.
", "rawMarkdown": "In the training set, the \"species\" field is labeled \"clover\". Normally, the \"Dry_Clover_g\" field should have a value, but it doesn't.There are a few explanations. First, the pasture might contain clover, but it's not visible in the images. Second, there could be prediction errors, or the amount of clover is so small that it's approximated as 0." } ] }, { "id": 3360750, "postDate": "2025-12-02T14:38:13.943Z", "rawMarkdown": "", "isDeleted": true }, { "id": 3362012, "postDate": "2025-12-04T04:52:28.533Z", "content": "It's an interesting thing to notice, so nice catch -- but manually inspecting a few of these images, I get the sense that they are mostly grass or other stuff, with a negligibly small amount of clover. So the biomass measurements are probably fine; the only issue is that the species labels probably shouldn't list clover.
", "rawMarkdown": "It's an interesting thing to notice, so nice catch -- but manually inspecting a few of these images, I get the sense that they are mostly grass or other stuff, with a negligibly small amount of clover. So the biomass measurements are probably fine; the only issue is that the species labels probably shouldn't list clover." }, { "id": 3362189, "postDate": "2025-12-04T09:32:33.627Z", "content": "Has anyone had success with leaving out possibly inconsistent training examples?
", "rawMarkdown": "Has anyone had success with leaving out possibly inconsistent training examples?", "replies": [ { "id": 3362221, "postDate": "2025-12-04T10:24:37.647Z", "content": "Define success… I am nervous about the public/private lb split, but currently masking/removing data improves CV but breaks LB for me.
", "rawMarkdown": "Define success... I am nervous about the public/private lb split, but currently masking/removing data improves CV but breaks LB for me." }, { "id": 3362642, "postDate": "2025-12-05T04:10:27.373Z", "content": "I removed 5 obviously bad images and it didn't do anything for me
", "rawMarkdown": "I removed 5 obviously bad images and it didn't do anything for me" }, { "id": 3362682, "postDate": "2025-12-05T05:58:03.307Z", "content": "I think removing will help the CV alone, but not the LB. In fact, this will make the LB even worse. It's better to leave them as they are so that the model can at least try to give some predictions for them in a reasonable range.
\nAlso, just a heads up, I statistically tried to separate these defective samples from the normal ones based on the predictions of my models and couldn't achieve a good enough F1 score to proceed.
", "rawMarkdown": "I think removing will help the CV alone, but not the LB. In fact, this will make the LB even worse. It's better to leave them as they are so that the model can at least try to give some predictions for them in a reasonable range.\n\nAlso, just a heads up, I statistically tried to separate these defective samples from the normal ones based on the predictions of my models and couldn't achieve a good enough F1 score to proceed.", "votes": 1 } ] } ] }, { "id": 3341081, "postDate": "2025-11-19T22:15:13.877Z", "content": "There are 18 images with the clover fraction, Clover/AllGreen, equal to 1; most are very clover-filled. But in addition to ID230058600 that you discovered, there is another one with little if any clover: ID1403107574 . Two others have patches of clover and also some grass, ID1337107565 and ID1761544403 ; it feels like the grass was ignored compared to the clover.
", "rawMarkdown": "There are 18 images with the clover fraction, Clover/AllGreen, equal to 1; most are very clover-filled. But in addition to ID230058600 that you discovered, there is another one with little if any clover: ID1403107574 . Two others have patches of clover and also some grass, ID1337107565 and ID1761544403 ; it feels like the grass was ignored compared to the clover.", "votes": 3, "replies": [ { "id": 3342333, "postDate": "2025-11-20T20:30:48.667Z", "content": "thanks @dan3dewey, adding ID681680726 to the list
", "rawMarkdown": "thanks @dan3dewey, adding ID681680726 to the list", "votes": 2 } ] }, { "id": 3374350, "postDate": "2025-12-12T17:36:23.053Z", "content": "I have found many samples on which each of my models trained so far are all getting negative R2 scores.
\n(For those who don't know, negative R2 scores is when your model predicts worse than simply giving the average of the targets. It is because when you simply predict the average, you would mathematically get 0 R2 score)
\nThere are multiple approaches I tried to fix these samples. Note that from fixing I mean trying to replace the predictions with mean if I am confident that the predictions is going to lead to a negative R2 score. These techniques mostly depends on finding pattern in the data, in the predictions of my models, even meta data but I couldn't get a reasonable F1 score to proceed.
\nThereafter, I tried two simple things.
\nBoth of these methods decreased both CV and LB.
\nAccording to my calculation, correctly fixing these samples would itself give a boost of 0.03 R2 score on LB, and if you somehow manage to get better than the average on these samples, you can also get 0.05+ R2 score boost on LB.
\nAmong these samples, I found that barely some overlapped with the many samples pointed out by @yogeshkd above, which means that while these samples (pointed out above) are indeed incorrectly labelled, there are many more among the other categories/species linked to other targets.
\nI may/should not ask about the techniques used by others to fix this but I would like to ask to the current toppers if they have found a fix and are using it or not so as to get an estimate of how far am I beyond poor data quality.
", "rawMarkdown": "I have found many samples on which each of my models trained so far are all getting negative R2 scores. \n\n(For those who don't know, negative R2 scores is when your model predicts worse than simply giving the average of the targets. It is because when you simply predict the average, you would mathematically get 0 R2 score)\n\nThere are multiple approaches I tried to fix these samples. Note that from fixing I mean trying to replace the predictions with mean if I am confident that the predictions is going to lead to a negative R2 score. These techniques mostly depends on finding pattern in the data, in the predictions of my models, even meta data but I couldn't get a reasonable F1 score to proceed.\n\nThereafter, I tried two simple things. \n1. Increasing the weight of these samples.\n2. Oversampling these samples. (Don't forget, each time the model is going to see a different augmented version of the oversampled samples)\n\nBoth of these methods decreased both CV and LB. \n\nAccording to my calculation, correctly fixing these samples would itself give a boost of 0.03 R2 score on LB, and if you somehow manage to get better than the average on these samples, you can also get 0.05+ R2 score boost on LB.\n\nAmong these samples, I found that barely some overlapped with the many samples pointed out by @yogeshkd above, which means that while these samples (pointed out above) are indeed incorrectly labelled, there are many more among the other categories/species linked to other targets.\n\nI may/should not ask about the techniques used by others to fix this but I would like to ask to the current toppers if they have found a fix and are using it or not so as to get an estimate of how far am I beyond poor data quality. ", "votes": 2, "replies": [ { "id": 3376149, "postDate": "2025-12-13T18:27:33.263Z", "content": "Nice findings. I have been using data as is to see how far that would get me. Will start experimenting with stuff like this going forward.
", "rawMarkdown": "Nice findings. I have been using data as is to see how far that would get me. Will start experimenting with stuff like this going forward. " } ] }, { "id": 3360839, "postDate": "2025-12-02T16:43:41.487Z", "content": "its looking great!
", "rawMarkdown": "its looking great!", "votes": -5 }, { "id": 3338183, "postDate": "2025-11-18T23:22:41.557Z", "content": "Hi there, Could you provide the file name of this image? Thanks.
", "rawMarkdown": "Hi there, Could you provide the file name of this image? Thanks.", "replies": [ { "id": 3338190, "postDate": "2025-11-18T23:33:50.400Z", "content": "It's train/ID230058600.jpg
I've checked the original data, and it should be an isolated bad case. Thank you for bringing it to our attention.
", "rawMarkdown": "I've checked the original data, and it should be an isolated bad case. Thank you for bringing it to our attention.", "votes": 3 }, { "id": 3342347, "postDate": "2025-11-20T20:56:02.940Z", "content": "@qiyuliaocsiro please have a look at the other comment from @dan3dewey, there is a rather strong common thread between these images and it leave a rather significant gap in the data. This state has zero dead matter in all images, and inconsistence counting of green. Is this representative of the types of error we should expect in the test data?
", "rawMarkdown": "@qiyuliaocsiro please have a look at the other comment from @dan3dewey, there is a rather strong common thread between these images and it leave a rather significant gap in the data. This state has zero dead matter in all images, and inconsistence counting of green. Is this representative of the types of error we should expect in the test data?", "votes": 10 } ] }, { "id": 3360488, "postDate": "2025-12-02T08:39:41.593Z", "content": "Are the plant types listed in the dataset's Species column:\nPresent in the images?\nSome present and some not?\nThe plant types of the entire pasture (not necessarily present in the images)?
", "rawMarkdown": "Are the plant types listed in the dataset's Species column:\nPresent in the images?\nSome present and some not?\nThe plant types of the entire pasture (not necessarily present in the images)?" } ] }, { "id": 3337165, "postDate": "2025-11-18T22:24:14.647Z", "content": "Interesting find! I guess we could remove this datapoint, and hope test doesn't have similar, or try guess a realistic green mass, either way this is the only data we have…
", "rawMarkdown": "Interesting find! I guess we could remove this datapoint, and hope test doesn't have similar, or try guess a realistic green mass, either way this is the only data we have..." } ], "comments": [ { "id": 3342435, "author_name": "yogesh", "author_url": "", "post_date": "2025-11-21T00:54:39.240000", "content": "there are also 27 rows that have clover specified in species but Dry_Clover_g set to 0.
\n
It is interesting. we should drop these sampels
", "votes": 0, "replies": [ { "id": 3360650, "author_name": "Komil Parmar", "author_url": "", "post_date": "2025-12-02T12:18:59.617000", "content": "And how exactly are you planning to drop them in the test set?\nDon't forget, similar samples are present in the test dataset as well.
", "votes": 0, "replies": [] }, { "id": 3360663, "author_name": "", "author_url": "", "post_date": "2025-12-02T12:36:22.203000", "content": "In the training set, the \"species\" field is labeled \"clover\". Normally, the \"Dry_Clover_g\" field should have a value, but it doesn't.There are a few explanations. First, the pasture might contain clover, but it's not visible in the images. Second, there could be prediction errors, or the amount of clover is so small that it's approximated as 0.
", "votes": 0, "replies": [] } ] }, { "id": 3360750, "author_name": "", "author_url": "", "post_date": "2025-12-02T14:38:13.943000", "content": "", "votes": 0, "replies": [] }, { "id": 3362012, "author_name": "particlebbq", "author_url": "", "post_date": "2025-12-04T04:52:28.533000", "content": "It's an interesting thing to notice, so nice catch -- but manually inspecting a few of these images, I get the sense that they are mostly grass or other stuff, with a negligibly small amount of clover. So the biomass measurements are probably fine; the only issue is that the species labels probably shouldn't list clover.
", "votes": 0, "replies": [] }, { "id": 3362189, "author_name": "thomas rost", "author_url": "", "post_date": "2025-12-04T09:32:33.627000", "content": "Has anyone had success with leaving out possibly inconsistent training examples?
", "votes": 0, "replies": [ { "id": 3362221, "author_name": "Tim Krige", "author_url": "", "post_date": "2025-12-04T10:24:37.647000", "content": "Define success… I am nervous about the public/private lb split, but currently masking/removing data improves CV but breaks LB for me.
", "votes": 0, "replies": [] }, { "id": 3362642, "author_name": "yogesh", "author_url": "", "post_date": "2025-12-05T04:10:27.373000", "content": "I removed 5 obviously bad images and it didn't do anything for me
", "votes": 0, "replies": [] }, { "id": 3362682, "author_name": "Komil Parmar", "author_url": "", "post_date": "2025-12-05T05:58:03.307000", "content": "I think removing will help the CV alone, but not the LB. In fact, this will make the LB even worse. It's better to leave them as they are so that the model can at least try to give some predictions for them in a reasonable range.
\nAlso, just a heads up, I statistically tried to separate these defective samples from the normal ones based on the predictions of my models and couldn't achieve a good enough F1 score to proceed.
", "votes": 1, "replies": [] } ] } ] }, { "id": 3341081, "author_name": "Daniel Dewey", "author_url": "", "post_date": "2025-11-19T22:15:13.877000", "content": "There are 18 images with the clover fraction, Clover/AllGreen, equal to 1; most are very clover-filled. But in addition to ID230058600 that you discovered, there is another one with little if any clover: ID1403107574 . Two others have patches of clover and also some grass, ID1337107565 and ID1761544403 ; it feels like the grass was ignored compared to the clover.
", "votes": 3, "replies": [ { "id": 3342333, "author_name": "yogesh", "author_url": "", "post_date": "2025-11-20T20:30:48.667000", "content": "thanks @dan3dewey, adding ID681680726 to the list
", "votes": 2, "replies": [] } ] }, { "id": 3374350, "author_name": "Komil Parmar", "author_url": "", "post_date": "2025-12-12T17:36:23.053000", "content": "I have found many samples on which each of my models trained so far are all getting negative R2 scores.
\n(For those who don't know, negative R2 scores is when your model predicts worse than simply giving the average of the targets. It is because when you simply predict the average, you would mathematically get 0 R2 score)
\nThere are multiple approaches I tried to fix these samples. Note that from fixing I mean trying to replace the predictions with mean if I am confident that the predictions is going to lead to a negative R2 score. These techniques mostly depends on finding pattern in the data, in the predictions of my models, even meta data but I couldn't get a reasonable F1 score to proceed.
\nThereafter, I tried two simple things.
\nBoth of these methods decreased both CV and LB.
\nAccording to my calculation, correctly fixing these samples would itself give a boost of 0.03 R2 score on LB, and if you somehow manage to get better than the average on these samples, you can also get 0.05+ R2 score boost on LB.
\nAmong these samples, I found that barely some overlapped with the many samples pointed out by @yogeshkd above, which means that while these samples (pointed out above) are indeed incorrectly labelled, there are many more among the other categories/species linked to other targets.
\nI may/should not ask about the techniques used by others to fix this but I would like to ask to the current toppers if they have found a fix and are using it or not so as to get an estimate of how far am I beyond poor data quality.
", "votes": 2, "replies": [ { "id": 3376149, "author_name": "Jonas Gjerris", "author_url": "", "post_date": "2025-12-13T18:27:33.263000", "content": "Nice findings. I have been using data as is to see how far that would get me. Will start experimenting with stuff like this going forward.
", "votes": 0, "replies": [] } ] }, { "id": 3360839, "author_name": "Khushi Yadav", "author_url": "", "post_date": "2025-12-02T16:43:41.487000", "content": "its looking great!
", "votes": -5, "replies": [] }, { "id": 3338183, "author_name": "Qiyu Liao", "author_url": "", "post_date": "2025-11-18T23:22:41.557000", "content": "Hi there, Could you provide the file name of this image? Thanks.
", "votes": 0, "replies": [ { "id": 3338190, "author_name": "Yoshio Sugiyama", "author_url": "", "post_date": "2025-11-18T23:33:50.400000", "content": "It's train/ID230058600.jpg
I've checked the original data, and it should be an isolated bad case. Thank you for bringing it to our attention.
", "votes": 3, "replies": [] }, { "id": 3342347, "author_name": "Tim Krige", "author_url": "", "post_date": "2025-11-20T20:56:02.940000", "content": "@qiyuliaocsiro please have a look at the other comment from @dan3dewey, there is a rather strong common thread between these images and it leave a rather significant gap in the data. This state has zero dead matter in all images, and inconsistence counting of green. Is this representative of the types of error we should expect in the test data?
", "votes": 10, "replies": [] } ] }, { "id": 3360488, "author_name": "", "author_url": "", "post_date": "2025-12-02T08:39:41.593000", "content": "Are the plant types listed in the dataset's Species column:\nPresent in the images?\nSome present and some not?\nThe plant types of the entire pasture (not necessarily present in the images)?
", "votes": 0, "replies": [] } ] }, { "id": 3337165, "author_name": "Tim Krige", "author_url": "", "post_date": "2025-11-18T22:24:14.647000", "content": "Interesting find! I guess we could remove this datapoint, and hope test doesn't have similar, or try guess a realistic green mass, either way this is the only data we have…
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3337159": "I was spot checking some training data and found some unexpected, potentially wrong annotations. For example, this image has Dry_Clover_g of 46.83 and Dry_Green_g of 0. However, clover is barely visible in the image and there is clearly other green mass in there.\n\n\n\n\n", "3342435": "there are also 27 rows that have clover specified in species but Dry_Clover_g set to 0.\n\n", "3341081": "There are 18 images with the clover fraction, Clover/AllGreen, equal to 1; most are very clover-filled. But in addition to ID230058600 that you discovered, there is another one with little if any clover: ID1403107574 . Two others have patches of clover and also some grass, ID1337107565 and ID1761544403 ; it feels like the grass was ignored compared to the clover.", "3374350": "I have found many samples on which each of my models trained so far are all getting negative R2 scores. \n\n(For those who don't know, negative R2 scores is when your model predicts worse than simply giving the average of the targets. It is because when you simply predict the average, you would mathematically get 0 R2 score)\n\nThere are multiple approaches I tried to fix these samples. Note that from fixing I mean trying to replace the predictions with mean if I am confident that the predictions is going to lead to a negative R2 score. These techniques mostly depends on finding pattern in the data, in the predictions of my models, even meta data but I couldn't get a reasonable F1 score to proceed.\n\nThereafter, I tried two simple things. \n1. Increasing the weight of these samples.\n2. Oversampling these samples. (Don't forget, each time the model is going to see a different augmented version of the oversampled samples)\n\nBoth of these methods decreased both CV and LB. \n\nAccording to my calculation, correctly fixing these samples would itself give a boost of 0.03 R2 score on LB, and if you somehow manage to get better than the average on these samples, you can also get 0.05+ R2 score boost on LB.\n\nAmong these samples, I found that barely some overlapped with the many samples pointed out by @yogeshkd above, which means that while these samples (pointed out above) are indeed incorrectly labelled, there are many more among the other categories/species linked to other targets.\n\nI may/should not ask about the techniques used by others to fix this but I would like to ask to the current toppers if they have found a fix and are using it or not so as to get an estimate of how far am I beyond poor data quality. ", "3360839": "its looking great!", "3338183": "Hi there, Could you provide the file name of this image? Thanks.", "3337165": "Interesting find! I guess we could remove this datapoint, and hope test doesn't have similar, or try guess a realistic green mass, either way this is the only data we have..." } }