![\"model\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F14349659%2F7f8e1ff15c87b96301e346f81ac3003c%2Fmodel2.png?generation=1769785563062926&alt=media\")
We would first like to express our gratitude to the competition host and the Kaggle staff for organizing this outstanding competition.\nBelow, we introduce the my solution
\n| Item | \nDescription | \n
|---|---|
| Model | \nDINOv3 ViT-L/16 distilled (Fine-tuned: regression head + last two transformer blocks) | \n
| Input | \nImages resized to 768×1536 | \n
| Output | \n5 target variables + auxiliary-loss features (Height, NDVI, and whether Species is Ryegrass (Ryegrass_Clover)) | \n
| Loss function | \nHuber loss weighted the same way as the evaluation metric + auxiliary loss | \n
| Preprocessing | \n- Removed 8 samples suspected as label noise after inspection - Standardized the target variables | \n
| CV | \nGrouped by State - Sampling-Date, then manually reviewed and merged similar groups. Split into 5 folds using a greedy method so that each target variable’s distribution is similar across folds. CV and LB seemed correlated, so I tried to improve both. | \n
| Anti-shake measures | \nData augmentation to reproduce unnatural artifacts present in the train set (red-dot artifacts, and a date stamp in the bottom-right corner) | \n
| Ensemble | \nTrained 3 models (with/without auxiliary losses) and averaged their predictions | \n
| Other | \n- EMA - TTA (horizontal flip) | \n
I used DINOv3 ViT-L/16 distilled. Images were resized to 768×1536 and fed into the model. The model outputs consist of the five target variables plus optional auxiliary-loss features, and the final prediction is the average of three models with different output configurations.
\n
| Pattern | \nOutput configuration | \n
|---|---|
| 1 | \n5 target variables | \n
| 2 | \n5 target variables + Height + NDVI | \n
| 3 | \n5 target variables + Height + NDVI + Species (binary classification: whether it is Ryegrass (Ryegrass_Clover)) | \n
For samples with State=WA and Sampling_Date=2015-08-21, the label indicates Dry_Green_g = 0. However, upon inspecting the images, there were clearly plants other than Clover present. I therefore treated these as label errors and removed the following 8 samples from the training data:
\n- ID681680726.jpg\n- ID697718693.jpg\n- ID40849327.jpg\n- ID230058600.jpg\n- ID1403107574.jpg\n- ID1337107565.jpg\n- ID1139918758.jpg\n- ID1761544403.jpg\nThis improved the score by +0.01 on CV, +0.01 on Public, and +0.01 on Private.\n(I corrected it because I had made a mistake.)
\nThe training data contains unnatural artifacts such as the following, so I applied data augmentation to reproduce them:
\nThis did not help on CV or Public, but improved the Private score by +0.01. When I tested them separately, the date/time augmentation did not improve the Public score, so this suggests the private set may contain some red-dot artifacts.
\nIf you have any questions, feel free to ask. \nThank you for reading!
", "messages": [ { "id": "3399501", "postDate": "01/30/2026 23:34:48", "content": "We would first like to express our gratitude to the competition host and the Kaggle staff for organizing this outstanding competition.\nBelow, we introduce the my solution
\n| Item | \nDescription | \n
|---|---|
| Model | \nDINOv3 ViT-L/16 distilled (Fine-tuned: regression head + last two transformer blocks) | \n
| Input | \nImages resized to 768×1536 | \n
| Output | \n5 target variables + auxiliary-loss features (Height, NDVI, and whether Species is Ryegrass (Ryegrass_Clover)) | \n
| Loss function | \nHuber loss weighted the same way as the evaluation metric + auxiliary loss | \n
| Preprocessing | \n- Removed 8 samples suspected as label noise after inspection - Standardized the target variables | \n
| CV | \nGrouped by State - Sampling-Date, then manually reviewed and merged similar groups. Split into 5 folds using a greedy method so that each target variable’s distribution is similar across folds. CV and LB seemed correlated, so I tried to improve both. | \n
| Anti-shake measures | \nData augmentation to reproduce unnatural artifacts present in the train set (red-dot artifacts, and a date stamp in the bottom-right corner) | \n
| Ensemble | \nTrained 3 models (with/without auxiliary losses) and averaged their predictions | \n
| Other | \n- EMA - TTA (horizontal flip) | \n
I used DINOv3 ViT-L/16 distilled. Images were resized to 768×1536 and fed into the model. The model outputs consist of the five target variables plus optional auxiliary-loss features, and the final prediction is the average of three models with different output configurations.
\n
| Pattern | \nOutput configuration | \n
|---|---|
| 1 | \n5 target variables | \n
| 2 | \n5 target variables + Height + NDVI | \n
| 3 | \n5 target variables + Height + NDVI + Species (binary classification: whether it is Ryegrass (Ryegrass_Clover)) | \n
For samples with State=WA and Sampling_Date=2015-08-21, the label indicates Dry_Green_g = 0. However, upon inspecting the images, there were clearly plants other than Clover present. I therefore treated these as label errors and removed the following 8 samples from the training data:
\n- ID681680726.jpg\n- ID697718693.jpg\n- ID40849327.jpg\n- ID230058600.jpg\n- ID1403107574.jpg\n- ID1337107565.jpg\n- ID1139918758.jpg\n- ID1761544403.jpg\nThis improved the score by +0.01 on CV, +0.01 on Public, and +0.01 on Private.\n(I corrected it because I had made a mistake.)
\nThe training data contains unnatural artifacts such as the following, so I applied data augmentation to reproduce them:
\nThis did not help on CV or Public, but improved the Private score by +0.01. When I tested them separately, the date/time augmentation did not improve the Public score, so this suggests the private set may contain some red-dot artifacts.
\nIf you have any questions, feel free to ask. \nThank you for reading!
", "rawMarkdown": "We would first like to express our gratitude to the competition host and the Kaggle staff for organizing this outstanding competition.\nBelow, we introduce the my solution\n\n# Solution\n\n| Item | Description |\n| --- | --- |\n| Model | DINOv3 ViT-L/16 distilled (Fine-tuned: regression head + last two transformer blocks) |\n| Input | Images resized to 768×1536 |\n| Output | 5 target variables + auxiliary-loss features (Height, NDVI, and whether Species is Ryegrass (Ryegrass_Clover)) |\n| Loss function | Huber loss weighted the same way as the evaluation metric + auxiliary loss |\n| Preprocessing | - Removed 8 samples suspected as label noise after inspectionIt was very clear and helpful. Thank you. Especially the model configuration diagram—it's a simple and interesting illustration.
", "rawMarkdown": "It was very clear and helpful. Thank you. Especially the model configuration diagram—it's a simple and interesting illustration.", "votes": null }, { "id": "3399666", "postDate": "01/31/2026 05:59:52", "content": "For samples with State=WA and Sampling_Date=2015-08-21, the label indicates Dry_Green_g = 0. However, upon inspecting the images, there were clearly plants other than Clover present. I therefore treated these as label errors and removed the following 8 samples from the training data:\nwhat was your OOF correlation between the model trained with/without these 8 samples? \nDid the same and for me was 0.99x, extremely similar…
", "rawMarkdown": "```\nFor samples with State=WA and Sampling_Date=2015-08-21, the label indicates Dry_Green_g = 0. However, upon inspecting the images, there were clearly plants other than Clover present. I therefore treated these as label errors and removed the following 8 samples from the training data:\n```\n\nwhat was your OOF correlation between the model trained with/without these 8 samples? \nDid the same and for me was 0.99x, extremely similar...", "votes": null }, { "id": "3399798", "postDate": "01/31/2026 10:25:43", "content": "I’m glad I could help!
", "rawMarkdown": "I’m glad I could help!", "votes": null }, { "id": "3399941", "postDate": "01/31/2026 15:03:30", "content": "The correlation between the predicted Dry_Total_g values from the model trained with the 8 samples and the model trained without them was 0.897. Also, I wrote in my solution that the CV score improved by 0.04, but that was a calculation mistake—the correct improvement was 0.01.
", "rawMarkdown": "The correlation between the predicted Dry_Total_g values from the model trained with the 8 samples and the model trained without them was 0.897. Also, I wrote in my solution that the CV score improved by 0.04, but that was a calculation mistake—the correct improvement was 0.01.", "votes": null }, { "id": "3400006", "postDate": "01/31/2026 17:14:19", "content": "ah makes sense now. thanks and congrats on the gold 🥇
", "rawMarkdown": "ah makes sense now. thanks and congrats on the gold 🥇", "votes": null } ], "comments": [ { "id": 3399572, "author_name": "minfuka", "author_url": "", "post_date": "01/31/2026 01:45:06", "content": "It was very clear and helpful. Thank you. Especially the model configuration diagram—it's a simple and interesting illustration.
", "votes": null, "replies": [ { "id": 3399798, "author_name": "owaowata", "author_url": "", "post_date": "01/31/2026 10:25:43", "content": "I’m glad I could help!
", "votes": null, "replies": [] } ] }, { "id": 3399666, "author_name": "yeoyunsianggeremie", "author_url": "", "post_date": "01/31/2026 05:59:52", "content": "For samples with State=WA and Sampling_Date=2015-08-21, the label indicates Dry_Green_g = 0. However, upon inspecting the images, there were clearly plants other than Clover present. I therefore treated these as label errors and removed the following 8 samples from the training data:\nwhat was your OOF correlation between the model trained with/without these 8 samples? \nDid the same and for me was 0.99x, extremely similar…
", "votes": null, "replies": [ { "id": 3399941, "author_name": "owaowata", "author_url": "", "post_date": "01/31/2026 15:03:30", "content": "The correlation between the predicted Dry_Total_g values from the model trained with the 8 samples and the model trained without them was 0.897. Also, I wrote in my solution that the CV score improved by 0.04, but that was a calculation mistake—the correct improvement was 0.01.
", "votes": null, "replies": [ { "id": 3400006, "author_name": "yeoyunsianggeremie", "author_url": "", "post_date": "01/31/2026 17:14:19", "content": "ah makes sense now. thanks and congrats on the gold 🥇
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3399501": "We would first like to express our gratitude to the competition host and the Kaggle staff for organizing this outstanding competition.\nBelow, we introduce the my solution\n\n# Solution\n\n| Item | Description |\n| --- | --- |\n| Model | DINOv3 ViT-L/16 distilled (Fine-tuned: regression head + last two transformer blocks) |\n| Input | Images resized to 768×1536 |\n| Output | 5 target variables + auxiliary-loss features (Height, NDVI, and whether Species is Ryegrass (Ryegrass_Clover)) |\n| Loss function | Huber loss weighted the same way as the evaluation metric + auxiliary loss |\n| Preprocessing | - Removed 8 samples suspected as label noise after inspection