{
  "id": 615516,
  "title": "How to tackle Dry Dead g target?",
  "url": "/competitions/csiro-biomass/discussion/615516",
  "author_name": "",
  "post_date": "2025-11-10T19:44:53.785895700Z",
  "votes": 14,
  "comment_count": 6,
  "views": 0,
  "content": "<p>My cross validation score across targets -</p>\n<p>Overall R2: 0.6721</p>\n<ul>\n<li>R2 Dry_Total_g: 0.6847</li>\n<li>R2 Dry_Green_g: 0.7753</li>\n<li>R2 Dry_Dead_g: 0.4448</li>\n<li>R2 Dry_Clover_g: 0.7086</li>\n<li>R2 GDM_g: 0.7468</li>\n</ul>\n<p>As you can see Dry Dead g is the poorest fit. I tried utilising the Dead = Total - GDM relationship but still couldn't cross 0.45</p>\n<p>I tried training my model with only 3 targets and then determining the other 2, but to no improvement :(</p>\n<p>I would appreciate any suggestions.</p>",
  "messages": [
    {
      "id": "3315832",
      "postDate": "11/10/2025 19:44:53",
      "content": "<p>My cross validation score across targets -</p>\n<p>Overall R2: 0.6721</p>\n<ul>\n<li>R2 Dry_Total_g: 0.6847</li>\n<li>R2 Dry_Green_g: 0.7753</li>\n<li>R2 Dry_Dead_g: 0.4448</li>\n<li>R2 Dry_Clover_g: 0.7086</li>\n<li>R2 GDM_g: 0.7468</li>\n</ul>\n<p>As you can see Dry Dead g is the poorest fit. I tried utilising the Dead = Total - GDM relationship but still couldn't cross 0.45</p>\n<p>I tried training my model with only 3 targets and then determining the other 2, but to no improvement :(</p>\n<p>I would appreciate any suggestions.</p>",
      "rawMarkdown": "My cross validation score across targets -\n\nOverall R2: 0.6721\n\n- R2 Dry_Total_g: 0.6847\n- R2 Dry_Green_g: 0.7753\n- R2 Dry_Dead_g: 0.4448\n- R2 Dry_Clover_g: 0.7086\n- R2 GDM_g: 0.7468\n\nAs you can see Dry Dead g is the poorest fit. I tried utilising the Dead = Total - GDM relationship but still couldn't cross 0.45\n\nI tried training my model with only 3 targets and then determining the other 2, but to no improvement :(\n\nI would appreciate any suggestions.",
      "votes": null
    },
    {
      "id": "3317973",
      "postDate": "11/10/2025 23:26:34",
      "content": "<p>Good question! </p>\n<p>Some dead matter hidden below may be invisible from the surface, which could make it difficult for the prediction to align with the ground truth.</p>\n<p>Perhaps we could consider using the metadata <strong>Height_Ave_cm</strong>? </p>\n<p>(This is just my personal opinion and has not been validated.)</p>",
      "rawMarkdown": "Good question! \n\nSome dead matter hidden below may be invisible from the surface, which could make it difficult for the prediction to align with the ground truth.\n\nPerhaps we could consider using the metadata **Height_Ave_cm**? \n\n(This is just my personal opinion and has not been validated.)",
      "votes": null
    },
    {
      "id": "3320372",
      "postDate": "11/12/2025 13:32:38",
      "content": "<p>hey!\nwhat did you even do to reach 0.45? I'm stuck at 0.22 for days and can't surpass, which is demoralizing…</p>",
      "rawMarkdown": "hey!\nwhat did you even do to reach 0.45? I'm stuck at 0.22 for days and can't surpass, which is demoralizing...",
      "votes": null
    },
    {
      "id": "3320377",
      "postDate": "11/12/2025 13:43:13",
      "content": "<p><a href=\"https://www.kaggle.com/code/paharpurailab/csiro-eva02-modified-targets-infer-lb-0-63\" target=\"_blank\">This notebook</a> has a very similar approach.</p>",
      "rawMarkdown": "[This notebook](https://www.kaggle.com/code/paharpurailab/csiro-eva02-modified-targets-infer-lb-0-63) has a very similar approach.",
      "votes": null
    },
    {
      "id": "3322602",
      "postDate": "11/13/2025 15:46:20",
      "content": "<p>Keep at it! a lot of good notebooks in the code section for some good ideas</p>",
      "rawMarkdown": "Keep at it! a lot of good notebooks in the code section for some good ideas",
      "votes": null
    },
    {
      "id": "3350829",
      "postDate": "11/27/2025 20:44:53",
      "content": "<p>But <strong>Height_Ave_cm</strong> are not available for test dataset, correct? </p>",
      "rawMarkdown": "But **Height_Ave_cm** are not available for test dataset, correct?",
      "votes": null
    },
    {
      "id": "3351455",
      "postDate": "11/28/2025 12:18:29",
      "content": "<p>Yeah exactly. I tried to use it as an auxiliary task and then using those predicted height to predict the Dry_dead, but that had a lot of noise, variance in it. Simply deriving it from the Dead = Total - GDM worked much better.</p>\n<p>I guess the only solution now is to train a MIM model on the dataset. </p>",
      "rawMarkdown": "Yeah exactly. I tried to use it as an auxiliary task and then using those predicted height to predict the Dry_dead, but that had a lot of noise, variance in it. Simply deriving it from the Dead = Total - GDM worked much better.\n\nI guess the only solution now is to train a MIM model on the dataset.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3317973,
      "author_name": "qiyuliaocsiro",
      "author_url": "",
      "post_date": "11/10/2025 23:26:34",
      "content": "<p>Good question! </p>\n<p>Some dead matter hidden below may be invisible from the surface, which could make it difficult for the prediction to align with the ground truth.</p>\n<p>Perhaps we could consider using the metadata <strong>Height_Ave_cm</strong>? </p>\n<p>(This is just my personal opinion and has not been validated.)</p>",
      "votes": null,
      "replies": [
        {
          "id": 3350829,
          "author_name": "jirkaborovec",
          "author_url": "",
          "post_date": "11/27/2025 20:44:53",
          "content": "<p>But <strong>Height_Ave_cm</strong> are not available for test dataset, correct? </p>",
          "votes": null,
          "replies": [
            {
              "id": 3351455,
              "author_name": "komilparmar",
              "author_url": "",
              "post_date": "11/28/2025 12:18:29",
              "content": "<p>Yeah exactly. I tried to use it as an auxiliary task and then using those predicted height to predict the Dry_dead, but that had a lot of noise, variance in it. Simply deriving it from the Dead = Total - GDM worked much better.</p>\n<p>I guess the only solution now is to train a MIM model on the dataset. </p>",
              "votes": null,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 3320372,
      "author_name": "romanmorozov",
      "author_url": "",
      "post_date": "11/12/2025 13:32:38",
      "content": "<p>hey!\nwhat did you even do to reach 0.45? I'm stuck at 0.22 for days and can't surpass, which is demoralizing…</p>",
      "votes": null,
      "replies": [
        {
          "id": 3320377,
          "author_name": "paritoshtripathi5",
          "author_url": "",
          "post_date": "11/12/2025 13:43:13",
          "content": "<p><a href=\"https://www.kaggle.com/code/paharpurailab/csiro-eva02-modified-targets-infer-lb-0-63\" target=\"_blank\">This notebook</a> has a very similar approach.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 3322602,
          "author_name": "evnsnclr",
          "author_url": "",
          "post_date": "11/13/2025 15:46:20",
          "content": "<p>Keep at it! a lot of good notebooks in the code section for some good ideas</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3315832": "My cross validation score across targets -\n\nOverall R2: 0.6721\n\n- R2 Dry_Total_g: 0.6847\n- R2 Dry_Green_g: 0.7753\n- R2 Dry_Dead_g: 0.4448\n- R2 Dry_Clover_g: 0.7086\n- R2 GDM_g: 0.7468\n\nAs you can see Dry Dead g is the poorest fit. I tried utilising the Dead = Total - GDM relationship but still couldn't cross 0.45\n\nI tried training my model with only 3 targets and then determining the other 2, but to no improvement :(\n\nI would appreciate any suggestions.",
    "3317973": "Good question! \n\nSome dead matter hidden below may be invisible from the surface, which could make it difficult for the prediction to align with the ground truth.\n\nPerhaps we could consider using the metadata **Height_Ave_cm**? \n\n(This is just my personal opinion and has not been validated.)",
    "3320372": "hey!\nwhat did you even do to reach 0.45? I'm stuck at 0.22 for days and can't surpass, which is demoralizing...",
    "3320377": "[This notebook](https://www.kaggle.com/code/paharpurailab/csiro-eva02-modified-targets-infer-lb-0-63) has a very similar approach.",
    "3322602": "Keep at it! a lot of good notebooks in the code section for some good ideas",
    "3350829": "But **Height_Ave_cm** are not available for test dataset, correct?",
    "3351455": "Yeah exactly. I tried to use it as an auxiliary task and then using those predicted height to predict the Dry_dead, but that had a lot of noise, variance in it. Simply deriving it from the Dead = Total - GDM worked much better.\n\nI guess the only solution now is to train a MIM model on the dataset."
  },
  "source": "meta"
}