{
  "id": 648906,
  "title": "How do you effectively utilize metadata (NDVI, height)?",
  "url": "/competitions/csiro-biomass/discussion/648906",
  "author_name": "",
  "post_date": "2025-12-01T14:34:01.019251800Z",
  "votes": 2,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Hi everyone,</p>\n<p>I am currently experimenting with ways to incorporate metadata, specifically NDVI and height, into my models.</p>\n<p>So far, I have tried the following two approaches:</p>\n<p>Multi-task learning: Adding an auxiliary loss to a single model to predict metadata alongside the main target.</p>\n<p>Two-stage modeling: First, training a model to predict metadata from images, and then using the predicted metadata as input features for the final prediction model.</p>\n<p>Unfortunately, neither approach has worked for me yet. In fact, adding these methods made my scores worse compared to the baseline.</p>\n<p>I’m curious to know how everyone else is utilizing the metadata. Do you have any suggestions or successful strategies?</p>\n<p>Thanks!</p>",
  "messages": [
    {
      "id": "3358329",
      "postDate": "12/01/2025 14:34:01",
      "content": "<p>Hi everyone,</p>\n<p>I am currently experimenting with ways to incorporate metadata, specifically NDVI and height, into my models.</p>\n<p>So far, I have tried the following two approaches:</p>\n<p>Multi-task learning: Adding an auxiliary loss to a single model to predict metadata alongside the main target.</p>\n<p>Two-stage modeling: First, training a model to predict metadata from images, and then using the predicted metadata as input features for the final prediction model.</p>\n<p>Unfortunately, neither approach has worked for me yet. In fact, adding these methods made my scores worse compared to the baseline.</p>\n<p>I’m curious to know how everyone else is utilizing the metadata. Do you have any suggestions or successful strategies?</p>\n<p>Thanks!</p>",
      "rawMarkdown": "Hi everyone,\n\nI am currently experimenting with ways to incorporate metadata, specifically NDVI and height, into my models.\n\nSo far, I have tried the following two approaches:\n\nMulti-task learning: Adding an auxiliary loss to a single model to predict metadata alongside the main target.\n\nTwo-stage modeling: First, training a model to predict metadata from images, and then using the predicted metadata as input features for the final prediction model.\n\nUnfortunately, neither approach has worked for me yet. In fact, adding these methods made my scores worse compared to the baseline.\n\nI’m curious to know how everyone else is utilizing the metadata. Do you have any suggestions or successful strategies?\n\nThanks!",
      "votes": null
    },
    {
      "id": "3360264",
      "postDate": "12/02/2025 02:33:06",
      "content": "<p>I feel that most people only use target vectors from images and metadata to construct strategies. There is no metadata in the public ranking list, so the feature vectors of metadata have no value (of course, there may be, I am not sure). You said 'training a model to predict metadata from images', which would treat all feature vectors as target vectors for prediction. How can this be achieved? Will increase the difficulty of the task</p>",
      "rawMarkdown": "I feel that most people only use target vectors from images and metadata to construct strategies. There is no metadata in the public ranking list, so the feature vectors of metadata have no value (of course, there may be, I am not sure). You said 'training a model to predict metadata from images', which would treat all feature vectors as target vectors for prediction. How can this be achieved? Will increase the difficulty of the task",
      "votes": null
    },
    {
      "id": "3360467",
      "postDate": "12/02/2025 07:55:47",
      "content": "<p>Inspired by the discussion at (<a href=\"https://www.kaggle.com/competitions/csiro-biomass/discussion/627302)\" target=\"_blank\">https://www.kaggle.com/competitions/csiro-biomass/discussion/627302)</a>, I built a model to predict metadata so I could use it during inference. However, so far it seems like it's just making the task more complicated.</p>",
      "rawMarkdown": "Inspired by the discussion at (https://www.kaggle.com/competitions/csiro-biomass/discussion/627302), I built a model to predict metadata so I could use it during inference. However, so far it seems like it's just making the task more complicated.",
      "votes": null
    },
    {
      "id": "3361280",
      "postDate": "12/03/2025 08:23:15",
      "content": "<p>You might wanna try confidence based inference.</p>\n<p>First, please check the gradients going to your predicted meta data or some other technique like Shapley to know how much your model depends on your predicted meta data. Is it is reasonably high than image features or comparable to importance of image features then, </p>\n<p>Train two models.</p>\n<ol>\n<li>Dependent only on image features</li>\n<li>Uses your intermediate meta data predictions.</li>\n</ol>\n<p>If the confidence of the predicted meta data is high, use 2. else 1. This is because the error propagates. If the predicted meta data is wrong, the final target predictions would be even more wrong. \n(This is just a suggestion given that you really wish to incorporate metadata, not a validated idea.)</p>",
      "rawMarkdown": "You might wanna try confidence based inference.\n\nFirst, please check the gradients going to your predicted meta data or some other technique like Shapley to know how much your model depends on your predicted meta data. Is it is reasonably high than image features or comparable to importance of image features then, \n\nTrain two models.\n1. Dependent only on image features\n2. Uses your intermediate meta data predictions.\n\nIf the confidence of the predicted meta data is high, use 2. else 1. This is because the error propagates. If the predicted meta data is wrong, the final target predictions would be even more wrong. \n(This is just a suggestion given that you really wish to incorporate metadata, not a validated idea.)",
      "votes": null
    },
    {
      "id": "3361877",
      "postDate": "12/03/2025 23:20:00",
      "content": "<p>Thanks! I'll give confidence-based inference a try.</p>",
      "rawMarkdown": "Thanks! I'll give confidence-based inference a try.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3360264,
      "author_name": "koukoumi",
      "author_url": "",
      "post_date": "12/02/2025 02:33:06",
      "content": "<p>I feel that most people only use target vectors from images and metadata to construct strategies. There is no metadata in the public ranking list, so the feature vectors of metadata have no value (of course, there may be, I am not sure). You said 'training a model to predict metadata from images', which would treat all feature vectors as target vectors for prediction. How can this be achieved? Will increase the difficulty of the task</p>",
      "votes": null,
      "replies": [
        {
          "id": 3360467,
          "author_name": "taka3776toshi",
          "author_url": "",
          "post_date": "12/02/2025 07:55:47",
          "content": "<p>Inspired by the discussion at (<a href=\"https://www.kaggle.com/competitions/csiro-biomass/discussion/627302)\" target=\"_blank\">https://www.kaggle.com/competitions/csiro-biomass/discussion/627302)</a>, I built a model to predict metadata so I could use it during inference. However, so far it seems like it's just making the task more complicated.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 3361280,
      "author_name": "komilparmar",
      "author_url": "",
      "post_date": "12/03/2025 08:23:15",
      "content": "<p>You might wanna try confidence based inference.</p>\n<p>First, please check the gradients going to your predicted meta data or some other technique like Shapley to know how much your model depends on your predicted meta data. Is it is reasonably high than image features or comparable to importance of image features then, </p>\n<p>Train two models.</p>\n<ol>\n<li>Dependent only on image features</li>\n<li>Uses your intermediate meta data predictions.</li>\n</ol>\n<p>If the confidence of the predicted meta data is high, use 2. else 1. This is because the error propagates. If the predicted meta data is wrong, the final target predictions would be even more wrong. \n(This is just a suggestion given that you really wish to incorporate metadata, not a validated idea.)</p>",
      "votes": null,
      "replies": [
        {
          "id": 3361877,
          "author_name": "taka3776toshi",
          "author_url": "",
          "post_date": "12/03/2025 23:20:00",
          "content": "<p>Thanks! I'll give confidence-based inference a try.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3358329": "Hi everyone,\n\nI am currently experimenting with ways to incorporate metadata, specifically NDVI and height, into my models.\n\nSo far, I have tried the following two approaches:\n\nMulti-task learning: Adding an auxiliary loss to a single model to predict metadata alongside the main target.\n\nTwo-stage modeling: First, training a model to predict metadata from images, and then using the predicted metadata as input features for the final prediction model.\n\nUnfortunately, neither approach has worked for me yet. In fact, adding these methods made my scores worse compared to the baseline.\n\nI’m curious to know how everyone else is utilizing the metadata. Do you have any suggestions or successful strategies?\n\nThanks!",
    "3360264": "I feel that most people only use target vectors from images and metadata to construct strategies. There is no metadata in the public ranking list, so the feature vectors of metadata have no value (of course, there may be, I am not sure). You said 'training a model to predict metadata from images', which would treat all feature vectors as target vectors for prediction. How can this be achieved? Will increase the difficulty of the task",
    "3360467": "Inspired by the discussion at (https://www.kaggle.com/competitions/csiro-biomass/discussion/627302), I built a model to predict metadata so I could use it during inference. However, so far it seems like it's just making the task more complicated.",
    "3361280": "You might wanna try confidence based inference.\n\nFirst, please check the gradients going to your predicted meta data or some other technique like Shapley to know how much your model depends on your predicted meta data. Is it is reasonably high than image features or comparable to importance of image features then, \n\nTrain two models.\n1. Dependent only on image features\n2. Uses your intermediate meta data predictions.\n\nIf the confidence of the predicted meta data is high, use 2. else 1. This is because the error propagates. If the predicted meta data is wrong, the final target predictions would be even more wrong. \n(This is just a suggestion given that you really wish to incorporate metadata, not a validated idea.)",
    "3361877": "Thanks! I'll give confidence-based inference a try."
  },
  "source": "meta"
}