{
  "id": 614756,
  "title": "Meta data use in test set",
  "url": "/competitions/csiro-biomass/discussion/614756",
  "author_name": "",
  "post_date": "2025-11-06T08:21:53.251138600Z",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hi all:\nIs there anyone using input variables other than the image for prediction on the test set? I assume we only have the image_id for the test data and cannot use the extra variable, such as \"Species\", \"Pre_GSHH_NDVI\", \"Height_Ave_cm\", in the model.\nCorrect me if I have missed anything here.</p>\n<p>Best Tej</p>",
  "messages": [
    {
      "id": "3312021",
      "postDate": "11/06/2025 08:21:53",
      "content": "<p>Hi all:\nIs there anyone using input variables other than the image for prediction on the test set? I assume we only have the image_id for the test data and cannot use the extra variable, such as \"Species\", \"Pre_GSHH_NDVI\", \"Height_Ave_cm\", in the model.\nCorrect me if I have missed anything here.</p>\n<p>Best Tej</p>",
      "rawMarkdown": "Hi all:\nIs there anyone using input variables other than the image for prediction on the test set? I assume we only have the image_id for the test data and cannot use the extra variable, such as \"Species\", \"Pre_GSHH_NDVI\", \"Height_Ave_cm\", in the model.\nCorrect me if I have missed anything here.\n\nBest Tej",
      "votes": null
    },
    {
      "id": "3312081",
      "postDate": "11/06/2025 12:07:45",
      "content": "<p>As we only have a test image and no metadata for it, we really have to  predict based only on the test image. In training we do have additional information, though based on the current status of the discussion board, there do seem to be some unhelpful fields in this.</p>",
      "rawMarkdown": "As we only have a test image and no metadata for it, we really have to  predict based only on the test image. In training we do have additional information, though based on the current status of the discussion board, there do seem to be some unhelpful fields in this.",
      "votes": null
    },
    {
      "id": "3382619",
      "postDate": "12/28/2025 04:58:05",
      "content": "<p>Hi Tej,</p>\n<p>You’re correct that the test set only provides the images, so features like Species, Pre_GSHH_NDVI, or Height_Ave_cm aren’t directly available. However, one strategy that some participants use is a two-stage approach:</p>\n<p>Predict the metadata from the images</p>\n<p>Train small regression/classification models on the training set to predict Pre_GSHH_NDVI, Height_Ave_cm, and Species using the images as input.</p>\n<p>This gives “predicted metadata” for any new image.</p>\n<p>Use the predicted metadata along with the image for biomass prediction</p>\n<p>Combine image features (from a CNN, ViT, or pretrained model like DINOv2) with the predicted metadata as inputs to your final biomass regression model.</p>\n<p>This allows the model to leverage the information from the metadata without having the real values at test time.</p>\n<p>This two-stage pipeline is quite common when metadata is available in training but missing in the test set. It helps the model focus on relevant signals like pasture height or species composition while still being fully test-set compatible.</p>\n<p>Hope this helps!</p>",
      "rawMarkdown": "Hi Tej,\n\nYou’re correct that the test set only provides the images, so features like Species, Pre_GSHH_NDVI, or Height_Ave_cm aren’t directly available. However, one strategy that some participants use is a two-stage approach:\n\nPredict the metadata from the images\n\nTrain small regression/classification models on the training set to predict Pre_GSHH_NDVI, Height_Ave_cm, and Species using the images as input.\n\nThis gives “predicted metadata” for any new image.\n\nUse the predicted metadata along with the image for biomass prediction\n\nCombine image features (from a CNN, ViT, or pretrained model like DINOv2) with the predicted metadata as inputs to your final biomass regression model.\n\nThis allows the model to leverage the information from the metadata without having the real values at test time.\n\nThis two-stage pipeline is quite common when metadata is available in training but missing in the test set. It helps the model focus on relevant signals like pasture height or species composition while still being fully test-set compatible.\n\nHope this helps!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3312081,
      "author_name": "timkrige",
      "author_url": "",
      "post_date": "11/06/2025 12:07:45",
      "content": "<p>As we only have a test image and no metadata for it, we really have to  predict based only on the test image. In training we do have additional information, though based on the current status of the discussion board, there do seem to be some unhelpful fields in this.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3382619,
      "author_name": "mugishapacifique",
      "author_url": "",
      "post_date": "12/28/2025 04:58:05",
      "content": "<p>Hi Tej,</p>\n<p>You’re correct that the test set only provides the images, so features like Species, Pre_GSHH_NDVI, or Height_Ave_cm aren’t directly available. However, one strategy that some participants use is a two-stage approach:</p>\n<p>Predict the metadata from the images</p>\n<p>Train small regression/classification models on the training set to predict Pre_GSHH_NDVI, Height_Ave_cm, and Species using the images as input.</p>\n<p>This gives “predicted metadata” for any new image.</p>\n<p>Use the predicted metadata along with the image for biomass prediction</p>\n<p>Combine image features (from a CNN, ViT, or pretrained model like DINOv2) with the predicted metadata as inputs to your final biomass regression model.</p>\n<p>This allows the model to leverage the information from the metadata without having the real values at test time.</p>\n<p>This two-stage pipeline is quite common when metadata is available in training but missing in the test set. It helps the model focus on relevant signals like pasture height or species composition while still being fully test-set compatible.</p>\n<p>Hope this helps!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3312021": "Hi all:\nIs there anyone using input variables other than the image for prediction on the test set? I assume we only have the image_id for the test data and cannot use the extra variable, such as \"Species\", \"Pre_GSHH_NDVI\", \"Height_Ave_cm\", in the model.\nCorrect me if I have missed anything here.\n\nBest Tej",
    "3312081": "As we only have a test image and no metadata for it, we really have to  predict based only on the test image. In training we do have additional information, though based on the current status of the discussion board, there do seem to be some unhelpful fields in this.",
    "3382619": "Hi Tej,\n\nYou’re correct that the test set only provides the images, so features like Species, Pre_GSHH_NDVI, or Height_Ave_cm aren’t directly available. However, one strategy that some participants use is a two-stage approach:\n\nPredict the metadata from the images\n\nTrain small regression/classification models on the training set to predict Pre_GSHH_NDVI, Height_Ave_cm, and Species using the images as input.\n\nThis gives “predicted metadata” for any new image.\n\nUse the predicted metadata along with the image for biomass prediction\n\nCombine image features (from a CNN, ViT, or pretrained model like DINOv2) with the predicted metadata as inputs to your final biomass regression model.\n\nThis allows the model to leverage the information from the metadata without having the real values at test time.\n\nThis two-stage pipeline is quite common when metadata is available in training but missing in the test set. It helps the model focus on relevant signals like pasture height or species composition while still being fully test-set compatible.\n\nHope this helps!"
  },
  "source": "meta"
}