{
  "id": 619477,
  "title": "Does Metadata Really Matter?",
  "url": "/competitions/csiro-biomass/discussion/619477",
  "author_name": "",
  "post_date": "2025-11-14T09:09:57.423473800Z",
  "votes": 17,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I ran a simple ablation study using the same EfficientNetV2-S model across five metadata configurations, and the results show that adding metadata consistently improves R² performance:</p>\n<table>\n<thead>\n<tr>\n<th>Experiment</th>\n<th>Image</th>\n<th>Date</th>\n<th>Height</th>\n<th>NDVI</th>\n<th>State</th>\n<th>Species</th>\n<th>R² (mean ± std)</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td><strong>baseline</strong></td>\n<td>✔</td>\n<td>✖️</td>\n<td>✖️</td>\n<td>✖️</td>\n<td>✖️</td>\n<td>✖️</td>\n<td><strong>0.445 ± 0.163</strong></td>\n</tr>\n<tr>\n<td><strong>temporal_features</strong></td>\n<td>✔</td>\n<td>✔</td>\n<td>✖️</td>\n<td>✖️</td>\n<td>✖️</td>\n<td>✖️</td>\n<td><strong>0.585 ± 0.147</strong></td>\n</tr>\n<tr>\n<td><strong>measurement_features</strong></td>\n<td>✔</td>\n<td>✖️</td>\n<td>✔</td>\n<td>✔</td>\n<td>✖️</td>\n<td>✖️</td>\n<td><strong>0.526 ± 0.154</strong></td>\n</tr>\n<tr>\n<td><strong>contextual_features</strong></td>\n<td>✔</td>\n<td>✖️</td>\n<td>✖️</td>\n<td>✖️</td>\n<td>✔</td>\n<td>✔</td>\n<td><strong>0.605 ± 0.118</strong></td>\n</tr>\n<tr>\n<td><strong>all_features</strong></td>\n<td>✔</td>\n<td>✔</td>\n<td>✔</td>\n<td>✔</td>\n<td>✔</td>\n<td>✔</td>\n<td><strong>0.676 ± 0.105</strong></td>\n</tr>\n</tbody>\n</table>\n<ul>\n<li>Find out more in the <a href=\"https://www.kaggle.com/code/amirmmahdavikia/csiro-does-metadata-really-matter\" target=\"_blank\">notebook</a></li>\n</ul>",
  "messages": [
    {
      "id": "3323519",
      "postDate": "11/14/2025 09:09:57",
      "content": "<p>I ran a simple ablation study using the same EfficientNetV2-S model across five metadata configurations, and the results show that adding metadata consistently improves R² performance:</p>\n<table>\n<thead>\n<tr>\n<th>Experiment</th>\n<th>Image</th>\n<th>Date</th>\n<th>Height</th>\n<th>NDVI</th>\n<th>State</th>\n<th>Species</th>\n<th>R² (mean ± std)</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td><strong>baseline</strong></td>\n<td>✔</td>\n<td>✖️</td>\n<td>✖️</td>\n<td>✖️</td>\n<td>✖️</td>\n<td>✖️</td>\n<td><strong>0.445 ± 0.163</strong></td>\n</tr>\n<tr>\n<td><strong>temporal_features</strong></td>\n<td>✔</td>\n<td>✔</td>\n<td>✖️</td>\n<td>✖️</td>\n<td>✖️</td>\n<td>✖️</td>\n<td><strong>0.585 ± 0.147</strong></td>\n</tr>\n<tr>\n<td><strong>measurement_features</strong></td>\n<td>✔</td>\n<td>✖️</td>\n<td>✔</td>\n<td>✔</td>\n<td>✖️</td>\n<td>✖️</td>\n<td><strong>0.526 ± 0.154</strong></td>\n</tr>\n<tr>\n<td><strong>contextual_features</strong></td>\n<td>✔</td>\n<td>✖️</td>\n<td>✖️</td>\n<td>✖️</td>\n<td>✔</td>\n<td>✔</td>\n<td><strong>0.605 ± 0.118</strong></td>\n</tr>\n<tr>\n<td><strong>all_features</strong></td>\n<td>✔</td>\n<td>✔</td>\n<td>✔</td>\n<td>✔</td>\n<td>✔</td>\n<td>✔</td>\n<td><strong>0.676 ± 0.105</strong></td>\n</tr>\n</tbody>\n</table>\n<ul>\n<li>Find out more in the <a href=\"https://www.kaggle.com/code/amirmmahdavikia/csiro-does-metadata-really-matter\" target=\"_blank\">notebook</a></li>\n</ul>",
      "rawMarkdown": "I ran a simple ablation study using the same EfficientNetV2-S model across five metadata configurations, and the results show that adding metadata consistently improves R² performance:\n\n| Experiment | Image | Date | Height | NDVI | State | Species | R² (mean ± std) |\n|-----------|:-----:|:----:|:------:|:----:|:-----:|:-------:|------------------|\n| **baseline** | ✔ | ✖️ | ✖️ | ✖️ | ✖️ | ✖️ | **0.445 ± 0.163** |\n| **temporal_features** | ✔ | ✔ | ✖️ | ✖️ | ✖️ | ✖️ | **0.585 ± 0.147** |\n| **measurement_features** | ✔ | ✖️ | ✔ | ✔ | ✖️ | ✖️ | **0.526 ± 0.154** |\n| **contextual_features** | ✔ | ✖️ | ✖️ | ✖️ | ✔ | ✔ | **0.605 ± 0.118** |\n| **all_features** | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | **0.676 ± 0.105** |\n\n* Find out more in the [notebook](https://www.kaggle.com/code/amirmmahdavikia/csiro-does-metadata-really-matter)",
      "votes": null
    },
    {
      "id": "3323528",
      "postDate": "11/14/2025 09:20:45",
      "content": "<p>This is interesting, how does this translate to the unseen test data? Are you able to achieve a similar upward trend on the LB? </p>\n<p>I find that using some metadata is  generalizable, others is quite hard to get right.</p>",
      "rawMarkdown": "This is interesting, how does this translate to the unseen test data? Are you able to achieve a similar upward trend on the LB? \n\nI find that using some metadata is  generalizable, others is quite hard to get right.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3323528,
      "author_name": "timkrige",
      "author_url": "",
      "post_date": "11/14/2025 09:20:45",
      "content": "<p>This is interesting, how does this translate to the unseen test data? Are you able to achieve a similar upward trend on the LB? </p>\n<p>I find that using some metadata is  generalizable, others is quite hard to get right.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3323519": "I ran a simple ablation study using the same EfficientNetV2-S model across five metadata configurations, and the results show that adding metadata consistently improves R² performance:\n\n| Experiment | Image | Date | Height | NDVI | State | Species | R² (mean ± std) |\n|-----------|:-----:|:----:|:------:|:----:|:-----:|:-------:|------------------|\n| **baseline** | ✔ | ✖️ | ✖️ | ✖️ | ✖️ | ✖️ | **0.445 ± 0.163** |\n| **temporal_features** | ✔ | ✔ | ✖️ | ✖️ | ✖️ | ✖️ | **0.585 ± 0.147** |\n| **measurement_features** | ✔ | ✖️ | ✔ | ✔ | ✖️ | ✖️ | **0.526 ± 0.154** |\n| **contextual_features** | ✔ | ✖️ | ✖️ | ✖️ | ✔ | ✔ | **0.605 ± 0.118** |\n| **all_features** | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | **0.676 ± 0.105** |\n\n* Find out more in the [notebook](https://www.kaggle.com/code/amirmmahdavikia/csiro-does-metadata-really-matter)",
    "3323528": "This is interesting, how does this translate to the unseen test data? Are you able to achieve a similar upward trend on the LB? \n\nI find that using some metadata is  generalizable, others is quite hard to get right."
  },
  "source": "meta"
}