{
  "id": 29430,
  "title": "Do pan-sharpened 3band images improve model accuracy",
  "url": "/competitions/dstl-satellite-imagery-feature-detection/discussion/29430",
  "author_name": "",
  "post_date": "2017-03-01T11:28:56.295376100Z",
  "votes": null,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I've searched through to find similar topics but with no results. My question is if kagglers are willing to share their experience with using the 3band images. Did stitching 3 band images to the 16 band images improve the model accuracy? Training a model with 3k by 3k and 800 by 800 images really makes a difference when it comes to time consumption :D</p>",
  "messages": [
    {
      "id": "164494",
      "postDate": "03/01/2017 11:28:56",
      "content": "<p>I've searched through to find similar topics but with no results. My question is if kagglers are willing to share their experience with using the 3band images. Did stitching 3 band images to the 16 band images improve the model accuracy? Training a model with 3k by 3k and 800 by 800 images really makes a difference when it comes to time consumption :D</p>",
      "rawMarkdown": "I've searched through to find similar topics but with no results. My question is if kagglers are willing to share their experience with using the 3band images. Did stitching 3 band images to the 16 band images improve the model accuracy? Training a model with 3k by 3k and 800 by 800 images really makes a difference when it comes to time consumption :D",
      "votes": null
    },
    {
      "id": "164531",
      "postDate": "03/01/2017 13:59:29",
      "content": "<p>In my case, using 3000*3000 resolution rather than 2000*2000 kind of decreases the accuracy. But yeah, 20 bands gives me significant accuracy improvement compared to 3 bands</p>",
      "rawMarkdown": "In my case, using 3000*3000 resolution rather than 2000*2000 kind of decreases the accuracy. But yeah, 20 bands gives me significant accuracy improvement compared to 3 bands",
      "votes": null
    },
    {
      "id": "164633",
      "postDate": "03/01/2017 20:39:54",
      "content": "<p>I might not have been clear. What I meant is 16 vs 20 bands rather than 3 vs 20.</p>",
      "rawMarkdown": "I might not have been clear. What I meant is 16 vs 20 bands rather than 3 vs 20.",
      "votes": null
    },
    {
      "id": "165564",
      "postDate": "03/06/2017 07:42:09",
      "content": "<p>3 band images are obtained from pansharpening MS (multispectral 8 bands) with P (panchromatic 1 band) images so 3b have more visual information than MS which has more spectral information\nSWIR gives other spectral information at much lower resolution</p>",
      "rawMarkdown": "3 band images are obtained from pansharpening MS (multispectral 8 bands) with P (panchromatic 1 band) images so 3b have more visual information than MS which has more spectral information\nSWIR gives other spectral information at much lower resolution",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 164531,
      "author_name": "lihaorocky",
      "author_url": "",
      "post_date": "03/01/2017 13:59:29",
      "content": "<p>In my case, using 3000*3000 resolution rather than 2000*2000 kind of decreases the accuracy. But yeah, 20 bands gives me significant accuracy improvement compared to 3 bands</p>",
      "votes": null,
      "replies": [
        {
          "id": 164633,
          "author_name": "philldonn",
          "author_url": "",
          "post_date": "03/01/2017 20:39:54",
          "content": "<p>I might not have been clear. What I meant is 16 vs 20 bands rather than 3 vs 20.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 165564,
      "author_name": "cogitae",
      "author_url": "",
      "post_date": "03/06/2017 07:42:09",
      "content": "<p>3 band images are obtained from pansharpening MS (multispectral 8 bands) with P (panchromatic 1 band) images so 3b have more visual information than MS which has more spectral information\nSWIR gives other spectral information at much lower resolution</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "164494": "I've searched through to find similar topics but with no results. My question is if kagglers are willing to share their experience with using the 3band images. Did stitching 3 band images to the 16 band images improve the model accuracy? Training a model with 3k by 3k and 800 by 800 images really makes a difference when it comes to time consumption :D",
    "164531": "In my case, using 3000*3000 resolution rather than 2000*2000 kind of decreases the accuracy. But yeah, 20 bands gives me significant accuracy improvement compared to 3 bands",
    "164633": "I might not have been clear. What I meant is 16 vs 20 bands rather than 3 vs 20.",
    "165564": "3 band images are obtained from pansharpening MS (multispectral 8 bands) with P (panchromatic 1 band) images so 3b have more visual information than MS which has more spectral information\nSWIR gives other spectral information at much lower resolution"
  },
  "source": "meta"
}