Segmentation wrong result

OTB Usage
1

Hello, everyone. I’m new to OTB and I need some help in segmentation. I don’t understand why don’t get a vector file with segments, and only a square covering the area. I input Landsat - 5, 7 and 8 and used the meanshift algorithm and maintained the default values on the parameters. Here is a pic of what I’m talking about:

image
image1366×720 111 KB

2

Dear @gabrieltolentino,

Thank you for using the OTB!

Maybe the default parameters are not adapted to you image. Do you want to share your image with us, so we can test?

3

Hello! Yeah, it can be.
https://drive.google.com/drive/folders/14e_wtKN5btC8jYNGsinxyNfbX9RM0Pov?usp=sharing

4

You need to adapt the Range radius parameter -ranger to the dynamic of your image.

When I use the application ComputeImageStatistics on the Mosaico image, I get

<?xml version="1.0" ?>
<FeatureStatistics>
    <Statistic name="mean">
        <StatisticVector value="0.0112117" />
        <StatisticVector value="0.0739337" />
        <StatisticVector value="0.0396668" />
    </Statistic>
    <Statistic name="stddev">
        <StatisticVector value="0.0207732" />
        <StatisticVector value="0.121765" />
        <StatisticVector value="0.067313" />
    </Statistic>
</FeatureStatistics>

This means that the value of the pixels are of the order of 10e-2. If you keep the default value of range radius (15), the algorithm will consider that all the pixels are the same. For example, if I take a pixel in the city, it has the value [0.04, 0.4, 0.15]. And if I take a pixel in the forest, it has the value [0.1, 0.15, 0.18]. The euclidean distance between those two pixels is 0.09. If you set a range radius higher than 0.09, the algorithm won’t separate city and forest pixels.

5

Hello,

May you explain how did you calculate the Euclidean distance between the two pixels ?
How is it helping in the configuration of the segmentation algorithm ?

Thank you !