• Enhancing satellite imagery with AI: superresolution

     

     

    Enhancing satellite imagery with AI: superresolution. - Dr. Daniel García Díaz

    Multispectral Satellites

    Usually composed of a multispectral sensor and a panchromatic sensor.

    • Multi spectral Imaging
      Capture images in various bands of the electromagnetic spectrum, allowing different types of features on the Earth’s surface to be detected.
    • Panchromatic Image
      A single band of great spectral width that typically covers part of the visible and the beginning of the infrared. The main usefulness of this image is that it is taken at a higher spatial resolution than the rest of the multispectral bands to complement them. It is important for the detection of small indistinguishable elements in multispectral images.



    Resolution of Multispectral Images

    • Spatial Resolution
      Describes the systems ability to distinguish objects based on their size. Designates the size of the smallest object that can be distinguished in an image.


    • Spectral Resolution
      Indicates the number and width of spectral bands that the sensor can discriminate. In this way, a sensor will have better spectral resolution the greater the number of bands it provides and the more specific their width.


    Sentinel-2A/B

    • Spatial Resolution
      10 meters in the visible and near-infrared bands and up to 20 meters in the short-wave infrared band.
    • Spectral Bands
      13 spectral bands covering from the visible to the near-infrared.
    • High Temporal Revisit Frequency
      Capture images of the same area every few days (~ 5 days).



    Landsat 8

    • High Spatial Resolution
      30 meters for most bands, while the panchromatic band offers a higher resolution of up to 15 meters.
    • Spectral Bands
      11 spectral bands covering a wide range of wavelengths from visible to short-wave infrared.
    • Temporal Revisit Frequency
      Landsat 8 revisits the same area every 16 days.



    MODIS

    • Spectral Bands
      36 bands covering wavelengths from visible to thermal infrared (0.40 - 14.40 μm).
    • Spatial Resolution
      Three different resolutions - 250 meters, 500 meters and 1 kilometer.

    VIIRS

    • Spectral Bands
      22 bands covering wavelengths from visible to thermal infrared.
    • Spatial Resolution
      Ranges from 375 meters to 1.5 kilometers, depending on the spectral band.

    Superresolution

    • Increase the spatial resolution of the low-resolution multispectral bands and bring the mto the same resolution as the panchromatic band.



    Traditional Superresolution Algorithms

    • Component Substitution
      This method transforms components of the original multispectral (MS) image to isolate spatial information. By replacing this component with the panchromatic (PAN) image, a fused image with high spectral and spatial resolution is achieved. Examples: IHS or PCA.
    • Arithmetic operations
      Fused images are generated through algebraic operations between bands of the MS image and the PAN image. A common technique is the Brovey method.
    • High frequency injection
      This approach extracts high-frequency components (details) from the PAN image and integrates them into the MS image. It often involves using high-pass filters.
    • Multiresolution analysis
      MS and PAN bands are decomposed at different scales or resolutions to extract spatial details, which are then incorporated into the MS bands. This technique frequently employs methods like the Discrete Wavelet Transform.

    Convolutional Neural Network for Superresolution

    It is assumed that the textures of the image behave similarly at different scales. This allows for the application of image processing methods that rely on spectral correlation to enhance resolution or perform other operations effectively.
    Learn to transfer high frequency content to low resolution bands.

    • Train
      Sampling from 40 m to 20 m.
      Sampling from 360 m to 60 m.
    • Predict
      Increase from 20 m to 10 m.
      Increase from 60 m to 10 m.



    Superresolution Tutorial

    The next part of the course is very practical, so it is recommended to go through it with the video at the beginning of this section.


    ◄Image segmentation with AI techniquesTrain an AI4EOSC module (using YOLO)►