Lesson 1, Topic 1
In Progress

# Preprocessing of Optical imagery 1

### Learning objectives of this topic

• Why do we preprocess remote sensing data?
• Processing levels and their specifications
• Preprocessing steps for optical remote sensing imagery

In this topic, we will take a look at the preprocessing of satellites imagery. We will pay attention to the effects, which are being corrected during the preprocessing process. These preprocessing steps are inherent to most satellite data sets, and they depend on the acquisition platform/technology and therefore can pose significant challenges to data analysis.

#### The purpose of preprocessing satellite imagery

Raw satellite imagery inherits systemic error of various kinds. These systematic errors heavily depend on the acquisition technique. These ‘unwanted’ effects can be of radiometric or spatial nature, meaning that the value that was recorded, does not relate to the actual ground observation or that the pixel we look at is distorted in some way. Using this as a basis for any kind of scientific analysis propagates errors, which can lead to misclassifications and wrong conclusions. Consequently, solely completely calibrated data allows to detect and precisely quantify trends of changes in land cover.

Let’s take a look at the preprocessing of optical data first.

### Preprocessing of optical data

#### Processing levels

With optical data sets, we can divide the data into what is called ‘processing levels’. This term describes how many different calibration and registration algorithms were performed on a certain data set. Below, you can see the breakdown of the structure of Sentinel-2 data that is created by the European Space Agency (ESA). The structure is exemplary for other commonly used optical satellites.

The figure on the right visualizes the levels of processing, which can be divided into ‘sensor geometry’ and ‘ground geometry’. This implies that levels 0, 1A as well as 1B represent the actual values measured at the sensor, which do not yet relate to ground information. These levels are not made publicly available by ESA, since they cannot be used for analysis right away. Product Level 1C represent radiometrically and geometrically corrected images, which could be used for spatial analysis. Finally, the highest processing level provided by ESA are the Level-2A data sets. These include scene classifications and additional outputs such as quality maps and, most importantly, are atmospherically corrected, meaning that the given value does represent a reflectance on the ground.

Using the Level-1C format you can preprocess that data on your own using, in case of Sentinel-2, the sen2cor preprocessor in order to create data in Level 2A format yourself. There are other preprocessing algorithms, which are not explained in detail here. However, ‘ready-to-use’ data can be downloaded from different data hubs, without the need for further preprocessing.

### Most important preprocessing steps for optical data

Now, let’s go through the most important steps, which are part of preprocessing. This process is also referred to as image restoration and rectification.

The raw digital numbers (DN’s) of optical data sets are not physically meaningful units per se but simply just numbers. These numbers are created using gain and offset factors that are unique to each sensor. However, in order to compare data sets from different sensors with each other, it is crucial to somehow normalize the information. This procedure is also referred to as ‘sensor calibration’ and enables the interpretation of data sets with varying origin by converting DN’s into physically meaningful variables such as radiance, reflectance or brightness temperature.
Fortunately, commonly used products such as Landsat data or Sentinel-2 data are already distributed in a radiometrically calibrated version by the respective Space Agencies.

#### Value conversion

The conversion from DN to radiance and reflectance is an important step in the preprocessing. In the case of Sentinel-2 and Landsat data, we are usually interested in the bottom-of-atmosphere (BOA) reflectance, which is, in contrast to the top-of-atmosphere (TOA) reflectance, corrected for atmospheric impacts and can thus directly be interpreted. However, the initially available data sets are being converted from DN to TOA format. For the conversion, the data is rescaled using sensor specific equations containing information about sun-related angles and other parameters. As an example for such a conversion, the expression used to correct Landsat 8 OLI data is given below.

\rho\lambda = \frac{M_{\rho}Q_{cal} + A_{\rho}}{\cos \theta_{SZ}}

ρλ = TOA reflectance
Mρ​ = multiplicative rescaling factor
Qcal = DN (quantized & calibrated)
θSZ = local solar zenith angle, θSZ = 90° – θSE = local sun elevation angle

Using these equations, we can compute TOA values which is are subject to strong illumination geometry difference-related effects and therefore should not be favored over BOA reflectances.

#### Atmospheric correction

The atmosphere plays an important role due to its interaction with the radiation. Especially, optical data can easily be altered by atmospheric influences, leading to heavy impacts on the intensity of the reflected light that is received at the sensor. Let’s quickly take a look at the possible paths of light in our atmosphere again.

Methods to minimize such effects are atmospheric models and dark object subtraction (DOS) or radiative transfer models such as LOWTRAN or MODTRAN. For modelling approaches, a number of parameters need to be known, such as the amount of water vapor or the distribution of aerosol in the atmosphere. Transfer models compute estimates of true reflectance (without atmospheric influences) by linking various physical parameters. The DOS method can be used when there is no data available on the atmospheric conditions mentioned earlier. The underlying assumption to DOS is, that every dark object’s reflectance contains a substantial amount of scattering caused by atmospheric effects. After detecting the darkest pixel in each band, this value can then be subtracted from every pixel in the respective channel. Exemplary, this method can be useful to correct for haze effects.

#### Other preprocessing steps

Below, you can explore a couple of additional algorithms and how they influence the remote sensing image visually (and statistically).

Stretching of digital numbers (rescaling over the range of the greyscale histogram) to allow better visibility and pixel distinction.

Geometric correction to remove spatial distortions. This process corrects the discrepancy between an image coordinate and the actual position on the Earth’s surface. Exemplary, this can be done using ground control points (GCPs) from known ‘correct’ images.

Low pass filter to smoothen satellite imagery using moving windows (e.g., 3×3 pixels), which are used to calculate local averages for kernel pixels.

High pass filter to sharpen edges and increasing contrast. They enhance spatial frequencies/components (e.g., edges). Subtracts the low pass filtered from the original image.

If you want to preprocess Landsat data, we can recommend an excellent guide for preprocessing compiled by Nicholas E. Young et al. 2017. You can download the PDF using the link below.

European Space Agency (ESA, 2021): Sentinel-2 MSI. Overview. <https://dragon3.esa.int/web/sentinel/user-guides/sentinel-2-msi/overview>.

Gottwald, M. & Bovensmann, H. (2011.). SCIAMACHY – Exploring the Changing Earth’s Atmosphere. Springer Heidelberg Dordrecht London New York.

Rees, W.G. (2010²). Physical Principles of Remote Sensing. Cambridge, USA: Cambridge University Press.

Schowengerdt, R.A. (2007³). Remote Sensing. Models and Methods for Image Processing. San Diego, USA: Academic Press.

### Learning objectives of this topic

• Why do we preprocess remote sensing data?
• Processing levels and their specifications
• Preprocessing steps for optical remote sensing imagery

In this topic, we will take a look at the preprocessing of satellites imagery. We will pay attention to the effects, which are being corrected during the preprocessing process. These preprocessing steps are inherent to most satellite data sets, and they depend on the acquisition platform/technology and therefore can pose significant challenges to data analysis.

#### The purpose of preprocessing satellite imagery

Raw satellite imagery inherits systemic error of various kinds. These systematic errors heavily depend on the acquisition technique. These ‘unwanted’ effects can be of radiometric or spatial nature, meaning that the value that was recorded, does not relate to the actual ground observation or that the pixel we look at is distorted in some way. Using this as a basis for any kind of scientific analysis propagates errors, which can lead to misclassifications and wrong conclusions. Consequently, solely completely calibrated data allows to detect and precisely quantify trends of changes in land cover.

Let’s take a look at the preprocessing of optical data first.

### Preprocessing of optical data

#### Processing levels

With optical data sets, we can divide the data into what is called ‘processing levels’. This term describes how many different calibration and registration algorithms were performed on a certain data set. Below, you can see the breakdown of the structure of Sentinel-2 data that is created by the European Space Agency (ESA). The structure is exemplary for other commonly used optical satellites.

The figure on the right visualizes the levels of processing, which can be divided into ‘sensor geometry’ and ‘ground geometry’. This implies that levels 0, 1A as well as 1B represent the actual values measured at the sensor, which do not yet relate to ground information. These levels are not made publicly available by ESA, since they cannot be used for analysis right away. Product Level 1C represent radiometrically and geometrically corrected images, which could be used for spatial analysis. Finally, the highest processing level provided by ESA are the Level-2A data sets. These include scene classifications and additional outputs such as quality maps and, most importantly, are atmospherically corrected, meaning that the given value does represent a reflectance on the ground.

Using the Level-1C format you can preprocess that data on your own using, in case of Sentinel-2, the sen2cor preprocessor in order to create data in Level 2A format yourself. There are other preprocessing algorithms, which are not explained in detail here. However, ‘ready-to-use’ data can be downloaded from different data hubs, without the need for further preprocessing.

### Most important preprocessing steps for optical data

Now, let’s go through the most important steps, which are part of preprocessing. This process is also referred to as image restoration and rectification.

The raw digital numbers (DN’s) of optical data sets are not physically meaningful units per se but simply just numbers. These numbers are created using gain and offset factors that are unique to each sensor. However, in order to compare data sets from different sensors with each other, it is crucial to somehow normalize the information. This procedure is also referred to as ‘sensor calibration’ and enables the interpretation of data sets with varying origin by converting DN’s into physically meaningful variables such as radiance, reflectance or brightness temperature.
Fortunately, commonly used products such as Landsat data or Sentinel-2 data are already distributed in a radiometrically calibrated version by the respective Space Agencies.

#### Value conversion

The conversion from DN to radiance and reflectance is an important step in the preprocessing. In the case of Sentinel-2 and Landsat data, we are usually interested in the bottom-of-atmosphere (BOA) reflectance, which is, in contrast to the top-of-atmosphere (TOA) reflectance, corrected for atmospheric impacts and can thus directly be interpreted. However, the initially available data sets are being converted from DN to TOA format. For the conversion, the data is rescaled using sensor specific equations containing information about sun-related angles and other parameters. As an example for such a conversion, the expression used to correct Landsat 8 OLI data is given below.

\rho\lambda = \frac{M_{\rho}Q_{cal} + A_{\rho}}{\cos \theta_{SZ}}

ρλ = TOA reflectance
Mρ​ = multiplicative rescaling factor
Qcal = DN (quantized & calibrated)
θSZ = local solar zenith angle, θSZ = 90° – θSE = local sun elevation angle

Using these equations, we can compute TOA values which is are subject to strong illumination geometry difference-related effects and therefore should not be favored over BOA reflectances.

#### Atmospheric correction

The atmosphere plays an important role due to its interaction with the radiation. Especially, optical data can easily be altered by atmospheric influences, leading to heavy impacts on the intensity of the reflected light that is received at the sensor. Let’s quickly take a look at the possible paths of light in our atmosphere again.

Methods to minimize such effects are atmospheric models and dark object subtraction (DOS) or radiative transfer models such as LOWTRAN or MODTRAN. For modelling approaches, a number of parameters need to be known, such as the amount of water vapor or the distribution of aerosol in the atmosphere. Transfer models compute estimates of true reflectance (without atmospheric influences) by linking various physical parameters. The DOS method can be used when there is no data available on the atmospheric conditions mentioned earlier. The underlying assumption to DOS is, that every dark object’s reflectance contains a substantial amount of scattering caused by atmospheric effects. After detecting the darkest pixel in each band, this value can then be subtracted from every pixel in the respective channel. Exemplary, this method can be useful to correct for haze effects.

#### Other preprocessing steps

Below, you can explore a couple of additional algorithms and how they influence the remote sensing image visually (and statistically).

Stretching of digital numbers (rescaling over the range of the greyscale histogram) to allow better visibility and pixel distinction.

Geometric correction to remove spatial distortions. This process corrects the discrepancy between an image coordinate and the actual position on the Earth’s surface. Exemplary, this can be done using ground control points (GCPs) from known ‘correct’ images.

Low pass filter to smoothen satellite imagery using moving windows (e.g., 3×3 pixels), which are used to calculate local averages for kernel pixels.

High pass filter to sharpen edges and increasing contrast. They enhance spatial frequencies/components (e.g., edges). Subtracts the low pass filtered from the original image.

If you want to preprocess Landsat data, we can recommend an excellent guide for preprocessing compiled by Nicholas E. Young et al. 2017. You can download the PDF using the link below.