WEEK 3

Remote sensing data

This is a learning diary of CASA0023 WEEK 3, the lecture presentation is here, and the practical material is here.

1 Summary: lecture

This week focuses on the pre-processing of remote sensing data, including the correction, joining and enhancement of various types of remote sensing data.

Mindmap of Week 3 Leacture

1.1 Pre-knowledge

Different sensors

  • MSS (Multispectral Scanner)
  • RBV (Return Beam Vidicon Camera)

Push broom vs Whisk broom

  • Whisk broom or spotlight or across track scanners: Mirror reflects light onto 1 detector - Landsat
  • Push broom or along track scanners: several detectors that are pushed along - SPOT, Quickbird

1.2 Corrections

Regression

\[ y_i=\beta_0+\beta_1x_i+\varepsilon_i \]

  • β0 is the intercept (the value of y when x = 0)
  • β1 the ‘slope’ the change in the value of y for a 1 unit change in the value of x
  • ϵi is a random error term (positive or negative)- if you add all of the vertical differences between the blue line and all of the residuals, it should sum to 0
  • Any value of y along the blue line can be modeled using the corresponding value of x

Geometric correction

What leads to image distortions

  • View angle (off-nadir)
  • Topography (e.g. hills not flat ground)
  • Wind (if from a plane)
  • Rotation of the earth (from satellite)

Geometric correction solution

Geometric correction. Source: Abdul Basith

  1. Identify Ground Control Points (GPS) to match known points in the image and a reference dataset

  2. Take the coordinates and model them to give geometric transformation coefficients, linear regression with our distorted x or y as the dependent or independent

    • Input to output (forward mapping): the issue with this is that we are modelling the rectified x and y which could fall anywhere on the gold standard map (e.g. not on a grid square or at a floating point)
    • Output to input (backward mapping): for every value in the output (gold standard) pixel we can get a value in the original input image. The images are distorted as so might not completely overlap. The goal is to match the distorted image with the gold standard image, so we want the pixels to line up

  3. Plot these and try to minimise the RMSE - Jensen sets a RMSE value of 0.5, typically might add more GCPs to reduce the RMSE (The model with the lowest RMSE will fit best)

    • RMSE: (observed - predicted (the residual))^2, sum them and divide by number of data points, square root that total
    • Resample methods: Nearest Neighbor, Linear, Cubic, Cubic spline

Atmospheric correction

Necessary and unnecessary atmospheric correction

  • Necessary
    • Biophysical parameters needed (e.g. temperature, leaf area index, NDVI)
    • Using spectral signatures through time and space
  • Unnecessary
    • Classification of a single image
    • Independent classification of multi date imagery
    • Composite images (combining images)
    • Single dates or where training data extracted from all data

Atmospheric correction in action

Atmospheric correction examples of three scenes (Bands 1, 2, and 3). Source: Liang et al. 2001

  • Absorption and scattering create the haze = reduces contrast of image
  • Scattering = can create the “adjacency effect”, radiance from pixels nearby mixed into pixel of interest

Atmospheric correction types

  • Relative (to something)
    • Normalize
      • Normalize intensities of different bands within a single image
      • Normalise intensities of bands from many dates to one date
    • Dark object subtraction (DOS) or histogram adjustment
      • Searches each band for the darkest value then subtracts that from each pixel
      • Landsat bands 1-3 (visible) have increased scattering vs longer wavelengths
    • Psuedo-invariant Features (PIFs)
      • Assume brightness pixels linearly related to a base image
      • Regression per band
      • Adjust the image based on the regression result
      • Here y is the value of our base. To get y we multiply our new date pixel (x) by the coefficient and add the intercept value
      • Apply this to the rest of the pixels
  • Absolute (definitive)
    • Method
      • Change digital brightness values into scaled surface reflectance. We can then compare these scaled surface reflectance values across the planet
      • We do this through atmospheric radiative transfer models and there are many to select from
      • However, nearly all assume atmospheric measurements are available which are used to “invert” the image radiance to scaled surface reflectance
      • The scattering and absorption information comes from atmopshierc radiative transfer code such as MODTRAN 4+ and the Second Simulation of the Satellite Signal in the Solar Spectrum (6S), which can now be used through python - called Py6S
    • Absolute Data requirements
      • An atmopsheric model (summer, tropical): usually you can select from the tool
      • Local atmopsheric visibility: from a weather station, like airports
      • Image altitude
    • Absolute Tools
      • ACORN: Atmopsehic CORection Now
      • FLAASH: Fast Line of-sight Atmopsheric Analysis
      • QUAC: Quick Atmopsheric Correction
      • ATCOR: The ATmospheric CORrection program
      • SMAC: Simplified Model for Atmospheric Correction (SMAC)
  • Empirical Line Correction
    • We can go and take measurements in situ using a field spectrometer, this does require measurements at the same time as the satellite overpass
    • Then use these measurements in linear regression against the satellite data raw digital number

Orthorectification / Topographic correction

A view captured from an oblique angle (for example, 25°, left) must be corrected for relief displacement caused by terrain to generate the orthorectified view (looking straight down, right). Orthoimagery is produced by calculating the nadir view for every pixel. Source: Esri Insider, 2016

A subset of georectification

  • Georectification = giving coordinates to an image
  • Orthorectification = removing distortions… making the pixels viewed at nadir (straight down)

Requires

  • Sensor geometry
  • An elevation model

Software / formulas to do this

  • Jensen covers the following formulas: Cosine correction, Minnaert correction, Statistical Empirical correction, C Correction (advancing the Cosine)
  • Software: QGIS, SAGA GIS, R package topocorr, R package RStoolbox

Solar location

  • Solar azimuth = compass angle of the sun (N =0°) 90° (E) at sunrise and 270° (W) at sunset
  • Solar zenith = angle of local zenith (above the point on ground) and sun from vertical (90° - elevation)

Radiometric calibration

  • Sensors capture image brightness and distributed as a Digital Number (or DN) - allows for efficient storage but has no units
  • Spectral radiance is the amount of light within a band from a sensor in the field of view (FOV), it is independent of the sensor, measured in Watts (power or light here)
  • DN to spectral radiance = radiometric calibration
  • Sensor calibration = the relationship between Gain and Bias are usually provided but we can calcaulte them
Terms
  • Radiance: refers to any radiation leaving the Earth (i.e. upwelling, toward the sensor
    • Might also be called Top of Atmosphere (TOA) radiance
    • How much light the instrument sees in meaningful units but still have effects of: Light source, atmosphere and surface material
    • We can remove the effects of the light source to generate Top of Atmosphere reflectance but usually this is combined within the radiance to reflectance step
  • Irradiance: is used to describe downwelling radiation reaching the Earth from the sun
  • Digital number (DN)
    • Intensity of the electromagnetic radiation per pixel
    • Pixel values that aren’t calibrated and have no unit
    • Have light source
    • Effects of sensor + atmosphere + material
    • Values range from 0 - 255 (Lansat 5) = 8 bit or 0 - 65536 Landsat 8 (12 bit)
  • Reflectance: We need to account for atmospheric and illumination effects to create reflectance. BUT this typically doesn’t deal with shadows and directional effects (e.g. viewing angles) = apparent reflectance However, this is often called reflectance
    • Reflectance is a property of a material (e.g. reflectance of grass is a property of grass)
    • The issue with radiance is that is contains physical properties AND is dependent on the light source
  • Hemispherical reflectance: all of the light leaving the surface goes straight to the sensor (nothing is intercepted or at an angle)
  • Path radiance: radiance reflected above the surface (e.g. scattering)
  • Atmospheric attenuation: absorption of EMR due to materials in atmosphere (e.g. water vapour)
  • Local: specific to pixel
  • `Neighbourhood: pixels within a range (nearby)

1.3 Data joining and enhancement

Joining data sets

  • This is termed “Mosaicking” in remote sensing - but it’s not much different to merging in GIS
  • In Remote Sensing we usually feather the images together
  • This creates a seamless mosaic or image(s)
  • The dividing line is termed the seamline
  • We have a base image and “other” or second image

How to join data sets

  • Standardization (dividing the SR value by a maximum value per band) and normalization (divide the standarised value by the sum of values across all bands) applied to each image
  • Undertake further relative radiometric normalization
  • Classify each image alone
  • Calculate other metrics from the image

Image enhancement

Contrast enhancement

Contrast enhancement in QGIS. Source: Atilo Francois

  • Minimum - Maximum
  • Percentage Linear and Standard Deviation
  • Piecewise Linear Contrast Stretch

Ratio

  • Band ratioing means dividing the pixels in one band by the corresponding pixels in a second band.
  • Example: Normalized Burn Ratio = (NIR - SWIR) / (NIR + SWIR)
    • In Landsat 4-7, NBR = (Band 4 – Band 7) / (Band 4 + Band 7)
    • In Landsat 8-9, NBR = (Band 5 – Band 7) / (Band 5 + Band 7)

Filtering

  • Low pass or low frequency (averages the surrounding pixels)
  • High pass or high frequency - enhance local variations
  • Edge enhancement

PCA (Principal Component)

  • Transform multi-spectral data into uncorrelated and smaller dataset
  • Has most of the original information
  • Reduces future computation “dimensionatliy reduction”
  • The first component will capture most of the variance within the dataset
  • In R this is from the RStoolbox packagerasterPCA()
  • PCA example, multi-date PCA - bands from both time points are combined into one image, then PCA

Texture

  • Images just use tonal (spectral) data not texture
  • Texture: spatial variation of gray values
  • First order (occurrence): use counts or occurrences
  • Second order(co-occurrence): relationship between pixel pairs “a function of both the angular relationship and distance between two (or kernel) neighboring pixels”

Fusion

Image fusion is where data from multiple sensors / sources is fused together

  • Pan sharpen
  • Data Fusion

2 Summary: practical

This week’s practical will consist mainly of the following:

  • Introduction and access to Landsat data
  • A deeper understanding of atmospheric correction and how it works in practice
  • Introduction to the principles of Radiance (or DN) to Reflectance
  • Practical exercises in image joining
  • A deep understanding of image enhancement and how it works in practice

For this week’s practical I chose to follow the guidance using Landsat data from Cape Town. I used R to do atmospheric corrections, merging, data enhancement and other operations on the data. Through practical, I gained a better understanding of remote sensing data pre-processing and was able to apply this to my workflow in R. Combined with the previous term’s CASA0005 course, I can now use R to analyse more types of data.

Practical output: texture of landsat data from Cape Town

3 Application

This week, the main focus was on the correction, joining and enhancement of remote sensing images, mostly applied to pre-processing before analysis, but remote sensing image enhancement has a wider range of applications.

3.1 Applications of remote sensing image enhancement

Remote sensing image enhancement refers to the improvement of the quality and information of remote sensing images through some technical means to make them more suitable for human vision or subsequent analytical processing. Remote sensing image enhancement can be applied to the following areas:

Contrast enhancement: By adjusting the grey level of a remotely sensing image, the contrast of the image is increased to make it clearer and brighter. Contrast enhancement can be done using methods such as histogram stretching, histogram equalisation and segmented linear stretching.

Band ratioing: Extracts specific information from remote sensing images, such as vegetation indices, water indices, soil indices, etc., by calculating the ratio between different bands. Band ratios can be calculated using methods such as band arithmetic and band combination.

Filtering: Remove noise or enhance edges and textures in an image by convolving the remote sensing image in the spatial or frequency domain. Filtering can use methods such as smoothing filtering, sharpening filtering and edge detection filtering.

Principal component analysis (PCA): By applying orthogonal transformations to multiple bands of remote sensing images, the main information in the image is extracted and the redundancy and correlation of the data is reduced. Principal component analysis can use statistical methods or methods such as wavelet transform.

Texture: Describes the surface roughness or structural features in an image by calculating the grey scale variation within a region of the remotely sensed image. Texturing can use methods such as grey level co-occurrence matrix, grey level distance matrix, grey level gradient matrix, etc.

Fusion: Increasing the spatial or spectral resolution of an image and increasing the amount of information in the image by combining remote sensing images from different sources or at different resolutions. Fusion can be done using methods such as HIS transform, wavelet transform, multi-resolution analysis, etc.

3.2 Application case

Sourse

PCA‐based land‐use change detection and analysis using multitemporal and multisensor satellite data , Sourse: Deng et al. (2008)

Remote‐sensing change detection based on multitemporal, multispectral, and multisensor imagery has been developed over several decades and provided timely and comprehensive information for planning and decision‐making. In practice, however, it is still difficult to select a suitable change‐detection method, especially in urban areas, because of the impacts of complex factors.

This paper presents a new method using multitemporal and multisensor data (SPOT‐5 and Landsat data) to detect land‐use changes in an urban environment based on principal‐component analysis (PCA) and hybrid classification methods. After geometric correction and radiometric normalization, PCA was used to enhance the change information from stacked multisensor data. Then, a hybrid classifier combining unsupervised and supervised classification was performed to identify and quantify land‐use changes. Finally, stratified random and user‐defined plots sampling methods were synthetically used to obtain total 966 reference points for accuracy assessment.

Example of land‐use changes from cropland to urban land (Economic and Technological Development Zone). (a) Pan image of ETM (2000); (b) aerial photograph (2000); (c) RGB composition image of SPOT‐5 (2003); (d) RGB composition image of IKONOS (2003); (e)–(h) first four principal components.

Although errors and confusion exist, this method shows satisfying results with an overall accuracy to be 89.54% and 0.88 for the kappa coefficient. When compared with the post‐classification method, PCA‐based change detection also showed a better accuracy in terms of overall, producer’s, and user’s accuracy and kappa index. The results suggested that significant land‐use changes have occurred in Hangzhou City from 2000 to 2003, which may be related to rapid economy development and urban expansion. It is further indicated that most changes occurred in cropland areas due to urban encroachment.

Land use and land‐use change detected using the PCA‐based approach.

3.3 Case comments

Advantages or contribution

  • Flexible classification strategies: The hybrid classification approach allows the authors to combine the accuracy of supervised classification methods while taking advantage of the automation of unsupervised classification methods. This flexibility allows the method to be better adapted to different types of land use change scenarios.
  • PCA-based feature extraction: By combining multi-temporal and multi-sensor data into a single Principal Component Analysis (PCA) model, the authors are able to extract more meaningful features that help distinguish between changing and non-changing areas. This approach has advantages in dealing with high-dimensional data and reducing data redundancy.
  • Correction and radiation normalisation: This paper corrected and radiation normalised the data prior to analysis, which helped to reduce errors due to sensor differences, variations in atmospheric conditions and land cover type, thus improving the accuracy and reliability of the results.

Disadvantages or potential

  • Limitations of Principal Component Analysis (PCA): While PCA is able to extract and enhance variation information from multi-sensor data, it may not be able to completely eliminate noise and other non-target factors. In addition, PCA may have limitations in handling non-linear data, which may lead to inaccurate detection results.
  • Limitations of sensor data: SPOT-5 and Landsat data were used in this paper. While these data sources are of value in land cover change detection, they may not capture the detailed information that finer spatial resolution and higher spectral resolution can provide. In addition, these data sources may be affected by issues such as cloud occlusion, atmospheric interference and timing inconsistencies.

4 Reflection

This week I learnt the principles and methods of correcting, joining and enhancing remotely sensed images and understood how to pre-process remotely sensed data. Although most of the data we acquired had already been pre-processed, this knowledge gave me a better understanding of how to apply remote sensing images to the actual analysis and what could cause errors and whether further processing was needed. In addition, the study of remote sensing image enhancement has taught me more about remote sensing data analysis methods and principles in a deeper way.

However, not all data pre-processing can be perfectly accurate, errors are allowed but need to be controlled according to the actual situation. Also, in practice, some R packages are no longer available due to version updates. This is a common problem in data analysis and can be solved by downloading a lower version of the package or replacing it.