L8: Digital Image Processing PDF

Module M.SIA.I14M GIS and Remote Sensing for Agriculture L8: Digital Image Processing Dr Jayan Wijesingha [email protected] +49 561 804-1245 What happen after image capture? Data transmission to the receiving station Directly...

Module M.SIA.I14M GIS and Remote Sensing for Agriculture L8: Digital Image Processing Dr Jayan Wijesingha [email protected] +49 561 804-1245 What happen after image capture? Data transmission to the receiving station Directly from the satellite Via communication satellite Data management at the receiving station System corrections Coarse geometric correction (orbit variations, earth curvature, etc.,) Quick-looks Precision correction to map reference system Data distribution to agent/users 2 What user need to after getting image data? - I Pre-processing Radiometric correction DN  Radiance  Reflectance Atmospheric correction Topographic correction Geometric correction Georeferencing 3 What user need to after getting image data? - II Image enhancement Colour composite Contrast enhancement Multispectral transformations Vegetation indices Filtering Image classification Quantitative analysis of image properties Post processing 4 DN  at sensor radiance (toARad) - I Two methods Method is depends on the meta data provided by the data provider Method – 1 Using “Gain” and “Bias” 𝐿𝐿𝜆𝜆 = 𝑔𝑔𝑎𝑎𝑖𝑖𝑛𝑛 × 𝐷𝐷𝑁𝑁 + 𝑏𝑏𝑖𝑖𝑎𝑎𝑠𝑠 Lλ = At sensor radiance (W sr-1 m-2 µm-1) 5 DN  at sensor radiance (toARad) - II 5.9206E-03 -29.60312 If the DN value for Landsat 8 band 5 is 15300; The ARad is; 𝐿𝐿𝑏𝑏5 = 5.9206 × 10−3 × 15300 + (−29.60312) = 60.98 𝑊𝑊 𝑠𝑠𝑟𝑟−1𝑚𝑚−2 𝜇𝜇𝑚𝑚−1 6  DN  at sensor radiance (toARad) - III Method – 2 Longer method with more parameters 𝐿𝐿𝑀𝑀𝐴𝐴𝑋𝑋𝜆𝜆 − 𝐿𝐿𝑀𝑀𝐼𝐼𝑁𝑁𝜆𝜆 𝐿𝐿𝜆𝜆 = × 𝐷𝐷𝑁𝑁𝜆𝜆 − 𝑄𝑄𝐶𝐶𝐴𝐴𝐿𝐿𝑀𝑀𝐼𝐼𝑁𝑁 + 𝐿𝐿𝑀𝑀𝐼𝐼𝑁𝑁𝜆𝜆 𝑄𝑄𝐶𝐶𝐴𝐴𝐿𝐿𝑀𝑀𝐴𝐴𝑋𝑋 − 𝑄𝑄𝐶𝐶𝐴𝐴𝐿𝐿𝑀𝑀𝐼𝐼𝑁𝑁 LMAX λ = spectral radiance scales to QCALMAX LMIN λ = spectral radiance scales to QCALMIN QCALMAX = the maximum quantized calibrated pixel value QCALMIN = the minimum quantized calibrated pixel value 7 DN  at sensor radiance (toARad) - IV 358.40497 -29.59720 65535 1 If the DN value for Landsat 8 band 5 is 15300; The ARad is; 358.40497 −(−29.59720) 𝐿𝐿𝑏𝑏5 = × 15300 − 1 + 65535 −1 −29.59720 = 60.98 𝑊𝑊 𝑠𝑠𝑟𝑟−1𝑚𝑚−2 𝜇𝜇𝑚𝑚−1 8 At sensor radiance  At sensor reflectance (toARef) - I Assumes Lambertian surface Sensor Zenith angle is changing with location and time d Distance Sun Earth measured by using θ Astronomical units (AU) δ Surface 9 At sensor radiance  At sensor reflectance (toARef) - II 𝜋𝜋 × 𝐿𝐿𝜆𝜆 × 𝑑𝑑2 𝜌𝜌𝜆𝜆 = 𝐸𝐸𝑒𝑒𝑥𝑥𝑜𝑜 × cos 𝜃𝜃 ρλ = At-sensor reflectance [unitless] 𝜋𝜋 = Mathematical constant equal to ~3.14159 [unitless] Lλ = Spectral radiance at the sensor [W sr-1 m-2 µm-1] θ = Solar zenith angle [degree] δ = Solar elevation angle [degree] d = Distance Earth-Sun [AU] Eexo = Exoatmospheric irradiance [W sr-1 m-2 µm-1] 10 Solar angles Solar zenith angles (θ) Summer Spring/ Autumn Solar elevation angles Winter (δ) 11 Earth-Sun Distances (d) Distance Sun-Earth (www.wikipedia.com) http://disc.gsfc.nasa.gov/julian_calendar.sh tml  At sensor radiance  At sensor reflectance (toARef) - III 148.94819400 58.90567162 1.0161872 If the ARad for Landsat 8 band 5 is 60.98 W sr-1 m-2 µm-1 The ARef is; 3.14159 ×60.98 × 1.01618722 𝜌𝜌𝑏𝑏5 = = 0.2212 = 22. 1 % 1044 × sin(58.90567162) 13  At sensor radiance  At sensor reflectance (toARef) - IV Landsat 8 collection 1 data DN values can be directly converted to the ARef values 𝜌𝜌ƴ 𝜆𝜆 = 𝑔𝑔𝑎𝑎𝑖𝑖𝑛𝑛 × 𝐷𝐷𝑁𝑁 + 𝑏𝑏𝑖𝑖𝑎𝑎𝑠𝑠 𝜌𝜌ƴ 𝜆𝜆 𝜌𝜌ƴ 𝜆𝜆 𝜌𝜌𝜆𝜆 = = cos 𝜃𝜃 sin 𝛿𝛿 https://www.usgs.gov/landsat-missions/using-usgs-landsat-level-1-data-product 14 At sensor reflectance  Surface reflectance (SRef) Three ways to correct for atmospheric effect: Empirical correction methods (Flat-field, IARR, Empirical line) Use of radiative transfer codes (i.e. model the atmosphere’s optical behaviour) (e.g. ATREM, ATCOR, FLAASH …) In-flight calibration (simultaneous detection of incoming irradiance and reflected radiance) Output of empirical approaches is the reflectivity measured relative to a standard target from the scene (relative reflectance) Output of other methods result in absolute reflectance values 15  Topographic correction Strong influence of the topography on spectral signal in rugged terrain Causes problems for a subsequent scene classification 18 Gupta, S.K., Shukla, D.P. Evaluation of topographic correction methods for LULC preparation based on multi-source DEMs and Landsat-8 imagery. Spat. Inf. Res. 28, 113–127 (2020). https://doi.org/10.1007/s41324-019-00274-0 Geometric distortions Geometric distortions can occur due to The earth rotation Viewing angle Earth curvature Scanning time Platform position Non-linear mirror speed Topography 19 Geometric corrections With geometric model of satellite / orbit / earth With control points and polynomials Georeferencing 20 Georeferencing - I Two transformation equations needed to transform pixels in a distorted image (row i and column j) to the right coordinates in the map (x and y) i =f1(x,y) and j =f2(x,y) Equations are „empirically“ calculated with statistical models Position of GCP‘s are determined in the image and in the geometrically correct reference (e.g. map) Data pairs are used for transformation 21  Georeferencing - II Quality of accordance between initial and corrected image is calculated with Root Means Square Error (RMSE) 𝑖𝑖𝑚𝑚𝑎𝑎𝑔𝑔𝑒𝑒 GCP 2 GCP 1 Δ𝑥𝑥𝑖𝑖 = 𝑥𝑥𝑖𝑖 − 𝑥𝑥𝑖𝑖𝑎𝑎𝑐𝑐𝑡𝑡𝑢𝑢𝑎𝑎𝑙𝑙 GCP 1 GCP 2 Δ𝑦𝑦𝑖𝑖 = 𝑦𝑦𝑖𝑖𝑖𝑖𝑚𝑚𝑎𝑎𝑔𝑔𝑒𝑒 − 𝑦𝑦𝑖𝑖𝑎𝑎𝑐𝑐𝑡𝑡𝑢𝑢𝑎𝑎𝑙𝑙 Δ𝐿𝐿𝑖𝑖 = Δ𝑥𝑥2 + Δ𝑦𝑦2 GCP 4 𝑖𝑖 𝑖𝑖 GCP 3 GCP 4 GCP 3 σ 𝑁𝑁𝑖𝑖=1 Δ𝐿𝐿𝑖𝑖 2 R𝑀𝑀𝑆𝑆𝐸𝐸 = Initial Image 𝑁𝑁 Geometrically corrected image  Georeferencing – III Rotation Offset Scaling Skewness 2 GCP 1 GCP More than 2 GCP The more GCP the better GCP should have the following properties: High contrast in the image Small size Temporally stable Height differences should not be too big Georeferencing - IV Artificial Elements normally most reliable (street crossing, field corners, corners of buildings…) GCPs should be spread over the entire image to even out local distortions Interpolate new pixel values – resampling Fill the new image matrix with values from the old image. 3 common methods 24 Radiometric interpolation Geometric errors cause a shift in pixel values during image recording  influences radiometric information Transformation of Pixels to their geometrically correct positions requires radiometric interpolation Methods for radiometric interpolation (Resampling): Nearest Neighbour Interpolation Bilinear interpolation Cubic interpolation Contrast stretching 26 Colour composite Satellite images are usually multi-band images But in computer we can only view 3 bands, because computer has only three primary colour channels which are red, green, and blue Different band combinations can be assigned to each channel and enhance the features https://gsp.humboldt.edu/OLM/Courses/GSP_216_Online/lesson3-1/composites.html in the image 27 Colour composite True colour composite False colour composite (TCC) (FCC) Satellite image red  Satellite image nir  Computer red Computer red Satellite image green Satellite image red   Computer green Computer green Satellite image blue Satellite image green  Computer blue  Computer blue 28 Spatial feature manipulation Spatial filtering – Emphasize or deemphasize image data of various spatial frequencies Spatial filtering is a local operation that pixel values of the original image is modified based on values of the neighbouring pixels Used moving window (AKA ‘Kernel’) to modify pixel values Window size can be 3 x 3 or 5 x 5 or …. 9 x 9... or bigger Based on the values in the window different features in the image can be enhanced Convolution filter Edge enhancement 29 Spatial filtering Lillesand, T. M., Kiefer, R. W., & Chipman, J. W. (2003). Remote sensing and image interpretation (Fifth Edit). Wiley. 30 Spatial filtering 31