Image interpretation and analysis Grundlagen Fernerkundung, Geo 123.1, FS 2014 Lecture 7a Rogier de Jong Michael Schaepman Why are snow, foam, and clouds white?
Why are snow, foam, and clouds white? Today s topics Book: Chapter 7, p. 482 623 Visualization and contrast Spectral analysis Band combinations and indices Image transformation Temporal analysis Change detection Spatial analysis Filtering (high-pass and low-pass) Pixels and objects
Visualization and contrast manipulation A sensor measures energy and has a certain dynamic range (DR) The lowest amount of energy that can be detected is recorded as 0 The highest amount of energy that can be detected is recorded as 255 Histogram stretching DR sensor (0-255) DR display (0-255) DR sensor (60-158) DR display (0-255) Lillesand Fig. 7.13
Histogram stretching Lillesand Fig. 7.13 Contrast
Level slicing Discontinuous color mapping Subdividing the continuous range of values into discontinuous but sequential groups (called bins or classes) Simplest way of making classes based on spectral values 0 10 aquamarine 11 50 sienna 51 100 dark green 101 255 color scale light green to white Continuous color mapping 0-255 = 256 values 256 values = 2 8 = 8 bit 00 00 00 00 11 11 11 11 Example: 01 10 01 01 =>
Band combinations Each spectral band represents a grey-scale image. Three of these bands can be assigned to the display colors red (R), green (G) and blue (B) to obtain a full-color image. 3 bands, 0-255 each = 24 bit color depth (16.7 million colors) blue green red True-color image B G R
Color-infrared image R B G Red = Near IR (band 4) Green = Red (band 3) Blue = Green (band 2) False-color images using 7 Landsat bands Landsat bands 1 0.45-0.52 µm Blue 2 0.52-0.60 µm Green 3 0.63-0.69 µm Red 4 0.76-0.90 µm Near IR 5 1.55-1.75 µm Mid-IR 6 10.40-12.50 µm Thermal IR 7 2.08-2.35 µm Mid-IR
Spectral analysis R,G,B = R,G,B R,G,B = NIR,R,G Lillesand Plate 3 Spectral analysis R,G,B = R,G,B R,G,B = SWIR,NIR,G
Band operations Per-pixel mathematical operations and transformations, where equation variables are represented by various spectral bands Simple band ratios HOTSPOT" DARKSPOT" Minimize topographic effects on spectral information." The DN ratio for a sunlit pixel is nearly identical to the DN ratio for a shadowed pixel." " Lillesand Fig. 7.25
Maximizing spectral contrast Color-infrared (CIR) images provide the best contrast for vegetation (lecture 2) Why? R = Near IR G = Red B = Green Vegetation in visible versus near infrared
Maximizing spectral contrast Green, dense For vegetation:" Brown or sparse Reflectance is highest in NIR Red Infrared Red Infrared and lowest in RED" Both change if vegetation changes (senescence animation in previous lecture)" The ratio NIR / RED gives therefore the highest contrast and the best way to quantify changes" " Spectral index A mathematical combination of spectral regions (or bands) is known as a spectral index. For vegetation, these indices represent greenness in terms of photosynthetic activity (not greenness in terms of color) Many other spectral indices exist, for example for soils/geology (texture, iron content, soil organic carbon), for snow and ice (grain size, contamination), for water (chlorophyll content, dissolved organic matter) and for the atmosphere (NO 2 concentration, aerosols, ozone)
Spectral index RED Color-mapped INDEX NIR 1 Southern France (Cabrieres, Languedoc) 0 Data transformations Aim at reducing the spectral redundancy (i.e. correlation between spectral bands) Compress the information content in fewer bands with decreasing variance Common example: Principal Component Analysis (PCA) Lillesand Fig. 7.27
Data transformations The resulting bands can be used for visual interpretation or as input for classification Apex, true color PC1 PC2 PC3 PC25 PC50 Temporal analysis Detecting changes is the first step towards assigning causes
Change detection Example: Mapping of forest / plantation disturbances (red = disturbed) 2002 2003 Google Earth, 2014 2004 2005 Temporal change: Aral Sea
Temporal change: Rondonia, Brazil 2000-2008 Temporal change: Dubai 2002-2011
Types of change Short-term change e.g. weather events Cyclic change e.g. seasonality Sustained change e.g. urban expansion (Dubai) Multidirectional change e.g. drought stress and recovery Change detection techniques Image differencing Calculate the difference or the ratio of two remotely sensed images. Regions that differ from 0 (difference) or 1 (ratio) have changed. Post-classification comparison Use same classification algorithm for two or more dates and detect pixels that are assigned to a different class
Change detection techniques Changed Vector Analysis Calculate the vector between the spectral response of a pixel at the initial date versus following date(s). Long vectors indicate large change. Composite image analysis Apply a classification algorithm or a transformation on an image stack of multiple dates. Pixels with similar changes show up in the same class. Lillesand Fig. 7.59 Change detection techniques Image regression Regress spectral bands or indices at date 1 versus date 2. Pixels that deviate from the 1:1 line have changed. Temporal regression For time series, changes occurred if the regression coefficient of observations against time is unequal to zero Lillesand Fig. 7.61
Image cube: spectral profiles for one pixel reflectance ~ wavelength 284 bands (Apex) 288 bands (Modis VI) spectral index value ~ time
Trend breaks Temporal regression: sustained and multidirectional Observation (one pixel) 1000 1000 3000 5000 Seasonality Changes 200 400 600 800 β = 21.834 p = 0.004 Y = ax + b gradual increase β = 12.617 p = 0.367 m = 234.773 fast decrease stable β = 6.680 p = 0.024 1985 1990 1995 2000 2005 2010 Mapping the change Y = ax + b Green increase (a > 0) Red decrease (a < 0)
Spatial analysis Spatial Image Filtering Filtering manipulates the image elements Filters are applied using kernels composed of size and weights Most common in image analysis: high-pass and low-pass filters
High-pass filter (sharpening) Used to enhance high-frequency variations Disadvantage: enhances noise as well Kernel typically has high central value surrounded by (partially) negative weights; sum is 0 or higher Special case: Laplacian (2 nd derivative) filter 0-1 0-1 5-1 0-1 0 High-pass filter 14 28 13 9 12 12 8 7 8 3x3 kernel 0-1 0-1 5-1 0-1 0 4 (0*14) + (-1*28) + (0*13) + (-1*9) + (5*12) + (-1*12) + (0*8) + (-1*7) + (0*8) = 4
Influence of kernel size (High-pass filter) Radius 1, 10, 50 pixel(s) Low-pass filter (smoothing) Used to suppress high-frequency variation and noise Kernel with small positive values. Simplest case: all values (1 / kernel size) and thus sum equals 1 Kernel size (3x3, 5x5, ) determines degree of smoothing Special case: Gaussian filter (or Gaussian blur ) 1/9 1/9 1/9 1/9 1/9 1/9 1/9 1/9 1/9
Low-pass filter 14 28 13 9 12 12 8 7 8 3x3 kernel 1/9 1/9 1/9 1/9 1/9 1/9 1/9 1/9 1/9 12.33 (1/9*14) + (1/9*28) + (1/9*13) + (1/9*9) + (1/9*12) + (1/9*12) + (1/9*8) + (1/9*7) + (1/9*8) = 12.33 OR 1/9 * (14+28+13+9+12+12+8+7+8) = 12.33 Low-pass example: de-noise noise 5% low-pass 30% 50%
Low-pass example: extract large objects original image 15 x 15 low-pass thresholding High-pass and low-pass filters, overview High-pass 0-1 0-1 5-1 0-1 0 Low-pass 1/9 1/9 1/9 1/9 1/9 1/9 1/9 1/9 1/9
Pixels versus objects Statistical approaches can be used to group pixels based on spectral similarity and on neighborhood criteria These segmentation techniques transform per-pixel images into discrete pixel groups Lillesand Fig. 7.57 Pixels versus objects Advantage: Close to the human ability to distinguish spatial relationships Disadvantage: Models require scale and shape parameters and therefore strongly depend on user choices The segmented image can be used for classification (lecture 12) Lillesand Fig. 7.57
To take home from this lecture Understand that the choice of analysis technique depends on spectral, spatial and temporal resolution Know which flavors of image-analysis techniques exist without understanding technical details Be able to select an approach given a problem and to provide example problems given an approach Thank you! Next week: image classification Grundlagen Fernerkundung, Geo 123.1, FS 2014 Lecture 7a Rogier de Jong Michael Schaepman