Mod. 2 p. 1. Prof. Dr. Christoph Kleinn Institut für Waldinventur und Waldwachstum Arbeitsbereich Fernerkundung und Waldinventur

Transcription

1 Histograms of gray values for TM bands 1-7 for the example image - Band 4 and 5 show more differentiation than the others (contrast=the ratio of brightest to darkest areas of a landscape). - Judging from these histograms: which band has highest mean brightness and which highest variance?, Georg-August-Universität Göttingen Slide No 1 - Obviously, for a given image the bands do not utilize the entire available range of possible DN. - However, the more similar the brightness values, the more difficult is is to distinguish different features. - Which values in this table give a good comparative idea of the contrast present?, Georg-August-Universität Göttingen Slide No 2 Mean and variance are univariate statistics (for one band). Histogram: univariate descriptive statistics. We are also interested in a simultaneous analysis of more than one band. Scatterplot: bivariate descriptive statistics: two bands simultaneneously depicted. 3D scatterplot: three bands simultaneously. What can be read from these graphs?, Georg-August-Universität Göttingen Slide No 3 Mod. 2 p. 1

2 Scatterplot Land cover classes have typical reflection patterns in different bands. A two-dimensional scatterplot helps defining rules to distinguish those classes: basis for automatic classification. The example given is an ideal case where classes are clearly separated., Georg-August-Universität Göttingen Slide No 4 As we observe in one image the same ground features in different spectral bands, we may expect that the DNs for the different bands are correlated. Scatterplots do also give an idea of the correlation between bands in multispectral imagery, Georg-August-Universität Göttingen Slide No 5 Example for a correlation matrix for TM spectral bands Interpretation? Select for analysis (and also for visualization by for RGB color composite) those bands that carry most information., Georg-August-Universität Göttingen Slide No 6 Mod. 2 p. 2

3 Principles of contrast stretching - A sensor needs to be able to record high (e.g. snow) and low brightness (water in IR) in each band. - However, for one specific region / image, the actual range utilized is much narrower than the (of the possible range of for a 8 bit image). - Contrast stretching stretches the observed range to the complete possible range and makes brightness differences better visible. - Also, many landscape features generate similarbrightness values within single bands (i.e. low contrast): stretching makes these small differences better visible. Objective: enhance the raw image so that it is better suited for a specific purpose., Georg-August-Universität Göttingen Slide No 7 Principles of contrast stretching - All methods intend to use the whole dynamic range of a video display (e.g ). - For each band per pixel a new brightness value must be calculated. - Input information for this calculation: - Minimum and maximum brightness value (BV min and BV max ) for the band. - Raw brightness value (BV raw )for the pixel. - Range of gray levels of the video display (V Range ). Algorithm for the min-max linear contrast stretch (calculating a new brightness value BVstretched for each pixel): BV stretched BV = BV raw max BV BV min min V Range, Georg-August-Universität Göttingen Slide No 8 - If the distribution of brightness values has long tails with low frequencies, the stretch by the min/max technique may result to be very little only. - Therefore, one may tail clip the distribution of the brightness values, cutting, for example, the 5% lowest and highest values, and apply then the stretch to the remaining core -distribution = 5% saturation linear contrast stretch., Georg-August-Universität Göttingen Slide No 9 Mod. 2 p. 3

4 Illustration of 2 linear contrast stretch algorithms, Georg-August-Universität Göttingen Source: Wilkie & Slide Finn No Nonlinear contrast stretching Using a non-linear stretching function Histogram equalization: attempts to equalize the number of pixels in each brightness level in order to maximize contrast. Gaussian stretching: similar basic idea like equalization, but the target is to achieve a stretched brightness value distribution which resembles a normal distribution where tails are clipped to ±2 or ±3 standard deviations., Georg-August-Universität Göttingen Slide No 11 Illustration of 2 non-linear contrast stretch algorithms Source: Wilkie & Finn 1996, Georg-August-Universität Göttingen Slide No 12 Mod. 2 p. 4

5 Illustration of changes in histograms by contrast stretching..., Georg-August-Universität Göttingen Slide No 13 Correcting for sensor errors A sensor consists of a series of single detectors which record the radiance for a given ground area (GIFOV) simultaneously (16 detectors in Landsat TM, 3000 in SPOT multispectral). Optimally, all such detectors are calibrated such that they register in exactly the same manner. If one detector fails, there will be regularly spaced black lines in the image (line drop). Then, one may interpolate a brightness value for each missing pixel from brightness values of the neighboring pixels. Correcting miscalibrated detectors is more complex and is done with the so-called transfer characteristics of the detectors The translation of radiation intensity to recorded brightness values should be the same for all detectors which is checked by comparing the detectors transfer characteristics. Salt and pepper effect (speckles): random pixels throughout the image are erroneously dropped (0) or saturated (255). After identification of those pixels: Correction by a filter: moving window algorithm., Georg-August-Universität Göttingen Slide No 14 Effect of atmosphere and topography The same feature at the earth surface may be recorded with different spectral and radiometric characteristics due to the atmospheric and topographic situation: Haze may blur the image and reduce contrast. The sun angle together with topography changes the intensity of reflected radiance: the same forest type looks different on a shadow slope and in a fully illuminated situation., Georg-August-Universität Göttingen Slide No 15 Mod. 2 p. 5

6 Influence of the atmosphere on radiance Only a part of the radiation that arrives at the outer atmosphere comes directly to the earth s surface A part of the radiation is scattered in the atmosphere and arrives as diffuse light. The portion of this diffuse light is between 10 and 100%, depending on the composition of the atmosphere (clouds, water vapor, aerosols), particularly in the range of shorter wavelengths. Sections of the spectrum with a good transmission of radiation are called atmospheric windows. Diffuse light makes that shadows are less dark (positive for remote sensing), but reduces contrast (negative for remote sensing)., Georg-August-Universität Göttingen Slide No 16 Scattering in atmosphere Necessary because of light scattering off - Air molecules (Rayleigh scattering); On a clear day, Rayleigh scattering makes the sky appear blue. - Particles associated with smoke and dust (0.1-10µm) (= Mie scattering), Affects also shorter wavelength, but lesser. Clouds, haze and fog appear white or blueish-white, because larga partilce scattering is not wavelength specific! - Water droplets (>10µm; nonselective scattering); causes object to lose color and contrast. It reduces detail and our ability to detect useful features in the imagery., Georg-August-Universität Göttingen Slide No 17 Atmospheric correction Atmospheric correction means to adjust pixel by pixel the brightness values such that the atmospheric disturbance is removed. If haze is uniform over the entire image one applies the atmospheric correction uniformly to the entire image. If not so, one needs to identify images areas affected by atmospheric disturbances and apply the corrections only to those. In general, atmospheric corrections improves the color contrast of an image markedly., Georg-August-Universität Göttingen Slide No 18 Mod. 2 p. 6

7 Two basic techniques of atmospheric correction Haze Removal: Rayleigh and Mie scattering affect short wavelength the most, so that we should expect that the visible blue and green and are shifted further from zero than the red and IR bands. To correct these errors we must determine for each band the histogram shift and then subtract this amount from the brightness level of each pixel. Using zero-brightness areas: Selecting an area of water or homogeneous shadow (near zero brightness) area and regress the brightness values of a visible to an IR band. Without atmospheric disturbances, the regression should pass through the origin. If this is not the case, then each pixel in the visible band should be adjusted by the amount of intercept shift., Georg-August-Universität Göttingen Slide No 19 Mod. 2 p. 7