Friday, April 18, 2014

Lab 6: Geometric Correction

Introduction

In this lab, students will expand on their knowledge of geometric correction by reforming image-to-map and image-to-image rectification on a spatially distorted image. These processes are commonly done to satellite imagery before data extraction or visual analysis is preformed. First, a USGS 7.5 minute digital raster graphic (DRG) of Chicago, IL will be used as a reference image for a slightly distorted satellite image of the metropolitan area. Second, a previously rectified satellite image of Sierra Leone, Africa will be used as a reference image for a heavily distorted satellite image of the same area.

Methods

Image-to-Map Rectification
This method uses a map with an established coordinate system as a reference to modify the spatial location of features in a distorted image to match that of the map.


Figure 1: With the distorted image to be rectified in an active window, navigate to Multispectral > Control Points.


Figure 2: The Set Geometric Model window will appear. For this lab, Select Geometric Model is set to polynomial. Click OK.



Figure 3: Next, the GCP Tool Reference Setup window will appear. For this lab, the default value of Image Layer (New Viewer) is kept. Click OK.



Figure 4: Next, the Reference Image Layer window will appear. Here the reference image will be chosen. For the first part of the lab, the Chicago USGS 7.5 minute DRG is selected. After clicking OK, a window will appear indicating the coordinate reference system of the image. Click OK. Then the Polynomial Model Properties (No File) window will appear. For the fist part of this lab, the Polynomial Order is set to 1, all other default values were accepted and the window was closed.


Figure 5: The Multipoint Geometric Correction window will now appear with both the distorted image and the reference image located within. Each image is portrayed in three different scales. In the upper right, the image is at full scale. In the upper left, the image is zoomed to the extent of the inquire box. Below the two smaller upper windows is a larger window where the image can be manually zoomed and panned. This larger window is where the ground control points (GCPs) will be added.



Figure 6: To add a GCP, select the Create GCP tool on the Geometric Correction toolbar and click an area on the distorted image. Then select the Create GCP tool again and click the same area in the reference image. After a desired number of GCPs have been added, zoom in and reposition the pairs of points to spatially match as close as possible. Continue with this process for all GCPs until the RMS error, read in the bottom right corner of the window, is less than 2 (requirement for lab). For the first part of this lab, four GCPs were collected and the RMS value was reduced to 0.124.



Figure 7: Once all GCPs are added and the RMS error is low enough, select the Multipoint Geometric Correction tool on the Geometric Correction toolbar. The Resample window will appear. For the first part of the lab, Nearest Neighbor is chosen as the Resample Method and all other default values were accepted. A rectified output image is generated.




Image-to-Image Rectification
This method uses a previously rectified image as a reference to modify the spatial location of features in a distorted image to match that of the rectified image.



Figure 8: The process to execute an image-to-image rectification is the same as image-to-map rectification. Because the distorted image is so heavily distorted a third degree polynomial will be used instead of a first degree polynomial like the first part of the lab. In the Polynomial Model Properties, change the Polynomial Order to 3.



Figure 9: Because a third degree polynomial was used, the number of minimum controls points increased from 3 (for first degree) to 9. For the second part of this lab, 10 GCPs were added and the RMS error was reduced to 0.0916.



Figure 10: After selecting the Multipoint Geometric Correction tool again, the Resample Method was set to Bilinear Interpolation. Click OK and a rectified output image is generated.




Results



Figure 11: The resulting image over the distorted image with the swipe tool activated enabling both images to be viewed and compared. The section of river highlighted within the yellow circle showcases the difference between the two images.




Figure 12: The resulting image is the lighter colored image. On the left is the output image over the distorted image, the difference between the two is evident. On the right is the output image over the reference image, the two seem to be spatially identical.



Data Sources
UWEC Department of Geography and Anthropology

Wednesday, April 16, 2014

Lab 5: Miscellaneous Image Functions 2 and Image Mosaic

Introduction

In this lab, students will build on their knowledge of analytical processes in remote sensing by exploring more image processing functions provided in ERDAS IMAGINE 2013. Students will experiment with spatial and spectral image enhancement, band ratio, binary change detection, and image mosaic.

Methods

Spatial Enhancement:
Spatial enhancement techniques will improve the appearance of imagery, mainly visual analysis purposes, by amplifying subtle differences in radiometric or spectral resolutions not perceived by the human eye. This can be accomplished through spatial filtering to adjust the imagery's spatial frequency, which is the change in brightness value per unit of distance for any specific area in the imagery. A low frequency image has few changes in brightness values over the specific area, while high frequency has significant changes. Spatial frequency can be increased or decreased depending on the nature of analysis.


Figure 1: To apply a low pass filter, navigate to Raster > Spatial > Convolution. This will open the Convolution Window seen above. Different filters can be applied to an input image by selecting the desired filter in the Kernel options. For this lab, firstly a 5x5 Low Pass and secondly a 5x5 High Pass filter were chosen. Thirdly a 3x3 Laplacian Edge Detection filter was applied, Fill was checked under Handle Edges by, and Normalize the Kernel was unchecked.



Spectral Enhancement:
Spectral enhancement techniques will improve the appearance of imagery, mainly visual analysis purposes, by increasing the contrast in the image. Low contrast imagery can result from detector saturation or spectral similarity in features. There are two types of spectral enhancement, linear and non-linear. In this lab, linear methods used include minimum-maximum contrast stretch and piecewise contrast stretch. Both this methods stretch the histogram of the image from a low-contrast state to the entire range of brightness values (for 8 bit images 0-255), howver minimum-maximum should be used for Gaussian histograms and piecewise should be used for non-Gaussian. One non-linear method is used in this lab, Histogram Equalization. Instead of stretching the histogram like the linear methods do, histogram equalization will redistribute pixel values so the pixels in the output image are equally distributed across the entire range of brightness values (for 8 bit images 0-255).


Figure 2: To apply a minimum-maximum contrast stretch, navigate to Panchromatic > General > General Contrast > General Contrast. The Contrast Adjust window will appear as seen above. Selecting Gaussian as the method and clicking Apply will apply the min-max contrast stretch.



Figure 3: To apply a piecewise contrast stretch, navigate to Panchromatic > General > General Contrast > Piecewise Contrast. The Contrast Tool window will appear as seen above. For this lab, the From: and To: values for low and middle were determined from the image's histogram and the To: value for high was set to 180.



Figure 4: To apply histogram equalization, navigate to Raster > Radiometric > Histogram Equalization. The Histogram Equalization window will appear as seen above. For this lab, all default values were accepted.



Band Ratioing:
Band Ratioing is considered a non-linear spectral enchantment technique. By applying different ratios to an image, environmental factors can be reduced, unique information can be obtained, and features and objects can be distinguished differently. One commonly used ratio is the normalized difference vegetation index (NDVI) to revel unique information about vegetation layers.



Figure 5: To apply the NDVI band ratio, navigate to Raster > Unsupervised > NDVI. The Indices window will appear as seen above. For this lab, Sensor was set to Landsat TM and Select Function was set to NDVI.



Binary Change Detection (Image Differencing)
Image differencing is used to analyze land cover change from images taken at different dates by subtracting brightness values of pixels in one image from the other. To preform image differencing both images must have almost identical radiometric characteristics, identical spatial and spectral resolutions, and they must by geometrically rectified.


Figure 6: To apply binary change detection, navigate to Raster > Functions > Two Image Functions. The Two Input Operators window will appear as seen above. In this lab, the operator was changed to minus (-) and Layer was changed to Layer 4 instead of All.


Figure 7: Model builder can also be used to preform image differencing. Navigate to Toobox > Model Maker > Model Maker. A window with a new blank model and a smaller window will modeler tools will appear as seen above.



Figure 8: For this lab, two input raster objects were selected and connected to a function which in turn is connected to an output raster object, as seen above. A simplified version of the function created is (the 2011 image - the 1991 image + 127). The constant is added to generate all positive numbers for the resulting image's histogram.



Figure 9: Next, model builder was used again to showcase just the areas that changed over the 10 year period. A raster object was connected to a function which was connected to a raster object, as seen above.



Figure 10: The function for this model is more complicated. Functions was set to Conditional  and Either () IF () OR () OTHERWISE was chosen.  The change/no change threshold value was calculated by multiplying the mean of the image's histogram by 3 standard deviations. Essentially this function will display values of change and mask values of no change.



Image Mosaic:
Image mosaic is used to combine individual images together to create one seamless image. This is necessary when an area of interest is large enough to cover multiple images or the area of interest crosses the boundary of two or more images. When mosaicking images, each image must have the same project coordinate system and have identical numbers of layers.

ERDAS offers two options to mosaic images, Mosaic Express and MosaicPro. For this lab, to add images to be mosaicked in ERDAS, the first image file was opened in a new view but before adding the image Multiple images in Virtual Mosaic was checked in the Multiple section and Background Transparent and Fit to Frame was checked in the Raster Options section. The same procedure is used for following image files.


Figure 11: To use Mosaic Express, navigate to Raster > Mosaic > Mosaic Express. The Mosaic Express window will appear as seen above. For this lab, the images to be mosaicked were added in the Input section and all default values for each other section were kept.




Figure 12: To use MosaicPro, naviagate to Raster > Mosaic > MosaicPro. The MosaicPro window will appear as seen above. To add images click the Add Images icon near the save icon. For this lab, when adding the images, Image Area Options was selected before the images were added and Compute Active Area was checked.



Figure 13: Now that the images are added and their outline is visible on the MosaicPro window, their radiometric properties were synchronized by selecting the Color Corrections icon near the Set Overlap Function icon. The Color Corrections window will appear as seen above. For this lab, Use Histogram Matching was selected and Overlap Areas was set for the matching method by first selecting Set in the Color Corrections window. Next, click on the Set Overlap Fucntion icon and check Overlay. To finish the mosaic, click Process in the MosaicPro window and than Run Mosaic.



Results

Figure 14: Result of the 5x5 Low Pass filter. Using this filter decreased the contrast in the image. As a result, when zoomed in to a large extent features are more blurry and hard to distinguish in the output image (on the right).


Figure 15: Result of the 5x5 High Pass filter. Using this filter increased the contrast in the image. As a result, the output image has a greater range of brightness values specifically darker tones.



Figure 16: Result of the 3x3 Lapcian Edge filter. Using this filter resulted in values of sharp contrast to be highlighted and areas of less contrast to be deemphasized. Rivers and roads become more pronounced.



Figure 17: Result of the historgram equalization. With the histogram equalized, the output image has greater range in brightness values, stretching across the entire available range, giving it more contrast.



Figure 18: Restul of the NDVI band ratio. Appling this band ratio to the image emphasized vegetation over other land covers.


Figure 19: Result of first binary change method, Two Image Functions. The resulting image's histogram has positive and negative values and is marked with its appropriate change thresholds in red.



Figure 20: Result of running the first model. The resulting image's histogram has only positive values because of the constant that was added to the function equation. Here the change threshold is only located on the upper end.



Figure 21: Map made in ArcMap using the result from the second model. Running the second model resulted in the red areas shown on the map. By bringing that image and the original image into ArcMap and changing the symbology accordingly, the map seen above was created. When comparing this map to the original imagery, it can be inferred that the areas of change correspond to changes in land use, specifically agriculture.



Figure 22: Result of Mosaic Express. This result is not good. The boundary between the images is very evident when ideally it would be seamless.



Figure 23: Result of MoscaicPro. This result is much better than the one generated through Mosaic Express. The boundary between the images is much more seamless and it is difficult to tell if the boundary exists in much of the output image.



Data Sources
UWEC Department of Geography and Anthropology