Friday, February 23, 2018

Lab 2

Introduction
     The goal of this was to be introduced to a variety of ways of correcting remotely sensed images for the effects of atmospheric interference. This was done in a variety of different methods. These methods are as follows 1) Empirical Line Calibration (ELC), 2) Dark Object Subtraction, and 3) Multidate Image Normalization. The images for this lab were provided by Cyril Wilson and were collected from Landsat 5 and Landsat povided by Cyril Wilson and were collected from Landsat 5 and Landsat 7.

Methods

Empirical Line Calibration
     The first objective of this lab was to use Empirical Line Calibration (ELC) to clear up atmospheric interference for images of northwestern Wisconsin. To complete an ELC atmospheric correction this equation (fig 1). This equation takes the images separate bands, gain and offset. This equation was calculated using a built-in model in Erdas_Imagine by collecting GCPs. These GCPs were collected on series of different materials, these include asphalt roads, coniferous forests, grass, aluminum roofs, and a lake. The values collected from the GCPs were then compared to values in the USGS V4 and ASTER spectral libraries. Once the GCPs are collected and their spectral profiles compared to that of those in the spectral libraries, the view-preprocessed-atmospheric adjustment tool was used to run regression equations between the results. Once this was completed, a new image was created that corrected for atmospheric interference.

Figure 1. Equation used for Empirical Line Calibration 

Dark Object Subtraction
     The next atmospheric correction method used was Dark Object Subtraction. To complete this method, it needs to be completed in two steps. The first step is to convert satellite image to at-satellite spectral radiance. That is done by using the equation from figure 2. The inputs are collected from the images metadata. This equation needed to applied to each of the six bands for the image (fig. 3). To do this, a model was created in model maker (fig. 4). The model was run to create new radiance images (fig. 5).   
Figure 2. Equation used to convert satellite image to at-satellite spectral radiance 
Figure 3. The at-satellite spectral radience conversion formula being applied to band 1
Figure 4. model created that incorporates the equations for the six different models
Figure 5. Radiance images created using the model in Figure 4.

  After the model was run, the next step was to convert the newly-created at-satellite radiance images to true surface reflectance images (figure 5.). The inputs were collected from the original images meta data and the radiance image histograms (fig 6). To complete this another model was created to implement the new equation (fig. 7)
Figure 5. Equation used to convert at-satellite radiance to true surface reflectance.
Figure 6. Histogram of band 1 of the Landsat TM image. The path radiance for the image is calculated from zero to the start of the histogram
Figure 7. Model used to convert at-satellite radiance to true surface reflectance for each of the six reflective bands


     After the model was run six new images were created and stacked together creating a DOS image (fig. 8.)
Figure 8. Final DOS image after stacking the different bands together. 
Multidate Image Normalization 
     The next part of this lab was to create a atmospheric corrected image for Chicago using Multidate Image Normalization. This method uses two different images collected at different dates and compare their spectral reflective values using regression equations to create a new normalized image. To do this, GCPs are collected for the two images and their spectral signatures are recorded into spectral profiles (fig. 9).
Figure 9a. GCPs collected from various surfaces in the two images
Figure 9b. The spectral signatures collected from the two images
    After spectral signatures were recorded they were then entered in to excel (fig. 10) where regression equations could be created for the each band between the two images after scatter plots were created (fig. 11)
Figure 10. The spectral signatures collected at the different GCPs for the six different reflective bands
Figure 11. Scatter plots created for each of the bands for both the Chicago_2000 and Chicago_2009 images.
     The equations from the scatter plots were entered into a model that created a new normalized each for each band. Once each bands image was created (fig. 12), they were then stacked to make one image that consisted of the different bands (fig. 13).
Figure 12. Normalized images of each of the reflective bands
Figure 13. Layer stacked Multidate Normalized image.
Results
ELC
     Once the ELC image was created the new ELC image was compared to the original uncorrected image (fig. 14). To the naked-eye there are is not a large amount of change between the two images (fig. 14). The ELC corrected image is slightly sharper and is more vibrant in color. To examine the effectiveness of the ELC method, spectral signatures between the two images were taken. When looking at the spectral profile for healthy vegetation in the two images (fig. 15), it can seen that the ELC image has less atmospheric interference in the band 1 (blue band). This is because atmospheric scattering has been removed from the image. Similar trends can be seen when looking at the spectral profile of a local lake (fig. 16).

Figure 14. The two images side-by-side. The original image on the left and the corrected ELC image on the right. Between the two images only a slight difference between the images can be seen as the ELC image is slightly darker. 

Figure 15. Spectral signatures collected for healthy vegetation for the original (left) and ELC corrected image (right).
Figure 16. Spectral signatures collected for lake Wissota in northwestern Wisconsin. The original image is on the left and the corrected image on the right. Decreases can be seen in band 1 in the corrected image as atmospheric interference has been removed. 
DOS
     Much like the ELC image discuessed above, when viewing the original image and corrected image side-by-side (fig. 15) the differences between the images isn't very noticeable but more noticeable than the ELC image. Again, to spectral profiles of the two images were collected. Looking at simailar materials as in the ELC method, we can see similar results when looking at the spectral profiles. Looking at the spectal profiles a decrease in band 1 can be seen again as the effects of atmospheric interference has been removed from the image.
Figure 15. DOS image (right) and original image (left) side-by-side

Figure 16. Spectral profile for healthy vegetation with the DOS image on the left and original image on the right.
Multidate Image Normalization 
     As in the previously discussed atmospheric correction methods the images are shown side-by-side (fig. 17). When comparing the spectral profiles of the two images (fig. 18), a decrease in the reflective signature of band 1 can be seen in the normalized image.
Figure 17. Original image (left) and the corrected image (right) compared side-by-side. 
Figure 18. Spectral profiles of deciduous forest in the two images (normalized profile on the left and original on the right)
Conclusions
     This lab was successful in demonstrating the effectiveness of the different atmospheric correction methods. When using the ELC method, an important factor to consider is that values obtained from the spectral libraries may not be entirely accurate for the image that is being corrected. This is because the material in the image may not be the same material in the spectral library. This difference will lead to inaccuracies in the final product. Of the three methods discussed above, the most effective atmospheric correction method is the DOS method.  This method requires the most information on the actual images to be accounted for.

Monday, February 12, 2018

Visualizing Sandbox Survey

Introduction
     In the previous lab, we collect elevation points for a 114x114 cm sandbox. The data points were collected using a systematic sampling method collecting sample points every 6 cm within the grid. Once the elevation data was recorded it was then transfered into an excel spreadsheet. The data then needed to be normalized to decreased error and ease processing. Normalization refers to cleaning up data so that it uniform and easy to work with. Because a systematic sampling method was chosen for the initial survey, the data was already organized in evenly spaced increments. This allowed for the normalization to be very easily (figure 1).
     The objective for this lab was to take the topology data collected in the previous lab and create 3D topographic profiles in ArcMap and ArcScene. This was done using five interpolation methods 1) spline, 2) IDW, 3) natural neighbor, 4) kriging, and 5) Tin. Each of the methods produce different results when given the same data, because of this each of the interpolation methods will be explained in further detail below.
Figure. 1 Excel sheet displaying the normalized data collected in the previous lab


Methods
     Before the data could be interpolated into 3D topographic profiles, the data needed to be imported into ArcMap. To do this the add X,Y data tool was used create a new shapefile within a newly created feature class in a new geodatabase. The data for this project was left unprojected because the data was not collected using a geographic coordinate system, rather our own coordinate system (see lab 1 for details). Once this was completed the data could then be interpolated. After being interpolated, the models were brought into ArcScene where they were turned into 3D models.
  • Spline: The first interpolation method was spline. Spline uses mathematical estimate values that reduce the curvature of a surface, passing through the center of the data points. This results a profile with a smooth surface. Spline is an effective method when there are a lot of data points but is not optimal if there are few data points as the model tends to over-correct, resulting in a overly simplified profile built upon generalizations. If there are large discrepancies in the elevation of data points that are close together, the model struggles to create realistic profiles. 
  • (Inverse Distance Weighted (IDW): The IDW method estimates cell values by averaging the values of the collected data points using a weighted scale based upon relative distance to the sample point. For example, if a cell is closer to the sample point, it will have a higher weight assigned to that cell in comparison to a cell further away.
  • Natural Neighbor: This method places a strong importance on the sample points themselves and creates regions surronding each point. 
  • Kriging: This method is more complex than the previous methods as it uses formulas to create an estimated surface based upon sample point values. This model takes into consideration the correlations between direction and distance to predict a surface.  
  • Triangulated Irregular Network (TIN): Tin models are created using set of vertices (sample points) to create a triangulated network using Delaunay triangulation. TIN models create high resolution areas where there is high amounts of variability between points and lower resolution models where data points have low variability. 
     Once the models were run, 2D topographic profiles were created, where they were later imported into ArcScene to create 3D topographic profiles.

Results
Spline: The first interpolation method was spline (figure 2). This created a smooth surface that accurately portrayed the surface of the sandbox. Areas in the southwest corner were over generalized as they were more flat in sandbox than the model portrayed. Of the five interpolation methods, this produced the most aesthetically pleasing model.
Figure 2. Spline 3D interpolation model

IDW: The second model used was the IDW interpolation method (figure 3). This model does portray the elevation changes in the sand box quite well, however the model fails to smooth the surface. This produces a model that has unusual looking bumps that make the model appear unrealistic.
Figure 3. IDW interpolation 3D model

Natural Neighbor: The next interpolation method used was natural neighbors (figure 4). This method produced a smooth surface that portrayed the surface accurately. Although the model is smooth, it lacks detail in areas of higher elevation changes than other models used.
Figure 4. Natural Neighbor 3D model

Kriging: The next model run was kriging (figure 5). This model gives a very basic profile of the sandbox. While the general surface of the model gives a general idea of what the topology of the sand box looked like, however the model leaves a lot to be desired in terms of detail.
Figure 5. Kriging 3D interpolation model

TIN: The final method used was TIN (figure 6). This method accurately portrayed the various elevations very accurately and provided an accurate representation of the sandbox. The model however doesn't portray the surface accurately as the triangulations give the model a more jagged look than the topology of the sandbox.
Figure 6. TIN  3D interpolation model

Summary
     Each of the methods used to create the 3D topographic profiles created unique models as each used different methods to achieve their final result. Of the five, the spline method created the most accurate representation of the sandbox. This was because of systematic sampling method combined with of samples relatively small area, allowed for the model to create a very accurate profile. Some of the other sampling methods would have been more appropriate had the study area been larger a more broad sampling method been applied. Interpolation models are not limited only to elevation models, but can be applied to precipitation, temperature, and even air pollution models.

Thursday, February 8, 2018

Lab 1: Surface Temperature Extraction

Introduction
The objective of this lab is to use different skills of extracting land surface temperature. This was done in three main ways 1) visual identification of variabity in  relative surface temperature through surface reflectance using spectral emittance from sensors, 2) qualitatively estimate land surface temperature from thermal bands, 3) building of simple and complex models to extract land surface temperatures from satellite images.


Methods
The lab was divided into four parts using Erdas Imagine and ArcMap. The steps were completed as follows, 1) indentify relative variations in land surface temperature, 2) calculate land surface temperature from ETM+ satellite images, 3) calculate surface temperatures from TM images, and 4) calculate land surface temperatures from Landsat 8 images. To begin the lab we were asked to identify areas of relative warm and cooler temperatures from satellite images taken from Landsat ETM+ . After identifying areas of relative temperatures were then asked to calculate land surface temperature from ETM+ images. This began by converting Digital Numbers (DN) to at-satellite radiance using the formula below (Figure 1.) that was incorporated in a model (Figure 2.). Once DN were converted into satellite radiance the new satellite radiance values were then converted into radiant surface temperature using a formula (Figure 3.) and a simple model (Figure 4.).
Figure 1. Formala used to covert DN to at-satellite radiance.
  • Grescale: amount of gain
  • Brescale: bias in the image
Figure 2. Model used to calculate land surface temperature from ETM+ images incorporating the formula in Figure 1.  

Figure 3. Formula used to convert at-satellite radiance to surface temperature.
  • TB: at-satellite temperature in degrees kelvin
  • Lλ: spectral radiance
  • K1, K2: pre-launch constraints
Figure 4. Simple model used to convert at-satellite radiance to surface temperature. 

Following the calculating land surface temperatures from ETM+, Land surface temperatures were then calculated for a TM using a similar methodology only this time instead of using a simple model a single more complex model was created (Figure 5.). 
Figure 5. Model used to create TM land surface temperatures.
The final task was to calculate land surface temperatures using Landsat 8 imagery using processes described above for the TM image

Results
1) Identification of relative land surface temperature

Figure 6. The image on the right is a  fale color satellite image displaying in the visiable spectrum. The image on the right is a thermal image of the same area. Areas that lighter are areas of relativily high temperature and darker areas are areas of relatively low temperature. 
2) Calculated land surface temperature from ETM+ images

Figure 7. Surface temperature calculated of ETM+ image in Erdas Imagine and then displayed in ArcMap. Areas in red display areas of relativily high temperature.

3) Calculated land surface temperature from TM images
Figure 7. Surface temperature calculated of a TM image in Erdas Imagine and displayed in ArcMap. Areas of dark red indicated areas of relitivly high temperatures
4) Calculated land surface temperature for Landsat 8 images
Figure 8.  Calculated surface temperature for a Landsat 8 image for Eau Claire and Chippewa County created in Erdas Imagine and displayed in ArcMap.
Sources
Landsat satellite image is from Earth Resources Observation and Science Center, United States Geological Survey. Area of interest (AOI) file is derived from ESRI counties vector features.

Lab 7: Object-based Classification

Introduction The purpose of this lab is to be introduced to the relativity new object-based classification scheme. This was done through t...