For our final project in Remote Sensing we were tasked with utilizing the concepts, data types, processes and techniques we had learned over the course of the semester in lecture and lab assignments. I chose to apply remote sensing data and image processing techniques to the suggested project topic of completing a LULC analysis of Lake Tahoe. The question I posed was to do a LULC comparison between provided 1992 data and 1999 data to discover if a change in LULC, specifically permeable to impermeable surfaces had increased dramatically during the specified decade. A rapid increase in impermeable surfaces could potentially be a cause for an increase in fine sediment particles in the Lake Tahoe Basin.
Lake Tahoe is the largest alpine lake in North America. The Lake Tahoe region has undergone rapid urbanization over the last 2-3 decades. As a result of this urbanization, the lake is increasingly becoming eutrophic.
For my LULC comparison I used two provided images. The first image from the MRLC consortium was an unsupervised classification of Landsat Thematic Mapper circa 1990's satellite data (NLCD 1992). The second image was provided by the USGS. Metadata provided for the final project identified the image as a Landsat 7 ETM+ image from April of 1999. However, after performing a multispectral analysis of the image I noticed some anomalies. Displaying the image in a RGB 642 and 653 band combination identified an area north of the lake that was fire. The provided image had 6 layers but layer 6 did not appear to be a true thermal band 6 but rather band 7 from the ETM+ satellite. If layer 6 was a thermal band then the fire feature would not have appeared in the multispectral analysis and areas of snow and ice on the southwest side of the lake would appear darker than other object emitting more energy when viewing Layer 7 in greyscale. Next, I compared the provided image with the original image from the USGS Glovis Visualization Viewer. I could not find an image from the Landsat 7 ETM+ for April 1999. I then discovered through a Google search that the satellite had been launched in April 1999 so the earliest available images began in June of 1999. However, the LandsatLook jpeg image from June 1999 did not visually match the provided image for the project. I then compared the July 1999 LandsatLook image and it visually appeared to be a match to our provided image. This sidestep in analysis made me realize how important accurate metadata is to an enduser using dowloaded images from different data sources. Of course to positively confirm the date of the provide image I would need to download the original and compare brightness values between the provided image and the original.
I performed an NDVI of the 1999 image in order to aid in classification techniques. Because I would be directly comparing the two images I chose to use the same classification legend provided for the NLCD image - 21 Level II classes. A majority of these classes were for vegetative features so having the NDVI would enable me to more easily identify those classes when applying classification techniques to the 1999 image. I performed an unsupervised classification of the image first using 50 classes which I then reclassified using the NLCD 1993 LULC legend. I recoded an additional time to ensure I had at least one wetland class to distinguish from the open water class. In the unsupervised classification groupings of pixels are created based on their brightness values. In the image Lake Tahoe many of the features appear mingled; low and high residential classes were interspersed with forest, transportation and water classes. I performed a supervised classification to see if there would be a difference in classification methods. The supervised classification was performed using signatures I had created from seeds and polygons within the image.
When comparing the two classified 1999 images with the 1992 unsupervised image an increase in impermeable surfaces was seen. Likewise there was a decrease in permeable surfaces over the decade with both classified 1999 images. It was difficult to make precise judgements based on the comparison of the two images. The total coverage area differed between the 1992 and 1999 images even though the 1999 image was clipped to the 1992 image. I think the difference in pixel size between the two images may have caused this difference in coverage area (28.5 m² vs. 30 m²). The 1992 image had 17 classes while my unsupervised classified 1999 image only had 13 of the same classes. Because of the difference in the number of classes being four, I may have misclassified or lumped missing classes into other classes. My supervised classification of the 1999 image clearly exhibits signs of spectral confusion between classes. The region southwest of Lake Tahoe is known as the Desolation Wilderness Area, but in my supervised classified image it appears that the Commercial/Industrial/Transportation class is spectrally confused with Bare Rock, Shrubland and Grasslands classes. Creating the signatures for the supervised classification was difficult, as I previously mentioned features were in close proximity to other classes making it extremely difficult to get signatures that contained enough pixels. When using my signature file I also should have used a RGB 543 band combination as the setting for Approximate True Colors as that band combination may have made classes more distinguishable than the 742 band combination I had chosen.
This project was challenging but enabled the individual course assignments from the semester to be used collectively to answer a question. This course by far has been the most difficult. The sheer volume of information about remote sensing devices and techniques at first seemed daunting, having no significant prior experience in remote sensing and photo interpretation. However, as each lab progressed I realized that I was learning and applying that knowledge to the various assignments. The course ended being very rewarding.
