Methods

In order to determine how these two methods compare and whether NDVI is sufficient for research purposes for this regional environment, I acquired six images - two scenes each over three years - and processed each image using both SMA and NDVI techniques. The six scenes were all virtually cloud-free. The dates were 1972, 1976, 1980 during the months of September and October, peak months for agricultural vegetation.

Data Quality Check:

Because the three years were each acquired by different satellites, the data quality varied significantly. Data quality can be checked by applying strong stretches to the data and examining spectra of known material, as well as visually examining the image for missing data lines or rows. Surprisingly the data acquired in 1972 was the best quality, while data acquired in 1980 was missing a large area of data as well as scattered lines throughout the scene. It was important to determine these data quality complications early so that they could be accounted and corrected for throughout the image processing procedures.

Georegistration:

All images were georegistered according to the information provided with the data from NASA. This step allows for specific pixels to be correlated to specific geographic areas on earth. Later this information can be integrated with geocoded statistical data or georegistered political lines and boundaries.

Radiometric Alignment:

Radiometric alignment is vital to any change analysis. Variations in radiance values detected by a sensor change slightly due to changes in atmosphere, solar elevation, azimuth and sensor performance (Elmore et al., 2000). These differences are even more acute when data is acquired by different sensors. In order to compare values of reflectance from one scene to another, it is necessary to align the DN values of each scene to a reference scene.[9] In this study 1976 was used as the reference year.

Ground Invariant Points (GIP's) were selected between scenes. These are areas that should remain spectrally constant over time and therefore should be represented by equal values in all scenes. Such areas generally include deep reservoirs, and pure sand piles or soil with little human or natural disturbance or vegetation. It is important to use a wide range of intensities for improved accuracy.

Each band in the 1972 and 1980 images was regressed against its corresponding band in the 1976 image based on the values of various GIP’s. A line was fit to the points and the constant and slope of the best-fit line were used to adjust the 1972 and 1980 bands. (For the calibration equations used see Appendix I.)

This first step analysis aligns all values in all scenes so that they are comparable over all three years. Because this was determined through a best-fit line of GIP's, some error does exist. The determination of GIP's is difficult in scenes with high vegetation abundance, dynamic water bodies, and low resolution. In some cases it may be that part of the differences in radiance between points over the various scenes may reflect real changes and not differences in sensor calibration/sensitivity. While this could slightly change overall vegetation change results over time, because both NDVI values and SMA fractions were determined from these calibrated images, this error should not be reflected in the differences in the results of these two methodologies.

Co-registration:

The spatial accuracy of Landsat's georeferencing information is approximately one kilometer. However, for vegetation change detection purposes accuracy is needed on the subpixel level (<57m). Therefore, in order to overlay images of various dates, all images must be co-registered. Approximately 80 ground control points (GCP’s) were found between the 1972 and 1976 scenes as well as between the 1976 and 1980 scenes. These points were then used to warp the 1972 and 1980 images to be spatially co-registered to the 1976 images with approximately less than one pixel root mean squared (RMS) error.

The largest source of potential error is during co-registration of multi-temporal images. While the RMS error was minimized and over 80 GCP were used, some registration errors do occur. Errors in co-registration mean that some pixels may be incorrectly aligned between dates and change values many not be accurate. Again, because NDVI and SMA were both performed on the same co-registered images, this error should not be reflected in the results comparing the two methods.

Spectral Mixture Analysis:

Spectral Mixture modeling involves the separation of pixels into classes of specified endmembers. Endmembers are spectral signatures extracted from a “pure pixel” of various land covers, which together represent the range of spectral variability within a scene. Since MSS has only four bands, a maximum of three endmembers may be used. In these scenes endmembers included light sand/soil, shade, and vegetation. The final endmembers were determined through trials and comparisons of various candidates (see Figure 5 for the spectral signatures of the endmembers used). The values or fractions for each endmember in the resulting data generally lie between 0 and 1 and the RMSE is nearly uniform, with virtually all pixels having less than a 2% error.

In the northern scene, patches of basalt (Guha 1999) were so dark that the spectra were virtually identical to deep reservoirs over the visible and near infrared wavelengths. Therefore, water, shade and dark soil are used interchangeably for this scene. Because many soil types are present over the large area covered by these scenes, choices in the soil endmember were made based on the priorities of the study. In the northern scene, sandy soils are abundant, and a sand endmember would be useful in evaluating desertification effects. In the southern scene where forest cover is more abundant, a localized, relatively darker soil was useful in distinguishing forested areas for the purposes of forest change analysis. Pure pixels of each endmember were found in the scene through a thorough analysis of the spectral signatures of highly vegetated areas, sandbars for sand, bare soil patches, and deep reservoirs for shade. This process is especially challenging in semiarid or densely populated and disturbed environments.

Shade, as an endmember, is used to account for illumination effects rather than for a physical land cover type. Because deep, clear reservoirs generally reflect very little in all bands, its spectral signature is a good indication of atmospheric scattering and therefore is a good measure of shade illumination. Therefore, this endmember is used as more of a neutral multiplicative scaling factor than an actual measure of land cover material (Mustard & Sunshine, 1999; Smith et al., 1990 a&b).

NDVI Generation:

The methods for NDVI data generation are identical to the methods for SMA for co-registration, georeferencing, and radiometric alignment. Because NDVI requires data to be converted from DN values to reflectance, scattering must be removed and divided by solar illumination. Scattering effects are determined by finding the darkest pixel in the scene, usually a deep reservoir, and determining the difference between the DN value of the infrared band (Band 4 is used in Landsat MSS data) and the visible red band (Band 2 in MSS data). Because water does not radiate much light at wavelengths over the visible through infrared spectrum, light detected by the satellite in these bands is generally due to path radiance, and not radiation from the water itself. However, atmospheric scattering is only significant in the lower wavelengths, primarily the visible bands. Therefore, the scattering over the wavelengths in Bands 1 and 2 is far greater than the scattering in Bands 3 and 4. In this study, because the data was already spectrally aligned, an estimate of scattered light was applied to all scenes equally. The NDVI algorithm was run and a new dataset was generated. The algorithm used was:

NDVI= (band 4 - band 2 - scattering)

(band 4 + band 2 - scattering)

Change data generation:

In order to overlay images of varying sizes and aerial coverage, each registered image was cut to the maximum area of overlap for all three years. This was considerably smaller than the original extent of the various images. The final size of images in the northern scene was 1934 samples (110.2 km) by 2974 lines (169.5 km). The southern scene’s final size was 1923 samples (109.6 km) by 2954 lines (168.4 km). Upon cutting the images to equal sizes over identical areas, the vegetation data was subtracted for each year in order to generate vegetation change statistics over the eight year time period.

Back to Index

For more information please contact Laurel_Firestone@yahoo.com