Maximum Likelihood Classification of the Waterpocket Fold; the Saline Valley; Saudi Arabia - Remote Sensing Application - Completely Remote Sensing, GIS, and GIP Tutorial - facegis.com
Maximum Likelihood Classification of the Waterpocket Fold; the Saline Valley; Saudi Arabia

This Waterpocket Fold scene was part of a 1984 study of how accurately the writer could identify rock units using Landsat TM data from arid (vegetation-poor) terrains. First presented is a Maximum Likelihood Supervised classification map of the scene from the summer of 1981, without Band 6, made on the IDRISI processing system.

 A 14 class (plus black as misclassifications) Maximum Likelihood Classification of the Waterpocket Fold using the 6 TM reflectance bands in the Summer 1984 image; this was made from the IDRISI program.

Its classification accuracy is an estimated 78%; that is, the rock units or formations are correctly identified over at least that percentage of surface exposures. This is high in view of shadow and slope effects, soil cover, and other sources of error. Among misidentifications and conflicts: The Navajo legend color also occurs in the prospect pits and along canyon walls on Tarantula Mesa; black areas such as associated with the localized Carmel/Entrada (combined because of a 15 class limitation) are scattered throughout the scene and generally represent misclassification ambiguities; Kayenta almost disappears in the south except where it shows up incorrectly west of the Chinle (which itself is not well mapped); colors assigned individually to the Morrison/Mancos and the Masuk and the Emery Formations are dispersed within each other. Note too that the Upper Moenkopi and Shinarump Formation were combined into a single training site because the latter is too thin in outcrop area to be mapped discretely at this resolution.

There can be considerable subjectivity in setting up and interpreting any classification. Two more classifications of the Waterpocket Fold are shown below to demonstrate how changing factors lead to modified results. Both were made on the IDIMS system at the Goddard Space Flight Center.

A classification (units not specified) of the Waterpocket Fold 1984 TM subscene made with the IDIMS processing program.

The above classification shows the same summer of 1984 scene but applies a different set of training sites and, unlike the IDRISI one above, includes thermal Band 6. The formations are exceptionally distinct. Red was assigned to alluvium, which, for the most part, was placed correctly (the large section of red occupying the Blue Gate shale member [in blue] of the Mancos formation corresponds to strongly weathered slope wash along steep slopes against Tarantula Mesa). The Dakota formation is a wine purple color (but the white strip near it is of unknown character). The Kayenta is singled out in reddish-brown, and the Chinle is pinkish-peach.

An accuracy assessment of the TM mapping was carried on on classifications using 1) all 7 TM bands, and 2) a 7 PCA image set. The b & W geologic map (page 2-3) was digitized as a comparison base and the classification maps were registered to it. On this basis, the overall accuracy for the 7 band classification was 69.4% and the PCA classification 65.8%. Maximum accuracies were 97, 94, 88, 86, and 85 percents for the Carmel, Moenkopi, Kayenta, Lower Chinle, and Navajo formations respectively, in the TM data set. The Masuk and Tununk Formations had low accuracies in the 20% range.

The illustration below is a classification of a Landsat scene acquired during the winter of 1985. Results are less positive, although many of the units are where they should be. The various formations within the monoclinal limb are discernible but with more irregular outcrop patterns and they include more patches of other units. Note the black patches which are shadows in erosional alcoves. The Moenkopi is not well subdivided into upper and lower areas in this winter version. The light grayish-tan patches along the left side of the image mark snow cover.

 IDIMS classification  similar to the one above but using a subscene of the Waterpocket Fold acquired in the winter of early 1985.

Here is a map of the winter image using the legend units symbols (page 2-3) next to the geologic map for this scene. Overall accuracies of 48.5 and 46.1 percent were obtained for the 7 TM and 7 PCA winter maps respectively. One might suppose that the summer imagery is superior in classifying a geologic scene over one in winter. But there can be a tradeoff: in much of the world, summertime means active vegetative cover not present in winter.

Classification of a winter TM scene of the Waterpocket Fold, compared with the geologic map of the corresponding area.

On the whole, the Waterpocket Fold has proved an exceptional test site. Its field characteristics are especially suited to enhancements like PCA and ratioing. The abundant exposures of rocks rather than soil and vegetation cover make discriminating them easier and hence improve the classification's accuracy over most geological units, particularly those in areas where the units are hidden by weathering products and organic growth. As we have always known for aerial photographs, space imagery has limited value for direct mapping in vegetated terrains, so that field observations requiring on-site inspection and even digging remain the traditional way to map rock units.

2-10: Why aren't rock units (formations) as easy to map in the eastern U.S. as in parts of the West; what might cause mapping problems in desert areas such as in Utah? ANSWER

After this Section was prepared, a new set of very impressive images that deal with rock type identification were released from the on-going Terra satellite program which is the kingpin of the Earth Systems Enterprise (see page 16-6). Although the main objectives of the ESE international effort is to get co-ordinated information about the atmosphere-oceans-biosphere interactions, with geologic studies remaining subordinate, several of Terra's sensors are proving very sensitive to certain characteristics of the land surface that translate into still another approach to stratigraphy and mineral exploration.

This is supported by three images from the Terra's ASTER sensor, shown below as color composites. The area imaged is the Saline Valley which lies to the west of Death Valley and is hemmed in by the Inyo Mountains to its west and the Eureka Range to its east. At the top is a false color composite (15 m resolution) made from ASTER Bands 3 (0.81 m), 2 (0.76 m), and 1 (0.61 m) registered as red, green, and blue (RGB) respectively. In the image, red is vegetation, snow and sand are white, and rocks are gray, brown, yellow, and blue.

ASTER false color composite of the Saline Valley area.

The next ASTER image is made with the SWIR Bands 4 (1.65 m), 6 (2.205 m), and 8 (2.23 m) as RGB. Limestone appears in yellow-greens and rocks altered to kaolinite as one of the clays are purple. Image resolution is 30 m.

ASTER color composite of the Saline Valley scene made with 3 SWIR bands.

ASTER also has several thermal channels. The image below is made from Bands 13 (10.6 m), 12 (9.1 m) and 10 (8.3 m) as RGB. Resolution is 90 m. In this rendition, silica-rich rocks are red, carbonates green, and silica-low (mafic) rocks are purple.

With ASTER and other Terra sensors now on active duty, and hyperspectral sensors (Section 13) becoming spaceborne in the near future, the ability to distinguish (and identify) various rock types has become an effective reality.

However, proper processing of the Landsat MSS and TM imagery can often separate similar appearing rock units as seen in natural and false color. In the top image (false color) below is a part of the Arabian Shield (see page 17-3) in which there are several dark rock types, including basalts, that are hard to separate. But when a color ratio image is made, units at C are separated from those at D, and two plutons at E are now emphasized..

False color image of part of the Arabian Shield in western Saudi Arabia
Color ratio image of the above subscene, in which once similar dark units are here rendered in different colors; this study was done by the late Dr. Herbert Blodget, Goddard Space Flight Center.

Hyperspectral remote sensing, first treated in the Introduction, promises to provide far better capability for identification (classification) of rock lithologies. This will be covered in some detail on pages 13-5 through 13-9.

We now move on from our evaluation of using remote sensing data to distinguish different rock types and where possible assign them to specific stratigraphic units (formations) to a consideration of the structural expressions of geologic units.

Source: http://rst.gsfc.nasa.gov