CLASSIFICATION OF CRYOSPHERIC IMAGES


We have now examined some of the remote sensing techniques used in our attempts to comprehend the nature of the cryosphere and the changes that are taking place within it. In the following sections we will examine some of the issues raised by using these remote sensing techniques and some of the difficulties encountered.


Issues in Differentiation

In the discussions on the various types of remote sensing used in the study of the cryosphere, the issues associated with how the analyses of the imagery data is performed was ignored. For the most part it was assumed that image interpretation is done by human interpretation of a remotely sensed image. However, it is expected that by the turn of the century, there will be a large volume of data being returned to Earth by satellites, on the order of 1 terabyte each and every day (Gershon and Miller, 1993). This is such an enormous amount of data that if it were stored on standard magnetic tapes (capable of holding roughly 120 Megabytes of data) it would take a stack of tapes as tall as the Washington monument (roughly 169 meters) for a single day (Gershon and Miller, 1993). Expecting that all of this data will be analyzed by humans in a continuous fashion is quite impractical to say the least. Therefore, one must ask how are we going to store and analyze so much data and get some useful parameters for global environmental research.

Regarding the storage and dissemination of the vast amounts of data and information, NASA has proposed to build the Earth Observing System Data Information System (EOSDIS) as part of its so called Mission to Planet Earth (Gershon and Miller, 1993). We will not go into even a cursory examination of the EOSDIS, referring the reader to the vast literature that now exists on the system, including the Gershon and Miller paper referenced. However, this matter is brought to the attention of the reader to highlight the massive feat ahead associated with the science of remote sensing. Thus, it appears that some form of automated fashion is required for at least recognizing the various sensor scenes that include the objects of interest, in this case, snow and ice.

Therefore, in the next section we will examine some of the new applications for classifying a remote sensing scene and using this as part of one's study of the imagery.


New Applications for Scene Classification

One recent approach to scene classification has been the application of artificial neural networks. Artificial neural networks are a class of computational models inspired by, but only partially modeling our present understanding of biological nervous systems (Geller, 1992). A detailed description of this artificial intelligence application is not appropriate for this paper but its applications to the study of cryospheric remote sensing imagery will be broached.

In 1991, there was a special session of the American Geophysical Union (AGU) in which it was reported how neural networks were tested on Landsat TM imagery for their use in discriminating between such features in a scene as land, water and clouds (Barker et al., 1991). Since then, there has been a large number of applications of artificial neural networks including the discrimination of SAR sea ice images (Fernandes and Jernigan, 1992).

Fernandes and Jernigan wanted to determine if they could develop an automated approach for classifying SAR imagery in terms of the presence of first year smooth ice, first year rough ice and multiyear ice. They decided to experiment with both standard linear classification algorithms and artificial neural networks. They arrive at the conclusion that neither the linear classification system nor the artificial neural network are suitable, as tested, for the discrimination of SAR ice images. In fact, there seems to be little difference in the accuracy of classification between the linear methods and the artificial neural network. However, their conclusions are based upon the use of what is known as a two dimensional Gabor filter specified in the frequency domain and the values known as the spatial/spatial frequency (S/SF) texture measure. Therefore, their results may be more of a testimony to the use of these S/SF texture measures than either the linear or neural network techniques.

The use of simple linear algorithms has certainly not been exhausted and their applications remain promising. One of the key parameters that global change researchers are interested in, is the snow cover of the globe. After all, snow is a major component of the hydrological cycle possessing a strong influence on the Earth's radiation budget because of its albedo, as noted previously in this report. This makes snow a major component in any global climate model. Thus, investigators are interested in determining what portion of the earth is covered with snow at any one time.

Working in preparation for the launch of NASA's Earth Observing System, the development of a snow cover algorithm for use with the Moderate Resolution Imaging Spectrometer (MODIS) has been reported (Riggs et al., 1992). The authors report the development of a function, analogous to the vegetative index, called the normalized snow difference index (NSDI). This index was developed by using Landsat Thematic Mapper bands 2 and 5. The NSDI can be expressed as follows:

             NSDI = (TM band 2 - TM band 5) / (TM band 2 + TM band 5)
The major advantage of this algorithm over the simple approach of identifying snow simply by establishing a threshold value for individual pixels based on single band reflectance values is that the NSDI approach can also detect shaded snow, whereas the simple threshold approach cannot.

Interestingly enough, while this investigative team is ultimately trying to determine a viable snow cover algorithm for the MODIS instrument, they have uncovered a analog to the vegetative index for use with Landsat TM imagery, they do not offer a true MODIS algorithm, since the MODIS bands are much more numerous than the TM bands. It will be interesting to see what this team develops based on the more extensive number of bands of the MODIS instrument. The MODIS bands are presented in Figure 3. Note that the TM band 2 covers the region from 0.52 to 0.60 micrometers and the TM band 5 covers the region from 1.55 to 1.75 micrometers. In the same region as TM band 2, the MODIS instrument possesses four bands, and in the same region as TM band 5, the MODIS instrument possesses a single band, MODIS band 6.

It is important to note that not all parties view the issue of snow discrimination the same as the Riggs investigative group. In the same volume as the Riggs report appeared, there also appeared a report (Duguay and Hurtubise, 1992) which determined snow cover by using Landsat TM band 7. This group, lead by Duguay, was in the process of developing a "geographic database which includes climatological data on air temperature, precipitation, wind speed and direction (1952-present), Landsat MSS (1973-1989) and TM (1984 and 1986) images, as well as digital terrain data." Their goal is to develop a system for monitoring climate change in the Niwot Ridge, Colorado region. However, one must take into account the fact that their definition of snow cover will differ from that developed by Riggs et al., confusing the issues associated with snow cover. It will be interesting to see how these teams progress in their studies, realizing that any comparison of results must take into account the different methods.

Figure 3 MODIS VIS/NIR/SWIR Bands

Figure 3 MODIS VIS/NIR/SWIR Bands (cont'd)