Modern sensors produce remotely sensed data with hundreds of spectral bands. Since this data is generated continuously, data transmission and archiving become significant challenges. Consequently, there is an increasing need for efficient compression algorithms. Traditional approaches focus on minimizing the mean squared errors between original and compressed data. However, this creates problems in the context of hyperspectral images, which can be used for land-cover classification. For example, we can use such images to detect roads, rivers, buildings, and other targets. In this classification application, we need to preserve the features that are useful for discrimination among classes. However, since these features may not necessarily have significant energy levels, they may not be well represented during the encoding process.
In the past, a number of algorithms have been used to compress hyperspectral images. One can use any 2D or 3D compression algorithm, such as JPEG1 (named after the Joint Photographic Experts Group, which created the standard), set partitioning in hierarchical trees (SPIHT),2 or JPEG2000.3 However, we noticed that the resulting classification accuracy does not correlate well with the mean squared errors. In other words, although the differences between the original and compressed images are very small, the classification accuracy associated with using the compressed images is lower than that of the original images. When we try to detect certain targets using the compressed images, the detection rates are still lower, even though the original and compressed images look almost identical.
Figure 1. Illustration of discriminant features that are low in energy. φ2, φ2: Feature vectors.
To address this problem, we enhanced the discriminant features of the hyperspectral images prior to compression. First, we needed to find the relevant discriminants using feature-extraction algorithms. Such algorithms generally produce a set of feature vectors that are useful for discrimination among classes. Enhancing these features is equivalent to stretching the data along the direction in which the classes can be better separated. For example, in Figure 1, vector φ2 will provide better classification results, although φ1 is better for signal representation. However, most compression algorithms encode the data along φ1. In our method,4,5 we first stretched the data along φ2(see Figure 2), which forced the conventional compression algorithms to better represent the data along that direction.
Figure 2. Stretching the images along the most useful direction for feature detection.
We applied our enhancement algorithm to classify agricultural areas (see Figure 3). Table 1 lists the class descriptions. First, we applied decision-boundary feature extraction6 and obtained discriminant feature vectors. We subsequently stretched the images along the dominant discriminant-feature vectors and compressed the stretched data using the JPEG2000 algorithm. The enhanced (stretched) images exhibit much better classification accuracy than the compressed images that were not first enhanced (see Figure 4).
Table 1. Class descriptions. EW: Original-data identifier.
|Class index||Class species||No. samples||No. training samples|
|3||Corn (clean till)||966||100|
|7||Soy beans (clean till)||1050||100|
|8||Soy beans (clean till-EW)||1722||100|
Figure 3. Selected classes.
Figure 4.Classification-accuracy improvement.
In summary, satellite sensors continuously produce ever larger amounts of hyperspectral data. Consequently, data compression has become an important issue. While traditional compression methods mainly focus on minimizing the differences between original and compressed images, it is also important to preserve the classification characteristics of hyperspectral images when they are aimed at distinguishing among various classes of discriminating features. By enhancing the latter, we can better preserve the relevant classification performance. In our future studies, we will combine unsupervised classification and our new enhancement method to provide automated means of improved preservation of the discriminant information in hyperspectral images.
This work was supported by grant R01-2006-000-11223-0 from the Basic Research Program of the Korea Science and Engineering Foundation.
Chulhee Lee, Jonghwa Lee
Department of Electrical and Electronic Engineering, Yonsei University
Seoul, Republic of Korea
Chulhee Lee received his BS and MS degrees in electronic engineering from Seoul National University (Republic of Korea) in 1984 and 1986, respectively, and a PhD degree in electrical engineering from Purdue University in 1992. His research interests include image/signal processing, pattern recognition, and neural networks.
Jonghwa Lee received his BS degree in electronic engineering from Yonsei University in 2005, where he is currently pursuing a PhD degree. His research interests include image/signal processing, pattern recognition, and video-quality measurement.