Classification by diagnosing all absorption features (CDAF) for the most abundant minerals in airborne hyperspectral images
© Mobasheri and Ghamary-Asl; licensee Springer. 2011
Received: 11 February 2011
Accepted: 14 November 2011
Published: 14 November 2011
Imaging through hyperspectral technology is a powerful tool that can be used to spectrally identify and spatially map materials based on their specific absorption characteristics in electromagnetic spectrum. A robust method called Tetracorder has shown its effectiveness at material identification and mapping, using a set of algorithms within an expert system decision-making framework. In this study, using some stages of Tetracorder, a technique called classification by diagnosing all absorption features (CDAF) is introduced. This technique enables one to assign a class to the most abundant mineral in each pixel with high accuracy. The technique is based on the derivation of information from reflectance spectra of the image. This can be done through extraction of spectral absorption features of any minerals from their respected laboratory-measured reflectance spectra, and comparing it with those extracted from the pixels in the image. The CDAF technique has been executed on the AVIRIS image where the results show an overall accuracy of better than 96%.
The image classifications are based entirely on the spectral signatures of the land cover types. This area of specialty has attracted the attention of remote sensing researchers in recent years and as a result, the techniques of classification have been improved considerably. These techniques have been divided into two general categories: supervised and unsupervised. In supervised classification, usually the statistical methods  and training samples are being used, whereas the unsupervised classification is based on the comparison between spectral signatures of a pixel and those of different materials collected in spectral libraries .
Spectral characteristics is a tool that has been used for decades to identify, understand, and quantify solid, liquid, or gaseous materials, especially in the laboratory. This is usually done through detection of absorption features due to the presence of specific chemical bonds, where its depth of absorption represents the abundance and physical state of the detected absorbing species [3–5]. Imaging spectroradiometer can acquire data with suitable spectral range, resolution, and sampling rate at every pixel in a raster image, so that individual absorption features can be identified and spatially mapped .
One of the most powerful methods in unsupervised classification is Tetracorder introduced by Clark et al. . There are five innovations in this method where two of these are used in this study. The first innovation in Tetracorder method is to identify materials by comparing a remotely sensed spectrum (here pixels reflectance spectrum) with a large number of spectra of well-known materials . Of course it involves some undesired signals when working with mixed pixels but we usually interested only on the portions of the spectrum that are known to be diagnostic of the reference materials. Since every spectral feature is due to an interaction of photons of particular energies with the atoms and electrons within the chemical under study, then the nature of the absorption is largely unique to the specific chemical structure where the concept of a diagnostic absorption feature is used for it. These diagnostic absorption features are unique to particular materials in shape but varies in intensity with wavelength over a narrow interval and usually are concentrated in limited ranges of wavelength by type of absorption . Of course the width of these absorption features may vary due to the phenomenon such as Doppler shift, presence of overtones, effects of mixed pixels, etc.
The second innovation that Tetracorder presents is quantitative comparison between an unknown spectrum (pixel) and the entries from the spectral library reflectance curves seeking for similar diagnostic features. Then those materials having the highest similarity to the unknown are the most probable substance that can be present in the pixel. Thus, Tetracorder not only compares the pixel's spectral properties to the spectral properties of each of the entries from the library, but also quantitatively assesses and judges to identify the material present in the pixels.
On the other hand, there might exist materials that show similar diagnostic features as perceived by our normalization process, but they are never similar in all other key absorption wavelengths, or in terms of the local spectral normalized parameters such as reflectance local slope and depth of absorption. Although Tetracorder is based on five hypotheses but in this study we only built up a technique partly based on the first two of them along with some calculations on fitting to reflectance curves and using their continuum removals (CR). For this, classification by diagnosing all absorption features (CDAF) is selected for simplicity and presented in details below.
Usually the CR is used to identify the spectral features through their wavelength position and shapes . Most of remotely sensed spectra are composed of mixtures and not necessarily pure materials, and then the spectral reflectance curves produce a continuum upon which diagnostic absorptions may be superimposed. The CR algorithm can remove the effects of these other absorption features from the spectrum [5, 8]. The depth or strength of an absorption feature in the continuum depends upon the intrinsic absorption strength, the grain size, and abundance of the material mixed in the sample . The absorption feature's depth is generally proportional to the abundance of the materials in the sample (for a fixed value of grain size). On the other hand, the depth of a feature may increase to a maximum with larger grain size, but decreases as the absorption dominates over scattering .
2. Region of the study and data
3.1. Data preparation
Our new technique is developed using the first two innovations of the Tetracorder, a special analysis. In what follows, the CDAF technique is introduced through its eight implementation stages. Here to prepare the images for implementing CDAF, the following few calculation steps were carried out first:
The fitness coefficient F is assumed to be a measure of how well the spectral features match each other.
It is worthy to note that, in this study the definition of diagnostic absorption feature is slightly different from one that is used in other studies. In this study, we consider the whole region between two consecutive values of 1 in CR curve as a diagnostic absorption feature; where it is fully independent of its area, shape, depth, etc.
where A i is the area confined between CR and line representing reflectance 1 (Figure 2) in i th absorption region, and A represents the whole area confined between CR and horizontal line 1, F i is the fitness coefficient (Equation 3) for the i th absorption feature and Fw represents weighted fitness for all Na absorption features present in the spectrum.
Step 4: Now the mineral corresponding to the highest weighted fitness is taken as the class of the pixel under consideration .
3.2. Implementation of CDAF technique
The CDAF technique is based on the first two innovations of Tetracorder method plus some manipulations on the pixel and material reflectance spectrum in spectral libraries and their CRs. This means that the CDAF technique takes all of these curves into its calculations. The main hypothesis of this technique is based on the fact that the minerals may continuously be present in small regions such as 20 × 20 m , i.e., the surface distribution of a particular material does not change abruptly and this distribution could be assumed more or less similar from one region to its neighboring regions. Then, if the majority of pixels certify the presence of a particular material, it would imply that this material could be present dominantly in the contiguous pixels too. This assumption could enable us to reduce the chance of misclassification.
In CDAF technique, two thresholds "Fitness Threshold" and "Frequency Threshold" are used where each one of these could be determined either analytically or through some experimental procedures (if the ground truths are available). This can be done for different images of different sensors and for different conditions. In this study, these thresholds have been found to be around 80 and 70%, respectively. The procedure of implementation of this technique is presented in eight stages as follows:
Stage 1: selecting a small region (5 × 5 pixels) on the AVIRIS image. It is true that we have used 5 × 5 pixels windows in the classification procedure but this does not necessarily mean that we have lowered the spatial resolution down to 100 m. The reason for selecting this window was to expedite the calculation (reduce calculation time), for example by using this window we could find 10 materials out of 481 different materials in the spectral library having more chance to be present in the pixels. These ten materials are used for second round of classification, and in this round the classification is run for each and every pixel. It is possible for a pixel to be assigned a class different from its neighbors in 5 × 5 windows in the second run.
Stage 2: applying first and second steps of Tetracorder on each pixel on the selected region using all materials present in the spectral library.
Stage 3: selecting those materials having fitness above fitness threshold (i.e, 80%) and dismiss the rest.
Stage 4: Using results of stage 3 for selection of materials having frequency of more than its threshold value (here 70%).
After this stage, the number of materials suspected to be present in image is decreased dramatically and this will expedite the remaining calculations. It is worth to consider that the number of materials with highest frequencies determined in stage 3 is not necessarily equal to that of stage 4. So, before moving to the next stage, a criterion will be imposed in order to optimize the results. This criterion consists of an "if condition" as:
"if the ratio of the highest frequency to the total number of pixels in the selected region is greater than the frequency threshold"
then the reflectance curves of these selected pixels would be used in the next stages (instead of their CRs), because the main hypothesis of the CDAF (i.e., the minerals may present as continuous in small regions) is fulfilled. In case the condition is not met, then CR curves of both material and pixels will be used in the next stages. Since usually this condition is always met, then in what follows the term reflectance spectrum will be used (instead of CR).
where r and t stand for the reference (material) and target (pixel), respectively. It is clear that these new curves are totally different with the original reflectance curves.
where a is linear relationship coefficient and is the transpose of (Figure 3c).
Now the material that minimizes these RMSEs is taken to be the class of that pixel. Here, Nb is the number of bands being used.
4. Results and analysis
AVIRIS images explained in Section 2.
Field collected data for five classes of materials (Table 1).
Reflectance curves for 481 different materials collected by USGS available in ENVI software. The spectral library of the minerals can also be found in the USGS website http://speclab.cr.usgs.gov/spectral-lib.html.
Characteristics of 5 materials in 130 samples used in method testing and analysis
Material number in the list of ENVI spectral library
Number of labeled samples
67 and 68
88, 91, 380-383, 345-346
4.1. Technique evaluation
Results of classification and overall accuracy of CDAF method
Number of Class Samples
Number of pixels classified correctly
Classification accuracy for each class (%)
Overall accuracy: 96.92% (126/130)
The robustness of the CDAF is more profound in some special circumstances. One of these cases is the class of silica. There are varieties of silicas different in names but more or less with high similarity in the absorption features in their reflectance curves. Adding to this, the soils reflectance with special characteristics due to their physical and chemical particulars while their main constituents are minerals, organic materials, air, and water [10, 16]. These pixels were classified with 100% accuracy using CDAF technique. The reason for this is the analytical procedure introduced in the technique. Table 2 shows that by removing calcite from the list of classes, the classification precision raises from 96.92 to 100% where this itself is a great step in the classification of hyperspectral images .
Finally, it is worth noting that we have run our technique based on the field collected data by other people in the Cuperite field campaign where we ourselves did not play any role in that. As can be seen from Table 2, the CDAF technique could detect four of the classes with 100% accuracy and only has weakness in detection of calcite class (73.33%). However, it is possible that the field work done by the other people on the class of calcite was not accurate enough and had some misinterpretations involved. For example, it is possible in some region, the siderite abundance was more than calcite but the people collected the samples gave the calcite class. On the other hand, the accuracy of 100% does not necessarily means that there are no other material present in the pixel but it could be concluded that the combination of abundance and absorption features of on material has influenced the pixel reflectance to render the shape to look like a particular class.
In this study, based on the first two stages of Tetracorder method, a new technique called CDAF is developed. This technique enables one to classify the minerals with high precision. The technique is based on the derivation of information from the image reflectance spectrum. This can be done through extraction of spectral absorption features of any minerals from their corresponding laboratory-measured reflectance spectra, and comparing it with those extracted from the image. The results of evaluation show acceptable and reliable performance of the suggested technique. In this study, along with the first two innovations of Tetracorder method, based on absorption depth and absorption areas in the CR of reflectance curves, some weighting coefficients are calculated. These weighting coefficients help classification of pixels through their spectral similarities to a particular substance. The significance of CDAF technique is that in this technique besides using absorption features in the material and pixel in the respective CR of reflectance curves, the reflectance curves themselves are being used as well. Considering the results of classification, it can be seen that CDAF can perform well in all classes and can be applied to the region with the high variety of minerals distributing in a continuous manner. On the other hand, the results shown in Table 2 prove that this technique has been able to perform 100% accuracy in classification for most of the cases. So, the CDAF technique is recommended as a good substitution for unsupervised classification techniques in hyperspectral images.
- Kuo BC, Landgrebe DA: Improved statistics estimation and feature extraction for hyperspectral data classification. PhD thesis, School of Electrical & Computer Engineering Technical Report. Purdue University, USA; 2001.Google Scholar
- Clark RN, Swayze GA, Livo KE, Kokaly RF, Sutley SJ, Dalton B, McDougal RR, Gent CA: Imaging spectroscopy: earth and planetary remote sensing with the USGS Tetracorder and expert systems. J Geophys Res 2003, 108: 1-44. [http://speclab.cr.usgs.gov/PAPERS/tetracorder]Google Scholar
- Hunt GR, Salisbury JW: Visible and near-infrared spectra of minerals and rocks. I. Silicate minerals. Mod Geol 1970, 1: 283-300.Google Scholar
- Clark RN, King TVV, Klejwa M, Swayze GA, Vergo N: High spectral resolution reflectance spectroscopy of minerals. J Geophys Res 1990, 95: 12653-12680. 10.1029/JB095iB08p12653View ArticleGoogle Scholar
- Clark RN: Spectroscopy of Rocks and Minerals, and Principles of Spectroscopy, Manual of Remote Sensing, Remote Sensing for the Earth Sciences. Volume 3. Edited by: Rencz AN. Wiley, New York; 1999:3-58.Google Scholar
- Mobasheri MR, Rezaie Y, Valadan-zouje MJ: A method in extracting vegetation quality parameters using hyperion images, with application to precision farming. World Appl Sci 2007, 2: 476-483.Google Scholar
- Clark RN, Swayze GA, Wise R, Livo KE, Hoefen TM, Kokaly RF, Sutley SJ: 2000-2007, USGS Digital Spectral Library.[http://speclab.cr.usgs.gov/spectral-lib.html]Google Scholar
- Clark RN, Roush TL: Reflectance spectroscopy: quantitative analysis techniques for remote sensing applications. J Geophys Res 1984, 89: 6329-6340. 10.1029/JB089iB07p06329View ArticleGoogle Scholar
- Kruse FA, Boardman JW, Huntington JF: Comparison between AVIRIS and hyperion for hyperspectral mineral mapping. IEEE Trans Geosci Remote Sens 2003, 41: 1388-1400. 10.1109/TGRS.2003.812908View ArticleGoogle Scholar
- Gao BC, Goetz AFH: Column atmospheric water vapor and vegetation liquid water retrievals from airborne imaging spectrometer data. J Geophys Res 1990, 95: 3549-3564. 10.1029/JD095iD04p03549View ArticleGoogle Scholar
- Boardman JW: Post-ATREM polishing of AVIRIS apparent reflectance data using EFFORT: a lesson in accuracy versus precision. In Proceedings of the 8th JPL Airborne Earth Science Workshop. Jet Propulsion Laboratory Publication 99-17; 1998:53.Google Scholar
- Clark RN, Swayze GA, Livo KE, Kokaly RF, King TVV, Dalton JB, Vance JS, Rockwell BW, Hoefen T, McDougal RR: Surface reflectance calibration of terrestrial imaging spectroscopy data: a tutorial using AVIRIS. In Proceedings of the 10th Airborne Earth Science Workshop. JPL Publication 02-1; 2002.Google Scholar
- Kokaly RF: Investigating a physical basis for spectroscopic estimates of leaf nitrogen concentration. Remote Sens Environ 2001, 75: 153-161. 10.1016/S0034-4257(00)00163-2View ArticleGoogle Scholar
- Swayze GA, Clark RN, Goetz AFH, Chrien TG, Gorelick NS: Effects of spectrometer band pass, sampling, and signal-to-noise ratio on spectral identification using the Tetracorder algorithm. J Geophys Res (Planets) 2003,108(E9):5105.View ArticleGoogle Scholar
- Zhang J, Rivard B, Rogge DM: The successive projection algorithm (SPA), an algorithm with a spatial constraint for the automatic search of endmembers in hyperspectral data. Sensors 2008, 8: 1321-1342. 10.3390/s8021321View ArticleGoogle Scholar
- Van Der Meer FD, Dejong SM: Imaging Spectrometry: Basic Principles and Prospective Applications. Delft University of Technology, Kluwer Academic Publishers; 2002:11-12. 47, 71-80View ArticleGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.