Using spectrum differentiation and combination for target detection of minerals

https://doi.org/10.1016/j.jag.2016.10.005Get rights and content

Highlights

  • Evaluating the potential value of derivative spectrum in target detection.

  • Using derivative spectrum in 5 different orders.

  • Having a spectral curve and its best derivative order in a unified approach introduced as ECEM.

  • A real airborne hyperspectral data were applied in comparison to previous LAB-based approaches.

Abstract

Among the techniques that have been developed in spectroscopy, derivative analysis is particularly promising for use with remote sensing data. In the first step of this research we apply the derivative spectrum in a real hyperspectral image and introduce a new target detection approach called “DCEM”. For this purpose, 1st to 5th orders of derivative spectrum were applied to the DCEM. The outcome of this research has shown that the application of derivative spectrum in target detection is perfectly advisable in a specific derivative order for each target. This order can be introduced as an optimized order or the Best DCEM. The spectrum differentiation eliminates low frequency components of the spectrum. Despite the little information included in those low frequency components of a signal or spectrum, their complete elimination cause an information loss problem. Hence, in the second step of this research an ensemble classifier approach was employed for the combined use of both spectra and the best derivative order. This simultaneous use of the derivative and zero order spectra is introduced as “ECEM”. Experiments were conducted via a HyMap hyperspectral airborne image in eastern Iran. The detection results show that both proposed methods significantly outperform CEM in ROC and AUC values. The best performance upgrade in DCEM detection was about 24% for Kaolinite target and about 28% for Alunite target in ECEM.

Introduction

Imaging Spectrometry takes the advantage of contiguous spectral channels to unveil signal sources that usually cannot be resolved by multispectral sensors (Chang, 2003, Chang, 2007, Clark, 1999). In hyperspectral data exploitation, most targets of interest are those with small spatial presence and low probability existence in either form of mixed pixel or sub pixel (Chang, 2007, Chang and Heinz, 2000, Schowengerdt, 2007). Therefore, literal based image processing techniques (Chang, 2007, Van der Meer, 2006, Van der Meer et al., 2012, Zhang and Pazner, 2007) are not effective in hyperspectral images even if they can be applied. Derivative spectroscopy and spectral analysis, on the other hand, offers a valuable alternative in this area.

Numerical differentiation is a common tool used in analytical chemistry for processing one-dimensional signals since a derivative is sensitive to important yet subtle details in a spectrum. Hyperspectral data nature allows the implementation of such techniques in this kind of image data (DemetridesShah et al., 1990). Application of this technique in remote sensing is due to its ability for resolving complex spectra of several target species within individual pixels (DemetridesShah et al., 1990), and the fact that derivatives of second or higher order are insensitive to illumination variations caused by sun angle, cloud cover, or topography (Tsai and Philpot, 1998, Tsai and Philpot, 2002). Additionally, it is worth considering that the spectral shape information is one of the most important measures in evaluation of the spectral similarity between two spectra (Clark et al., 2003, Hungate et al., 2008, Van der Meer and De Jong, 2011). As it was mentioned in (Dehnavi et al.,, 2014, Tsai and Philpot, 1998, Zhang et al., 2010 Tsai and Philpot, 1998; Zhang et al., 2010), derivatives of different orders contain various levels of spectral shape information in a spectrum. Therefore, it is not unexpected to have better detection on signal sources in an individual pixel by the usage of derivative spectrum and consequently its spectral shape information. Moreover, spectrum matching techniques are sometimes ruined, in as much as the library spectrum is less fitted to the field spectrum in contrast to its derivative (Kim, 2011). Taken together, it is of great value applying information content of derivative spectrum in target detection.

Several investigators hinted that employing derivative signature has improved the results of their research in different application areas, including but not limited to the following: Torrecilla et al. took the advantage of derivative spectrum for identification of water constituents in oceanographic regions (Torrecilla et al., 2009). Algal chlorophyll concentration was estimated in (Han, 2005) by application of derivative analysis. The first derivatives were computed and correlated with the chlorophyll-a concentration in a shallow water in (Jensen, 2009). Second derivative approximation was applied in (Becker et al., 2005) in order to help identifying spectral regions where the spectral bands are the most botanically explanative. Partial abundances of spectrally similar minerals in complex mixtures were estimated in (Debba, 2009, Debba et al., 2006), applying first and second orders of derivative spectra. Toxic algal bloom was detected in (Craig et al., 2006) from the analysis of the fourth derivative of phytoplankton absorption spectra, estimated from in situ hyperspectral measurements of reflectance. First and second orders of derivative spectrum were also used in previous researches for un-mixing purposes (see (Bieniarz et al., 2012, Hengqiana et al., 2012, Zhang et al., 2004)). In an earlier research, second order derivative spectrum was applied in linear spectral mixture model (LSM) aimed to estimate minerals’ abundance (Zhang et al., 2004). Results of their research suggested an increase in the accuracy of the abundance estimates. However as is clear, the usage of higher order derivatives is neglected in these researches, specifically in target detection.

The spectrum differentiation concept was also previously employed in binary encoding (Kim, 2011), spectral derivative based feature coding (SDFC) concepts (Chang, 2013, Chang et al., 2009), and also the determination of absorption features by using gradient changes in the spectral reflectance curve (Rezaei et al., 2012) for endmember extraction purposes. However, in such coding methods the information content of whole derivative signature was not applied.

In sum, previous researches have proven the practical usefulness of derivative spectrum in different fields of research. However, many of the earlier works were carried at the laboratory level (see for example (Dehnavi et al., 2015), whereby some orders of derivative spectrum were introduced as good discriminators for mineral targets). On the other hand, they rarely or never discussed application of higher order derivative spectrum, i.e. 3rd to 5th orders, in target detection of hyperspectral images. The first idea of this work is thus, putting in various orders of derivative spectrum (1st to 5th orders) into CEM algorithm, introduced as “Derivative Constrained Energy Minimization” or “DCEM”, and exploring the detection maps in available image data.

Moreover, one of the points that we have to pay special attention is that spectrum differentiation eliminates low frequency components of the spectrum. Despite the little information included in the low frequency components of a signal or spectrum, their complete elimination cause an information loss problem. Having a spectra and its best derivative order (as a result of DCEM) in a unified approach cause an increase in the information we could obtain from a spectral curve. Hence, in the second step of this research an ensemble classifier approach (Dietterich, 2000, Dietterich, 2002, Rokach, 2010, Zhang et al., 2010) was applied for the combined use of both spectra and the best derivative order. This combined uses of derivative and zero order spectra is introduced as “Ensemble Constrained Energy Minimization” or “ECEM”. Thereupon, it can be claimed that the majority of the information contained in a target's spectral curve is used in ECEM.

To justify the above-mentioned frameworks, this study was carried out with an airborne hyperspectral data in comparison to the previous works whereby the standard laboratory images were applied (Zhang et al., 2004). Both proposed detection algorithms were used for identification of four mineral targets including alunite, kaolinite, epidote and hematite which were located in a hydrothermally altered mineral region in Iran east. Due to the possible economic importance of various minerals within the study area better mapping performance was required. To this end, our proposed method was used for better mapping of the region.

A description on the methods applied in this work and the evaluation methodology is presented in Section 2. Section 3 focuses on the study area, experimental and ground truth data; laboratory measurements, pre-processing step and experimental results are discussed in Sections 4–6 respectively. Finally, we draw out our conclusion in Section 7.

Section snippets

Method

This section provides a full description of the proposed three-step approach for target detection. First, derivative spectrum was used for target detection in the proposed DCEM approach. Second, one of the derivative orders was selected as the best order for the identification of each target. Third, a combination of both spectra and its derivative spectrum was applied in the proposed ECEM approach.

Experimental data

Study was conducted via a HyMap airborne hyperspectral dataset (Bannon, 2009, Cocks et al., 1998; Rica, 2005; Schapfer, 2006) in Gonabad county, Khorasan Razavi Province, Iran east (Lat. 34°25'N and Lon. 58°34'E). The HyMap images were acquired on September 11th 2006 with a spatial resolution of 5 × 5 m. HyMap sensor has been developed in Australia by Integrated Spectronics, HyVista Corporation and now deployed for commercial operations around the world. The specific aspects of the sensor is

Laboratory measurements

Laboratory measurements are required not only for the detection of the mineral targets, but also for the validation of the results. Thus, in what follows three different lab measurements required in this research are presented and explained in detail.

After the completion of the field sampling procedure, samples were transferred to the laboratory. They were firstly prepared by some expert geologists prior to the spectral measurements. For this purpose, all samples were first round to rock powder

Pre-processing

Pre-processing phase is a basic and important step toward the success of target detectors especially when derivative spectrum is used. The data products delivered by HyVista Corporation were radiance calibrated (at sensor), and geometric correction look up tables. All data were delivered as ENVI compatible files. Since the spectral and radiometric calibration of the HyMap sensor were accomplished prior to the survey, this information helped convert the raw DN counts to radiance values in μWcm2nm

Results and discussion

Detection maps were generated based on the proposed DCEM and ECEM approaches. Then, a model-based geometric correction approach was used for geo-correction and geo-referencing of the detection maps. For this, a geometric lookup table (.GLT) file prepared by HyVista Corporation based on HyMap sensoring model was used. Georeferenced detection maps for the targets of interest are presented in Fig. 10, Fig. 11, Fig. 12, Fig. 13 . Fig. 10, Fig. 11, Fig. 12, Fig. 13(a)–(c) show the detection results

Conclusion

The primary purpose of this research was to evaluate the potential of derivative spectrum in target detection. For this, derivative spectra was in CEM target detection algorithm and a new approach introduced as “DCEM”.

This study was carried out with an airborne hyperspectral data in comparison to the previous works whereby the standard laboratory images were applied. Study was conducted via an airborne hyperspectral HyMap data in a hydrothermally altered mineral region in Iran East. This region

References (61)

  • Chang, C.I., 2005. Orthogonal subspace projection (OSP) revisited: a comprehensive study and analysis. Geoscience and...
  • C.I. Chang

    Hyperspectral Data Exploitation: Theory and Applications

    (2007)
  • C.I. Chang

    Hyperspectral Data Processing: Algorithm Design and Analysis

    (2013)
  • R.N. Clark et al.

    Imaging spectroscopy: earth and planetary remote sensing with the USGS Tetracorder and expert systems

    J. Geophys. Res.

    (2003)
  • Clark, R.N., 1999 Spectroscopy of Rocks and Minerals, and Principles of Spectroscopy Manual of Remote Sensing 3,...
  • Cocks, T., Jenssen, R., Stewart, A., Wilson, I., Shields, T., 1998. THE HYMAPTM AIRBORNE HYPERSPECTRAL SENSOR:THE...
  • Craig, S.E., Lohrenz, S.E., Lee, Z., Mahoney, K.L., Kirkpatrick, G.J., Schofield, O.M., Steward, R.G., 2006. Use of...
  • P. Debba et al.

    Abundance estimation of spectrally similar minerals by using derivative spectra in simulated annealing

    IEEE Trans. Geosci. Remote Sens.

    (2006)
  • P. Debba

    Abundance estimation of spectrally similar minerals

  • Dehnavi, S., Maghsoudi, Y., ValadanZouj, M.J., Baniadam, F., 2014. Investigation of Hydrothermally Altered Zones...
  • S. Dehnavi et al.

    High-order derivative spectrum in hydrothermally altered minerals discrimination

    J. Appl. Remote Sens.

    (2015)
  • T.H. DemetridesShah et al.

    High resolution derivative spectra in remote sensing

    Remote Sens. Environ.

    (1990)
  • T.G. Dietterich
    (2000)
  • Dietterich, T.G. 2002. Ensemble learning. The handbook of brain theory and neural networks 2, 110–125....
  • J. Fürnkranz

    Pairwise classification as an ensemble technique

    European Conference on Machine Learning

    (2002)
  • Goldberg, H.R., 2007. A performance characterization of kernel-based algorithms for anomaly detection in hyperspectral...
  • P.A. Gorry

    General least-squares smoothing and differentiation by the convolotion (Savitzky-Golay) method

    Anal. Chem.

    (1990)
  • L. Han

    Estimating chlorophyll-a concentration using first-derivative spectra in coastal water

    Int. J. Remote Sens.

    (2005)
  • Harsanyi, J.C., Chang, C.I., 1994. Hyperspectral image classification and dimensionality reduction: an orthogonal...
  • J.C. Harsanyi et al.

    Detection of subpixel spectral signatures in hyperspectral image sequences

    Annual Meeting Proceedings of American Society of Photogrammetry & Remote Sensing

    (1994)
  • Cited by (7)

    • Deep learning-based spectral reconstruction in camouflaged target detection

      2024, International Journal of Applied Earth Observation and Geoinformation
    • Quantitative analysis of the oil mixture using PLS combined with spectroscopy detection

      2021, Optik
      Citation Excerpt :

      It becomes more urgent to find a rapid detection method to determine the oil mixture [11–14]. Spectral detection has attracted wide attention because of its non-contact, high speed and accuracy [15–17]. Which makes up the shortcomings of traditional detection methods, and many scholars have applied spectroscopy combined with PLS to detect substances [18–21].

    • A new cluster tendency assessment method for fuzzy co-clustering in hyperspectral image analysis

      2018, Neurocomputing
      Citation Excerpt :

      The advantages of the hyperspectral image are the high spectral resolution, providing spectral characteristics and spatial information simultaneously, the large band number which contains knowledge of the target spectrum for target detection and recognition. Some recent typical studies on the hyperspectral image are for target detection of minerals[38], environmental management [39], military [40], target detection [41,42] and the hyperspectral image classification [43–45]. As a powerful model, convolutional neural networks (CNNs) have demonstrated remarkable performance in various image representation and recognition problems.

    View all citing articles on Scopus
    View full text