Varying atmospheric conditions, differences in the sun geometry, topographic effects and the sensor scanning system strongly influence the recorded signal. And these influences modify the true spectrum of the ground features, Which means that the original data cannot be used quantitatively. Therefore, to obtain the real spectrum of the ground objects from imagery and carry out alteration extraction, preprocessing such as radiometric calibration, accurate atmospheric correction and geometric correction must be performed. After image preprocessing, the typical mineral mapping process, including Minimum Noise Fraction (MNF), Pixel Purity Index (PPI), N-dimensional visualization and mineral mapping techniuqes, was used to extract alteration information in 400-1000 nm and 2058-2361 nm range respectively. Aditionally, PCA and GLCM were used to extract texture information in the study area. The detailed methodology flowchart is shown in Fig. 2.
Hyperspectral image processing. The preprocessing of GF-5 AHSI imagery mainly included radiometric calibration, atmospheric correction, band selection, destriping, orthorectification and spectral denoising.
(1) Radiometric calibration. It includes relative calibration and absolute calibration. The relative calibration is used to remove the effects of the non-uniformity of the detector response. The absolute radiometric calibration is used to convert the detector response digital number (DN) values to radiance received by the optical aperture. In this paper, the Radiometric Calibration tool from ENVI (Environment for Visualizing Images) software version 5.3 was used to calibrate AHSI data. During processing, band interleaved by line (BIL) and scale factor of 0.1 were adopted.
(2) Atmospheric Correction. Gas molecules and aerosols in the atmosphere absorb and transmit the incoming light in regards to its wavelength and this causes distortons in the image. To eliminate the influence of atmosphere in the image, an atmospheric correction is necessary in which the raw radiance image is coverted to a reflectance image. In this paper the Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes (FLAASH) atmospheric correction model from ENVI software version 5.3 was adopted. The FlAASH modelis based on the MODREAN radiative transfer model and is widely used in hyperspectral[12,13] and multispectral imagery. During the correction process, Tiled Processing was not performed.
(3) Band selection. To obtain well results, the bands without calibration, contaminated by strips and overlapped between VNIR (1006.68-1028.98nm) and SWIR (1004.57-1029.85 nm) were removed. Additionally, considering the difference of the diagnostic spectral characteristics of anions and cations, we selected the data from 400-1000 nm and 2058-2361 nm respectively.
(4) Destriping. The problems of stray light interference, slit contamination, and instrument instability casued stripe noise. To remove the obvious vertical stripes, the global destriping method was adopted. After processing, the vertical stripes of the image can be effectively eliminated (Fig. 3).
(5) Orthorectification. Orthorectification, as the highest level of geometric correction, can simultaneously remove the effects of image perspective and relief effects for the purpose of creating a planimetrically correct image. After correction, the AHSI imagery was georeferenced to a UTM projection using the WGS-84 datum (UTM zone 49N).
(6) Spectral denoising. To improve the signal to noise ratio (SNR), The first 15 Minimum Noise Fraction (MNF) bands were selected for inverse MNF rotation. MNF rotation can be used to judge the intrinsic dimension of imagery data, separate noise, and reduce the demand for processing and calculation, which is extensively used in hyperspectral research.
Alteration mineral mapping. Alteration mineral mapping is one of the main research contents of HRS, which has been extensively studied[2,4-10,23,24]. In general, the diagnostic absorption features of anions are mainly located in the 2000-2500 nm range, and the diagnostic absorption features of Fe2+, Fe3+ and Mn2+ are mainly located in the 400-1200 nm range[25,26]. Therefore, The data of two spectral range was used to extract the altered minerals respectively. The typical mineral mapping process is the earliest technique for hyperspectral information extraction and has been extensively used in mineral mapping[5,9,27,28], which is implemented in the ENVI software.
After MNF, PPI and n-dimensional visualizer processing, the endmembers of the VNIR and SWIR spectral range were extracted respectively. A total of five endmember spectra were obtained in the VNIR spectral range. To identify these spectra, the Spectral Analyst model in ENVI software was used by comparing their spectral signatures with these from the USGS spectral library. Additionally, expert knowledge was also adopted to comprehensively analyze the characteristic absorption position, absorption depth, absorption width and overall shape of each endmember spectrum. After that, one goethite spectrum was finally identified in the VNIR spectral range, which is highly consistency with the USGS reference spectrum on waveform and absorption position (Fig. 4(a)). In the SWIR region, A total of seven endmember spectra were obtained, and the characteristic absorption position, absorption depth, absorption width, waveform and the analysis results from the Spectral Analyst model were comprehensively considered. Finally, four altered mineral endmembers were obtained. Among them, there are two endmember spectra that are similar in waveform, characteristic absorption position and absorption depth. Therefore, These two endmember spectra were merged.
Numerous studies have shown that the aluminum hydroxyl (Al-OH) absorption of sericite (white mica) near 2200 nm can be used to estimate its AlVI content[15,29-33]. And the absorption position varies from 2190 nm to 2225 nm, which can be used to distinguish Al-poor, Al-medium and Al-high sericite[15,16]. However, there is still no cosistent metric for distinguishing different aluminum content sericite (wihte mica). For instance, some researchers distinguished long wavelength micas and short wavelength micas with 2220 nm as the boundary. However, some researchers used the primary absorption characteristic of Al-OH at 2195 nm, 2210 nm and 2225 nm in combination with the common secondary absorption characteristic at 2345 nm to identify Al-high, Al-medium, and Al-poor sericite respectively. In this paper, the three endmembers in the SWIR region have common secondary absorption characteristic near 2345 nm, and the main absorption characteristic appearing at 2201 nm, 2210 nm, and 2218 nm, respectively. Based on it, the Al-high, Al-medium and Al-poor sericite are identified (Fig. 4(b)).
Mineral mapping is one of the ultimate aim of the utilization of hyperspectral data in mineral exploration. After acquiring and identifying the endmember spectra, the mapping algorithm can be used to analyze the relationship between the reference spectra and image spectra, so as to achieve alteration mineral mapping and inversion of their abundance Mapping algorithms can be subdivided broadly into per pixel and subpixel methods. Among them, pixel-based spectral angle matching (SAM) and spectral information divergence (SID), and subpixel-based hybrid modulation matched filtering (MTMF) and matched filtering (MF) algorithms have been extensively applied. MF is a partial unmixing or spectral decomposition technique to find the abundances of target(s) of interest in each pixel of a hyperspectral image, which can maximize the response of the target of interest and suppress the response of the compound unknown background. And the output is the matched filter score. The closer the score is to 1, the better the pixel spectrum matches the reference spectrum.
Threshold segmentation is commonly used to separate and extract the anomalous information by intensity. In this study, The threshold values of MF outputs were mean + 2 (standard deviation). Additionally, a median filtering with 3 × 3 pixels window was applied to remove isolated anomalous pixels and optimize the segmentation results.
Texture information extraction. Texture is an important image spatial feature, which refers to the frequency of tonal change in an image. Surface rocks show specific textures on the image due to the combined effect of internal minerals and structures, as well as external geological structures, weathering and denudation. The author believes that the texture in the area with less artificialities mainly reflects the information such as geological structure and lithology, including faults, joints, veins, rock bedding and contact plane. Therefore, the direct use of texture information instead of geological structures may improve work efficiency, while partly avoiding the influence of subjectivity during visual interpretation of geological structures. Additionally, compared to automatic recognition of geological structures, it can retain more original information.
Many texture descriptors have been developed in the past, and the gray-level co-occurrence matrix (GLCM) is the most commonly implemented. GLCM provides information in image gray direction, interval and change amplitude, so that 14 kinds of texture features can be effectively defined based on it. In this paper, based on the PC 1 obtained from PCA of VNIR data (400-1000 nm) and the GLCM with 3 × 3 pixels window, the contrast textural measure was calculated from four directions of 0°, 45°, 90° and 135°, respectively. And the final texture image of the study area was calculated by the average operation. To obtain texture complexity, the texture image was first converted to a binary image by the threshold value selected by the interpreter by observing arbitrary transects on the texture image. And then the texture complexity was generated by converting raster to point, deleting zero-value points and generating point density in ArcGIS software.