International Journal of Applied Earth Observation and Geoinformation 13 (2011) 616–625
Contents lists available at ScienceDirect
International Journal of Applied Earth Observation and Geoinformation journal homepage: www.elsevier.com/locate/jag
Land cover mapping at Alkali Flat and Lake Lucero, White Sands, New Mexico, USA using multi-temporal and multi-spectral remote sensing data Habes A. Ghrefat a,∗ , Philip C. Goodell b a b
King Saud University, Department of Geology and Geophysics, P.O. Box 2455, Riyadh 11451, Saudi Arabia University of Texas at El Paso, Department of Geological Sciences, El Paso, TX 79968, USA
a r t i c l e
i n f o
Article history: Received 18 March 2010 Accepted 31 March 2011 Keywords: Spectral mixture analysis Landcover Mapping Principal components Minimum-noise fraction Multi-dimensional data processing
a b s t r a c t The goal of this research is to map land cover patterns and to detect changes that occurred at Alkali Flat and Lake Lucero, White Sands using multispectral Landsat 7 Enhanced Thematic Mapper Plus (ETM+), Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER), Advanced Land Imager (ALI), and hyperspectral Hyperion and Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) data. The other objectives of this study were: (1) to evaluate the information dimensionality limits of Landsat 7 ETM+, ASTER, ALI, Hyperion, and AVIRIS data with respect to signal-to-noise and spectral resolution, (2) to determine the spatial distribution and fractional abundances of land cover endmembers, and (3) to check ground correspondence with satellite data. A better understanding of the spatial and spectral resolution of these sensors, optimum spectral bands and their information contents, appropriate image processing methods, spectral signatures of land cover classes, and atmospheric effects are needed to our ability to detect and map minerals from space. Image spectra were validated using samples collected from various localities across Alkali Flat and Lake Lucero. These samples were measured in the laboratory using VNIR–SWIR (0.4–2.5 m) spectra and X-ray Diffraction (XRD) method. Dry gypsum deposits, wet gypsum deposits, standing water, green vegetation, and clastic alluvial sediments dominated by mixtures of ferric iron (ferricrete) and calcite were identified in the study area using Minimum Noise Fraction (MNF), Pixel Purity Index (PPI), and n-D Visualization. The results of MNF confirm that AVIRIS and Hyperion data have higher information dimensionality thresholds exceeding the number of available bands of Landsat 7 ETM+, ASTER, and ALI data. ASTER and ALI data can be a reasonable alternative to AVIRIS and Hyperion data for the purpose of monitoring land cover, hydrology and sedimentation in the basin. The spectral unmixing analysis and dimensionality eigen analysis between the various datasets helped to uncover the most optimum spatial–spectral–temporal and radiometric-resolution sensor characteristics for remote sensing based on monitoring of seasonal land cover, surface water, groundwater, and alluvial sediment input changes within the basin. The results demonstrated good agreement between ground truth data and XRD analysis of samples, and the results of Matched Filtering (MF) mapping method. © 2011 Elsevier B.V. All rights reserved.
1. Introduction Monitoring the spatial and temporal changes of land cover in semi-arid and arid land regions is of importance for hydrologists, ecologists, and agronomists. Land cover information is used by hydrologists to update surface conditions affecting stream flow, infiltration and evapotranspiration (Su, 2000). Agronomists used information of land cover for acreage and yield prediction (Fang, 1998). The relationships between land degradation, human activities, and global climate change are utilized by ecol-
∗ Corresponding author. E-mail address:
[email protected] (H.A. Ghrefat). 0303-2434/$ – see front matter © 2011 Elsevier B.V. All rights reserved. doi:10.1016/j.jag.2011.03.009
ogists (Chehbouni et al., 2000). Detection of changes in the land cover involves use of at least two period data sets (Jensen, 1986). The use of multi-temporal images not only results in higher classification but also gives consistent in all classes being mapped. Multi-temporal data is especially advantageous in areas where vegetation or lands use changes rapidly. The capability of capturing changes in land cover and extracting the change information from satellite data requires effective and automated change detection techniques (Roy et al., 2002; Shalaby and Tateishi, 2007). The mapping of the spatial distribution of land cover types in playa lakes using airborne and spaceborne remotely sensed were conducted by few researchers (Chapman et al., 1989; Crowley, 1993; Bryant, 1996; Crowley and Hook, 1996; Giri et al., 2005;
H.A. Ghrefat, P.C. Goodell / International Journal of Applied Earth Observation and Geoinformation 13 (2011) 616–625
617
Fig. 1. Landsat 7 ETM+ image showing the location of the study area and the samples locations.
Shalaby and Tatersh, 2007; Giannico, 2007). These environments are difficult environments in which to undertake fieldwork due to problems of poor accessibility (Bryant, 1996). Further hindrances to field work include the subdued topography of many playa lakes, the sporadic nature of flooding of their surfaces and their harsh climate. Remote sensing data are useful for overcoming these problems in the playas environment. Second, multispectral and hyperspectral data make large scale land cover type’s identification and mapping the spatial distribution of these types feasible (Chapman et al., 1989; Crowley, 1993; Crowley and Hook, 1996; Bryant, 1996). The goal of this study was to evaluate multi-temporal and multi-spectral data from Landsat 7 ETM+, ASTER, and ALI, and hyperspectral data from Hyperion and AVIRIS for detecting and mapping various land cover classes at Alkali Flat and Lake Lucero, New Mexico. Third, it is relatively easy to obtain cloud free airborne and spaceborne imagery over arid areas. Specific objectives of the current study are: (1) to compare the information dimensionality limits of Landsat 7 ETM+, ASTER and ALI data with high spectral resolution, low signal-to-noise Hyperion data, using AVIRIS data as a high spectral resolution, high signal-to-noise standard, (2) to determine the spatial distribution and fractional abundances of land cover endmembers in the study area using a variety of multi-dimensional image processing methods, and (3) to check ground correspondence with satellite data. A greater understanding of the spatial and spectral resolution of these sensors, optimum spectral bands and their information contents, appropriate image processing methods, spectral signatures of land cover classes, and atmospheric effects are needed to our ability to detect and map minerals from space.
2. Literature review Recently, a major effort has been made to study and monitor land cover classes using satellite multispectral sensors such as Landsat, ALI, and ASTER (Castan˜eda et al., 2005; Stefanov and Netzband, 2005; Xu and Gong, 2007; French et al., 2008; Bakr et al., 2010; Kahya et al., 2010). The capability of ALI data were evaluated for discriminating different land cover classes in Fremont, California, USA using a Mahalanobis distance classifier (Xu and Gong, 2007). French et al. (2008) used thermal infrared observations from ASTER data to detect land cover changes at the Jornada Experimental Range, New Mexico, USA. Spaceborne and airborne hyperspectral sensors, such as AVIRIS and Hyperion were also used for land cover mapping (Crowley, 1993; Pignatti et al., 2009). Hyperspectral Hyperion data were evaluated for discriminating and mapping land cover classes in Fremont, California (Xu and Gong, 2007), and in a Pollino National Park, Italy (Pignatti et al., 2009). Linear Spectral Unmixing (LSU) technique was applied to Hyperion data to derive the abundance fractions of land cover endmembers. Table 1 Acquisition dates for remote sensing data used in this study. Sensor
Acquisition data
Landsat 7 ETM+ ASTER ALI Hyperion AVIRIS
September 12, 1999 May 9, 2000 December 6, 2001 December 6, 2001 October 10, 2002
618
H.A. Ghrefat, P.C. Goodell / International Journal of Applied Earth Observation and Geoinformation 13 (2011) 616–625
Table 2 Summary characteristics for AVIRIS, Hyperion, ALI, ASTER, and Landsat 7 ETM+. Sensor
Subsystem
Band no.
Spectral range (m)
Radiometric resolution (S/N)
Spatial resolution (m)
Swath width (km)
AVIRIS
VNIR SWIR VNIR SWIR
Continuous Continuous Continuous Continuous Pan 1 2 3 4 5 6 7 8 9 1 2 3N 3B 4 5 6 7 8 9 Pan 1 2 3 4 5 7
0.38–0.900 0.900–2.500 0.400–0.900 0.900–2.400 0.480–0.690 0.433–0.453 0.450–0.515 0.525–0.605 0.633–0.690 0.775–0.805 0.845–0.890 1.200–1.300 1.550–1.750 2.080–2.350 0.52–0.60 0.63–0.69 0.78–0.86 0.78–0.86 1.600–1.700 2.145–2.185 2.185–2.225 2.235–2.285 2.295–2.365 2.360–2.430 0.52–0.90 0.450–0.515 0.525–0.605 0.630–0.690 0.750–0.900 1.550–1.750 2.090–2.350
800/1 500/1 161/1 50/1 160/1 270/1 410/1 380/1 340/1 230/1 210/1 140/1 190/1 180/1 370/1 306/1 202/1 183/1 466/1 254/1 229/1 234/1 258/1 231/1 225/1 255/1 275/1 280/1 223/1 155/1 155/1
18.5
10.5
30
7.7
Hyperion
ALI
VNIR
SWIR
ASTER
VNIR
SWIR
Landsat 7 ETM+
VNIR
SWIR
36 30
15 60 30
15
185 30
Mapping land cover on playas by remote sensing techniques is much less frequent and it is usually tested with in situ observations. Chapman et al. (1989) applied decorrelation stretch methods to Landsat TM data in order to map halite, gypsum and their mixtures in Chilean salt flat. Epema (1990) defined several surface types within Tunisian playas by comparing Landsat TM images and simultaneous field reflectance measurements. These surface types represented various combinations of soil moisture, roughness and chemistry. Crowley (1993) used Airborne Visibile/Infrared Imaging Spectrometer data (AVIRIS) and least-square spectral band-fitting methods to map efflorescent salts such as three borate minerals, hydroboracite, pinnoite, rivadavite and damp halite at Death Valley, California. In addition, Crowley (1993) also showed that crusts composed of anhydrite, glauberite, and thenardite were not mapped because of their generally spectral signatures. Bryant (1996) used Landsat TM imagery and spectral mixture analysis to map gypsum, halite, vegetation, alluvial material, shade and moisture in the Chott el Djerid salt playa, Tunisia. Thermal Infrared Multispectral Scanner (TIMS) was used by Crowley and Hook (1996) to map playa evaporite minerals such as gypsum, halite, and thenardite at Death Valley, California. Castan˜eda et al. (2005) used Landsat TM and ETM+ images between 1985 and 2000 to map five surface facies classes at the Monegros desert and its saladas, Spain. These facies were identified using unsupervised classification method.
The Alkali Flat area is located to the west of the dune field of the White Sands National Monument (Fig. 1). Wide areas of Alkali Flat consist of thin layers of eolian sabkha sediments overlying an even more widespread wind scour surface atop either cemented dunes or former Lake Otero sediments. Eolian sabkhas consist of flat-lying sediments. It is commonly influenced by evaporative processes, that have been brought to the site of deposition by wind (Fryberger, 2000). Eolian sabkhas are common at White Sands due to the shallow, saline groundwater and the high annual rate of evaporation (Fryberger, 2000). Lake Lucero (Fig. 1) is the lowest point in the Tularosa Basin at an elevation of 1186 m above mean sea level (Allmendinger, 1971). Lake Lucero is a hypersaline, ephemeral playa (Allmendinger and Titus, 1973) and is located on the south western boundary in the Alkali Flat of White Sands National Monument (Fig. 1). It is the remnant of the once larger Lake Otero that existed between 10,000 and 24,000 years ago during the Pleistocene (Herrick, 1904). It is flooded during the rainy season with fresh water, and it precipitates out saline salts at the surface of the lake in the dry season (Allmendinger, 1971). At Lake Lucero, the water table is very shallow and varies from one season to another (Allmendinger, 1971). Because of its shallow depth, water with dissolved gypsum moves upward to the surface of Lake Lucero through capillary action (Fryberger, 2000).
3. Study area and geologic setting
4. Data analysis methods
The study area (Fig. 1) is located in Alkali Flat and Lake Lucero in New Mexico. The Alkali Flat area and Lake Lucero are located within the Tularosa Basin. The Tularosa Basin is a downfaulted, arid to semi-arid area covering about 17,000 km2 of south-central New Mexico. The Tularosa Basin is bounded on the east by the Sacramento Mountains, Sierra Blanca, and Otero Mesa, and is bounded to the west by the Oscura, San Andres, and Organ Mountains (Orr and Myers, 1986; Lozinsky and Bauer, 1991; Hawley, 1993).
4.1. Remote sensing datasets and resolution characteristics The acquisition dates of the datasets used in the current study are summarized in Table 1. The Landsat 7 Enhanced Thematic Mapper Plus (ETM+) Level-1G data have eight bands sensitive to various wavelengths of Visible Infrared (VNIR), Short Wavelength Infrared (SWIR), and Thermal Infrared (TIR) (Table 2) (Goward et al., 2001; Thome, 2001). Advanced Spaceborne Thermal Emission and Reflec-
H.A. Ghrefat, P.C. Goodell / International Journal of Applied Earth Observation and Geoinformation 13 (2011) 616–625
Fig. 2. Matched Filtering and Mixture Tuned Matched Filtering mapping results for the VNIR–SWIR bands (0.4–2.5 m) of Landsat 7 ETM+ covering Alkali Flat.
tion Radiometer (ASTER) Level-1B 003 data have three bands in the VNIR, six bands in the SWIR, and five bands in the TIR with 15, 30, and 60-m spatial resolution, respectively (Table 2) (Fujisada, 1995; Yamaguchi et al., 1998). The ALI Level1 data measures the electromagnetic radiation in 10 spectral bands from 0.43 to 2.35 m with spatial resolution of 10 and 30 m (Hearn et al., 2001) (Table 2). The Hyperion Level1B data have 242 spectral bands spanning the VNIR spectrum between 0.4 and 2.5 m (Table 2) (Folkman et al., 2000; Pealman et al., 2001). Hyperion has a spatial resolution of 30 m and approximately 10 nm spectral resolution. High-altitude AVIRIS simultaneously collects 224 spectral bands, each with a 10 nm bandwidth distributed over the 0.4–2.5 m range (Table 2) (Vane and Goetz, 1993; Green et al., 1998). ASTER level-1B data have been registered to UTM using a first-order rotation from the level-1A data. An image-to-image registration method was used to co-register ASTER, ALI, Hyperion, and AVIRIS, with the Landsat 7 ETM+ image used as a base image. 4.2. Atmospheric correction and reflectance calibration The Landsat 7 ETM+ Level-1G data were radiometrically and geometrically corrected to the same map projections, image orientations, and spatial resolution. The ASTER Level-1B 003 data were registered radiance at the sensor, which were radiometrically calibrated and geometrically registered using telemetry and
619
engineering data from the three imaging sub-systems during acquisition of Level-1A data. The four strips of ALI data were mosaicked into a continuous scene. The Hyperion data were radiometrically calibrated to radiance at the USGS EROS Data Center. The AVIRIS data were radiometrically calibrated to radiance at the NASA’s Jet Propulsion Laboratory (JPL). The AVIRIS scenes were mosaicked together to produce a single image. ENVI 4.4 software package was used to process the data used in this study. The empirical line (EM) approach (Roberts et al., 1985; Conel et al., 1987; Kruse, 1988, 1990) was used to calibrate Landsat 7 ETM+, and ALI data to surface reflectance. VNIR–SWIR ASTER bands 1, 2,3N, 4, 5, 6, 7, 8, and 9 were combined into a single file after the VNIR bands (1, 2, and 3N) were resampled to 30 m resolution to match that of the SWIR bands. ASTER radiance data were calibrated to surface reflectance using the ACORN atmospheric correction program (AIG, 2001). The Atmosphere Removal Program (ATREM) (Gao et al., 1993) was used to calibrate Hyperion and AVIRIS and data to surface reflectance. All 224 bands of AVIRIS data were used in the ATREM calibration, whereas 196 bands of Hyperion were selected after subsetting out overlapping and non-used bands. 4.3. Minimum Noise Fraction (MNF) eigen analysis, data dimensionality, and endmember determination MNF transformation process is a data reduction method designed to increase apparent signal-to-noise by estimating noise statistics from the data, segregating it to higher order eigen channels, while still retaining much of the original signal (Green et al., 1988). The MNF transformation is similar to Principal Components Analysis (PCA) transformation, but differs in that MNF considers the signal separately from the noise, while PCA considers the overall data variation using a single covariance matrix (Smith et al., 1985; Richards, 1994). Generally, the signal (i.e. high information content) is concentrated in the lower ordered eigenchannels, whereas the noisy bands are concentrated in the higher ordered eigenchannels. The spectral bands covering the VNIR–SWIR (0.4–2.5 m) of Landsat 7 ETM+, ASTER, ALI, Hyperion, and AVIRIS data were subjected to MNF to facilitate estimation of their “information dimensionalities”. The data dimensionality constrains the number of endmembers that can be detected and mapped. Dimensionality estimation and eigen-analysis were carried out for the Alkali Flat and Lake Lucero sites separately. We find eigen-thresholds that distinguish the lower order eigenchannels with spatial coherency, from nosier eigenchannels with no spatial coherency. The following criteria were used to determine information contents and dimensionality cutoff limits for Landsat 7 ETM+, ALI, ASTER, Hyperion, and AVIRIS: 1. The amount of spatially coherent information retained by visual inspection of each individual MNF eigenchannel. 2. The break in slope or inflection points on MNF eigenvalue versus eigenchannel plots, defining the point at which eigenvalue decreases little compared to increasing eigennumber. 3. PCA statistics to determine the overall cumulative percent variance change versus eigenchannel.
Fig. 3. Matched Filtering and Mixture Tuned Matched Filtering mapping results for the VNIR–SWIR bands (0.4–2.5 m) of Hyperion covering Lake Lucero. Water is only appeared in Hyperion and ALI data.
Once the information dimensionality of the data was determined, the Pixel Purity Index (PPI) was applied to the noisewhitened data. PPI is a data reduction technique used to identify the most spectrally pure, or extreme, pixels as vertices in ndimensional space (Boardman et al., 1995). Endmember classes of remote sensing data can be viewed graphically using n-D Visualization methods where n is the number of noise-reduced bands accounting for the majority of the scene’s spectral variability. Spectral signatures can be extracted in various units such as radiance or calibrated reflectance.
620
H.A. Ghrefat, P.C. Goodell / International Journal of Applied Earth Observation and Geoinformation 13 (2011) 616–625
Table 3 Summary of eigenvalue including information dimensionality, and MNF cumulative % variance of AVIRIS, Hyperion, ALI, ASTER, and Landsat 7 ETM+ for Alkali Flat and Lake Lucero, White Sands. Alkali Flat
Lake Lucero
Sensor
Information dimensionality thresholds
% of cum variance (MNF)
Information dimensionality thresholds
% of cum Variance (MNF)
AVIRIS Hyperion ALI ASTER Landsat 7 ETM+
24 19 9 9 6
76.3 65.2 95.8 100 100
24 16 9 9 6
57.1 60.5 92.0 100 100
4.4. Spectral mixture analysis
4.5. Field, laboratory spectral analysis and ground truth
Matched Filtering (MF) was applied to the datasets used in this study to map spectrally varying components in the study area. MF is a spectral mapping procedure that maximizes the response of known endmembers and suppresses the response of the composite unknown background using least square regression methods (Harsanyi and Chang, 1994; Boardman et al., 1995). The results of MF are as grayscale images with values ranging from 1, which provides a means of estimating the relative degree of match between a reference and image spectra, where ≥>1 is perfect fit. Because the scores of MF vary quite rapidly on a pixel-to-pixel basis, the best fit grayscale values (≥>1) are stretched for displayed as white, while the poorest fit grayscale values (
