Using hyperspectral remote sensing data for identifying geological and soil units in the Alora Region, Southern Spain Zenga, Y., Bartholomeus a, H.M., De Bruina, S., Epemaa, G.F., Clevers a, J.G.P.W. a
WUR-CGI, Droevendaalsesteeg 3, 6708 PB Wageningen, email:
[email protected]
ABSTRACT Optical remote sensing can be used for identifying geological and soil units, since the soil surface in combination with limited fieldwork gives enough information to distinguish major soil units. The main purpose of this research is to investigate the possibilities of mapping soil and geology in the partly vegetated Alora region with an acceptable accuracy. For this purpose DAIS-images in combination with spectral angle mapping (SAM) are used. This results in a 9-class soil and geology classification with an overall accuracy of 73 %. The SAM algorithm performed good, even pixels up to a ground cover of 90%, were classified correctly. However more research has to be done, especially on a larger dataset, the applicability of SAM for the determination of soil and geology appears to be promising in partly vegetated areas. Keywords: Spectral Anlge Mapping, Soil, Geology, Vegetation
1 INTRODUCTION Mapping soils and geology is a time and money consuming activity. For an exact identification a conscientious study of many profiles is required. This information has to be extrapolated to achieve a spatially continuous map of the soil units and geology. Remote sensing is a promising technique above field and statistical mapping techniques to map a large area. Aerial photographs have been used since the 50’s, nowadays more advanced systems are available. Unfortunately the possibilities to acquire information about the subsurface are limited to general information about moisture content and structure by radar [2]. Still optical remote sensing can be used, since the soil surface in combination with limited fieldwork gives enough information to distinguish major soil units. Imaging spectrometry, images with a large number, narrow, contiguous spectral bands, has proven to be a useful tool for identification of soil and geological surface information [3]. SAM has been used successfully in many research projects for classifying geology and soil by hyperspectral imagery. [8] used spectral angle mapping (SAM) to map the minerals in Nevada. [9] and [1] used the same method and AVIRIS images to investigate the mineral distributions in Cuprite and Southern Cedar mountain area. However this research is mainly done in bare areas. The main purpose of this research is to investigate the possibilities of mapping soil and geology in the partly vegetated Alora region (Spain) with an acceptable accuracy. For this purpose DAIS-images in combination with SAM are used.
2 METHODS AND MATERIALS 2.1.
Area description
The study area is located near the village Alora, 15 kilometers west of Malaga, Southern Spain and covers an area of approximately 150 km2. The climate is dry Mediterranean. Elevation ranges from 80 m above
Presented at the 3rd EARSeL Workshop on Imaging Spectroscopy, Herrsching, 13-16 May 2003
297
sea level in the Guadalhorce valley, to 1195 m on the top of the Sierra de Valle de Abdalajis in the Northern part of the area. Geologically, the area is part of the Betic Cordillera, an Alpine mountain range. The Sierra de Aguas (northwest of Alora) and the Montes de Malaga (southwest of Alora) consist of metamorphic rocks, respectively belonging to the Alpujarride and the Malaguide complexes. The southern part of the Sierra de Aguas contains an intrusion of peridotites and serpentinites. The Sierra de Valle Abdalajis mainly consists of Jurassic limestone and the gently sloping hills on both sides of the river valley consist of Cretaceous marls and flysch deposits. At some places, large Miocene conglomerate outcrops occur, e.g. El Hacho near Alora. Quaternary fluvial deposits are present on the level floodplain and terraces of the Guadalhorce river [6][7]. Agricultural land use in the area is closely related to geology and topography [4]: cultivation of surfaceirrigated citrus is concentrated on the floodplain and terraces of the Guadelhorce river. Some drip-irrigated citrus cultivations can be found on sloping landforms. Rain fed arable crops make up the main land use on the gentle slopes of flysch deposits; olives and almonds cover the higher and steeper parts. The infertile conglomerates and limestones are covered by herbaceous vegetation with scattered shrubs and trees. These are used for extensive grazing. On the peridotites and serpentinites a reforestation project is being carried out.
2.2.
Materials
On the 28th of June 2001, airborne hyperspectral images were taken of the area around Alora. Two flight paths were flown with the DAIS 7915 (Digital Airborne Imaging Spectrometer) sensor, which measures the surface reflectance in 79 spectral bands. The first 72 bands cover the atmospheric windows between 450 and 2450 nm. The other 7 bands cover the thermal infrared region, but are not used during this research [5]. Measuring the reflectance in this large amount of contiguous, very small spectral bands results in a good description of the spectral response of the surface at each specific wavelength. Depending on the mineralogical composition, moisture content and amount of organic matter, the soil spectrum will show characteristic absorption features. Two field campaigns were carried out. During the first field campaign, around the time of flight, field spectra were obtained, using an ASD fieldspectrometer. This instrument measures a contiguous spectrum between 400 and 2500 nm, with a bandwidth of 1 nm. Measurements were taken of three reference fields for the radiometric correction of the DAIS-images. Moreover spectra were measured of 15 plots with varying soil types. The second field campaign was carried out one year after the data-acquisition. Surface descriptions of several homogeneous plots were made to use as training dataset. Besides that, reference plots were described in regard to soil-type and ground cover, in order to assess the reliability and accuracy of the tested mapping method. To do so not only homogeneous spots were described, but also spectrally mixed locations were taken into account, in order to test the robustness of the method.
2.3.
Methods
Spectral Angle Mapping (SAM) is a technique, which uses the characteristic absorption dips of the soil spectrum. Since the reflectance information obtained by a sensor can be considered as an N-dimensional dataset, in which N is the amount of spectral bands, the reflectance of each pixel will have it’s specific place within this N-dimensional space. The location of the pixel can thus be described with a vector from
298
the origin to the pixel. SAM compares this vector with the vector of a reference spectrum by calculating the angle between both vectors. The smaller the spectral angle, the better the match between the two spectra. Instead of one reference spectrum a larger set of pure spectra, so-called endmembers, can be used to map the area.
3 RESULTS 3.1.
Endmember spectra selection
A set of 9 endmembers was made. The greater part of this set was derived from the image; only 3 were selected from field measurements. The advantage of choosing endmembers from the image is the absence of atmospheric differences and differences in moisture content between the moment of image acquisition and the moment of measuring the field spectra.
3.2.
Spectral Angle Mapping
Applying the SAM algorithm on the DAIS image results in a spectral angle image. A small angle indicates a good resemblance between the two spectra; therefore the dark areas in the SAM image show the places where the endmember concerned probably occurs. To come to a final soil type classification a pixel will be assigned to the endmember with the smallest spectral angle. This results in a 9-class soil and geology map (figure 1). The classification has an overall accuracy of 73 %. A confusion matrix is shown in table 1. A small overestimation of the terrace area is shown. Both limestone (31%) and conglomerate (50%) are not classified accurately. All other classes show a high accuracy, mostly 100%. Also a parking lot, of which a spectrum was measured but not used during analysis, was put in the correct class, namely unclassified. This indicates that the threshold values were correctly chosen. SAM classes Fieldwork classes 1. Limestone 2. Conglomerate 3. Sandstone 4. Gneiss 5. Serpentinite 6. Schist 7. Flysch 8. Terrace 9. River bed 10. Unclassified/ Parking lot
1. L 4
2 3. C S
4. G 1
5. S 1
6. S
7. F
5 1
18
2
8. T 5 2
9. R
10. U 2
1 3 1
3 3
1 7 2 1
Total 13 4 1 3 4 5 23 7 2 1 46/ 63 =73%
Table 1. Confusion matrix of the geological classification result
In figure 2 the classification accuracy is plotted against the amount of vegetation cover. As expected the accuracy decreases when the vegetation cover reaches a certain level, in this case more than 90%. The accuracy is low when vegetation cover is between 50 and 60%. This class only contains two observations.
299
SAM classification accuracy related to % vegetation cover
% accuracy
100 80 60 40 20 90100
80-90
70-80
60-70
50-60
40-50
30-40
20-30
10-20
0-10
0
% vegetation cover
Figure 2. SAM classification accuracy related to % vegetation cover
4 CONCLUSIONS AND DISCUSSION The SAM algorithm performed good, when mapping soils and geology in the partly vegetated Alora region. Even pixels up to a ground cover of 90%, were classified correctly. However more research has to be done, especially on a larger dataset. The applicability of SAM for the determination of soil and geology appears to be promising in partly vegetated areas. The relation between the fraction of vegetation cover and the accuracy of the classification requires further research. In order to examine the accuracy properly more ground truth is required. The sandstone area for example only contains one control point, which is insufficient for a reliable accuracy assessment. For the same reason it is hard to attach a value judgment to the relation between the classification accuracy and the amount of vegetation cover. The profile in figure 2 is as expected, the accuracy should decrease when the amount of vegetation exceeds a certain value. The low accuracy between 50% and 60% is a result from the low amount of observations within this class. Nonetheless from these first results it can be concluded that SAM is a useful technique for mapping soils and geology in a vegetated area.
REFERENCES [1] BAUGH, W.M. & KRUSE, F.A., ATKINSON, W., 1998: Quantitative geo-chemical mapping of ammonium minerals in the southern Cedar Mountains, Nevada, using airborne visible/infrared imaging spectrometer (AVIRIS). Remote Sensing of Environment. 65, pp. 292-308. [2] BINDLISH, R., BARROS, A.P., 2001: Parameterization of vegetation backscatter in radar-based, soil moisture estimation. Remote Sensing of Environment. 76: pp. 1-130. [3] CHABRILLAT, S., PINET, P.C., CEULENEER, G., JOHNSON, P.E., MUSTARD, J.F., 2000: Ronda peridotite massif: methodology for its geological mapping and lithological discrimination from airborne hyperspectral data. International Journal of Remote Sensing, Vol. 21, No. 12, pp. 2363-2388 [4] DE BRUIN, S., GORTE, B.G.H., 2000: Probabilistic image classification using geological map units applied to land-cover change detection. International Journal of Remote Sensing, Vol. 21, No.12, pp. 23892402. [5] HABERMEYER, M.H., 2001: Airborne Imaging Spectroscopy at DLR. DAIS homepage: http://www.op.dlr.de/dais/dais.htm [6] INSTITUTO GEOLOGICO Y MINERO DE ESPANA (IGME), 1978: Mapa Geologico de Espana 1:50.000, No. 105 (Alora). Servico de Publicaciones Ministerio de Industria Madrid. [7] INSTITUTO TECNOLOGICO GEOMINERO DE ESPANA (ITGE), 1991: Mapa Geologico de Espana 1:50.000, No. 1038 (Ardales). ITGE, Madrid.
300
[8] MIYATAKE, S., LEE, K., 1997: Mapping alteration minerals in northern Cuprite and Goldfield, Nevada, with JERS-1 OPS data. Proceedings of the Twelfth International Conference and Workshop on Applied Geologic Remote Sensing, Denver, Colorado, 17-19 November, Volume II, pp. 17-37 [9] YANG, H., VAN DER MEER, F., BAKKER, F., TAN, Z.J., 1999: A back-propagation neural network for mineralogical mapping from AVIRIS data. International Journal of Remote Sensing, 20(1), pp. 97-110.
301