Advanced
Development of a Fusion Vegetation Index Using Full-PolSAR and Multispectral Data
Development of a Fusion Vegetation Index Using Full-PolSAR and Multispectral Data
Journal of the Korean Society of Surveying, Geodesy, Photogrammetry and Cartography. 2015. Dec, 33(6): 547-556
Copyright © 2015, Korean Society of Surveying, Geodesy, Photogrammetry and Cartography
This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/3.0) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited
  • Received : November 11, 2015
  • Accepted : December 12, 2015
  • Published : December 31, 2015
Download
PDF
e-PUB
PubReader
PPT
Export by style
Share
Article
Author
Metrics
Cited by
TagCloud
About the Authors
Kim Yong-Hyun
Member, Dept. of Civil and Environmental Engineering, Seoul National University (E-mail:yhkeen@gmail.com)
Oh Jae-Hong
Corresponding Author, Member, Dept. of Civil Engineering, Chonnam National University (E-mail:ojh@chonnam.ac.kr)
Kim Yong-Il
Member, Dept. of Civil and Environmental Engineering, Seoul National University (E-mail:yik@snu.ac.kr)
Abstract
The vegetation index is a crucial parameter in many biophysical studies of vegetation, and is also a valuable content in ecological processes researching. The OVIs (Optical Vegetation Index) that of using multispectral and hyperspectral data have been widely investigated in the literature, while the RVI (Radar Vegetation Index) that of considering volume scattering measurement has been paid relatively little attention. Also, there was only some efforts have been put to fuse the OVI with the RVI as an integrated vegetation index. To address this issue, this paper presents a novel FVI (Fusion Vegetation Index) that uses multispectral and full-PolSAR (Polarimetric Synthetic Aperture Radar) data. By fusing a NDVI (Normalized Difference Vegetation Index) of RapidEye and an RVI of C-band Radarsat-2, we demonstrated that the proposed FVI has higher separability in different vegetation types than only with OVI and RVI. Also, the experimental results show that the proposed index not only has information on the vegetation greenness of the NDVI, but also has information on the canopy structure of the RVI. Based on this preliminary result, since the vegetation monitoring is more detailed, it could be possible in various application fields; this synergistic FVI will be further developed in the future.
Keywords
1. Introduction
The global to local vegetation patterns needs to be studied to understand effects in human societies by environmental changes and phenomena. Moreover, since the human activity has profoundly affected ecosystems, vegetation changes must be monitored and predicted ( Pettorelli , 2005 ). Remote sensing data acquired from different sensors have been commonly used to monitor and characterize vegetation information ( Shi ., 2008 ). Among the various parameters, the VI (Vegetation Index) derived from satellite data is the most commonly used for analyzing the characteristics of a vegetated land surface ( Zhang, 2015 ). The VI is a mathematical combination or transformation of specific bands that accentuate the properties of green plants in order to distinguish them from other land covers ( Jackson and Huete, 1991 ).
Two types of VIs are used for vegetation applications. The first VI is an OVI that uses a multispectral and hyperspectral sensor. It is not only a key source of information in vegetation conditions but also in forest studies and crop monitoring ( Kuenzer and Knauer, 2013 ). The use of the OVI has been proposed for determining the vegetation conditions from different portions of the wavelength. Among the various OVIs, the NDVI is a popular and standard index ( Gao, 1996 ). It is increasingly used in indirectly study of the biophysical properties of vegetation. However, the NDVI has a few problems, such as soil background variations and saturation in the high vegetation densities ( Huete, 1988 ).
In microwave remote sensing, the RVI was introduced as an index of the volume scattering media, such as vegetation canopies ( Kim and Van Zyl, 2009 ). This parameter also measures the randomness in the scattering. The RVI was evaluated by modeling a vegetation canopy by collecting random oriented cylinders with different lengths and diameters. It has been used as the level of crop growth indicator, particularly when time series data is available ( Kim , 2012 ). In particular, RVI measures the vegetation structure, which is independent of the vegetation greenness and vigor, both of which properties measured by OVIs ( McColl ., 2014 ).
The studies on OVI and RVI have had some promising results for various applications. The OVI is mainly related to the photosynthetic activity of vegetation, while the RVI is primarily related to the canopy structure of vegetation. In other words, the OVI and the RVI of representing complementary information are useful for vegetation monitoring, but there was less efforts in fusing them as a single VI. Furthermore, the possible synergic fusion of the OVI and RVI has not been explored, although it has been expected to contribute in vegetation monitoring. To address this issue, we present a novel FVI that utilizes the NDVI and the RVI as a single VI. To fuse the NDVI and RVI, the proposed FVI was firstly simulated for numerical analysis. At next, experiments were conducted by using C-band Radarsat-2 and RapidEye data in two different areas. Finally, the experiment results were evaluated and demonstrated that the FVI has high separability in different types of vegetation, unlike the NDVI or the RVI. This paper organized by following sections; in section 2, the NDVI and RVI will be briefly introduced with presenting the proposed FVI, while each of the section 3 and 4 will draw results and discussions and conclusions.
2. Methodology
This section reviews the NDVI and the RVI with brief discussion their characteristics, so as introduces a novel FVI. For achieving ideal status, the reflectance values to compute the NDVI must be radiometrically and atmospherically corrected. Thus, the RapidEye data, the FLAASH (Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes) algorithm was used to correct these effects ( Matthew ., 2000 ). Additionally, a Refined Lee filter was applied to reduce the speckle noise of Radarsat-2 ( Lee ., 1991 ). An increased filter size could reduce the speckle noise, and also lowers the spatial resolution. For that reason, we selected a 7x7 filter size.
- 2.1 Normalized difference vegetation index
Those OVIs are usually dimensionless measurement that derived from radiometric data, which are primarily used to indicate the amount of green vegetation. Among the various OVIs, the NDVI has been the most widely used in various applications. The NDVI indicates the difference between the NIR (Near-InfraRed) and the red bands, divided by their sum. It is defined as.
PPT Slide
Lager Image
where ρNIR and ρRed are the reflectance values of the NIR and red bands, respectively. The NDVI, which depends mainly on the green leaf material of the vegetation cover, is a foundation of the most recent indices ( Huete, 1988 ). Although the NDVI remains one of the most effective VI, it has some disadvantages; it is the inherent non-linearity of ratio-based indices, and vulnerable to additive noise effects. Also, it indicates saturated values over high biomass conditions. Moreover, a major limitation of the NDVI and those similar indices is that the optical sensors can only monitor a very thin layer of the canopy ( Shi ., 2008 ). In fact, those problems can be solved by fusing the NDVI with the RVI.
- 2.2 Radar vegetation index
In the microwave regions of electromagnetic spectrum, the intensity of the incident energy scattered by vegetation is primarily a function of the canopy architectures, such as the size, shape, and orientation of the canopy components and the dielectric properties ( Koppe ., 2013 ). The RVI has proposed in applying to estimate vegetation properties relevant to radar-only soil moisture retrieval algorithms ( Kim and Van Zyl, 2009 ). The RVI is a measure of volume scattering that is typically caused by the structural elements of vegetation canopies, and is defined as;
PPT Slide
Lager Image
where σHV nis the cross-polarization backscattering cross-section, while σHV and σW refer the co-polarization backscattering cross-sections represented in power units. The RVI generally ranged from 0 to 1. It is near zero for a smooth bare surface, but it is increasing following to the vegetation growth. Also, it has shown some promising results in the vegetation field, especially in crop monitoring ( Kim ., 2012 ). Unfortunately, the RVI also has some disadvantages; the speckle noise occurs in the SAR (Synthetic Aperture Radar) data. It is an inherent consequence in the coherent radar, which arises due to electromagnetic interactions with multiple scattering centers within a single resolution element of the radar rather than actual variation. The RVI showed vulnerable particularly to errors in the calibration offset term over sparsely vegetated regions and overestimated in some arid regions ( McColl ., 2014 ). Also, the RVI can result in an incorrect value in urban areas due to the POA (Polarization Orientation Angle) shifts. In other words, if the direction of normal to vertical wall of a building is off the incidence plane, the dihedral structure contributes to the yielding of the cross-polarization components ( Iribe and Sato, 2007 ). For that reason, the POA shifts lead to high RVI values in some oriented urban areas. In fact, those problems of the RVI can be solved by fusing it with the NDVI.
- 2.3 Fusion vegetation index
The VI should theoretically reflect those three; the amount of vegetation, the architecture of the vegetation, and the degree of vegetation vigor ( Zhang ., 2007 ). To benefit from both optical and SAR data, this paper presents an FVI that uses complementary information from both systems. In vegetation monitoring, the important goal of the using the VI is on enhanced the measurement of the greenness and the vegetation structure. The greenness information and the photosynthetic activity can be recognized from the OVIs, such as the NDVI. Additionally, the canopy architectures that may prove useful in distinguishing between different types of canopies can be extracted from the RVI. However, the RVI has speckle noise and often indicates urban areas as being vegetation areas. The first problem is naturally in the SAR system, but the second problem is an unwanted false alarm. Therefore, the direct sum of NDVI and RVI may not be suitable for fusing the NDVI and RVI, which led us assuming that the FVI can be obtained by the injection of necessary RVI into the NDVI. As the FVI can be less noisy than the RVI, we can tell the assumption is reasonable. The FVI can have lower values in the POA shifted urban areas due to below 0 representing urban area in the NDVI. In fact, Many mathematical representations of VIs are available. Similarly, many mathematical formations of FVI are possible. To fuse the NDVI with the RVI, this study proposes a novel FVI that can be formulated as follows:
PPT Slide
Lager Image
wherein α is the modulation coefficient for the trade-off. Because the maximum value of the NDVI is 1, RVI+ α must be greater than or equal to 1 to effectively enhance the vegetation structures. Equation (3) can generate the FVI by injecting the canopy information of the RVI into the NDVI. If the α values increases. the FVI converges towards the NDVI. If RVI+ α =1, the FVI is equal to the NDVI, while the FVI ranges from +( α +1) to -( α +1) in non-negative α values. Therefore, we expect a reasonable contrast in the FVI between the vegetated and non-vegetated land-covers. To our knowledge, this is the first attempt to fuse the RVI and the NDVI. This fusion could suggest a solution to the saturation problem of the NDVI. In the next subsection, we simulate the FVI with changes in the α values.
- 2.4 Simulation results
A numerical analysis was conducted to simulate the FVI in accordance with the changes in the α values. values. The Fig. 1 described the theoretically discussed results with the contour plot of showing a constant FVI along the RVI and the NDVI. The contour plot in Fig. 1 reflects that contour lines of the FVI are symmetrical to the NDVI and equal to the zero line. In addition, the FVI converges towards the NDVI, as the α values increase. Thus, a considerable α value is meaningless and does not represent any RVI information. It was observed from that the increased RVI+ α value led to a large FVI and that the rate of growth of the FVI was higher for lower values of the RVI (see α =1 in Fig. 1 ). In α =0, the FVI does not have any information of the RVI in less than 0.2 RVI. Also, in α =0, the speckle noise of RVI could be directly inserted into the FVI, whereas the maximum value of RVI and NDVI is 1. To adequately enhance the vegetation condition, the appropriate α value was set to fuse the NDVI with the RVI. In this study, since the simulation results showed that the α =1 value effectively injected the RVI information into the FVI information, we set α =1. In the following section, the experiment results for the two study areas are presented and discussed.
PPT Slide
Lager Image
Contour plot of FVI corresponding to change in the α values with the RVI and the NDVI
3. Results
- 3.1 Study areas
The Study Area 1 is Dangjin City in South Korea. The test site mainly covers a large number of agricultural fields, as well as several forested areas, urban areas, and bodies of water. In this area, the RapidEye data were acquired on August 5, 2012, as the Radarsat-2 data were acquired on August 12, 2012. The Study Area 2 is Daejeon City in South Korea, where includes mainly urban areas, as well as forested areas and bodies of water. In this area, the RapidEye data and the Radarsat-2 data were acquired on May 5, 2013 and May 3, 2013, respectively. Those two Study Areas are shown in Fig. 2 ; those RapidEye and Radarsat-2 have short time lags in these two datasets that are suitable for a multisensory fusion study. Also, the Radarsat-2 single-look complex data were georeferenced to the UTM (Universal Transverse Mercator) Zone 52 South projection with the WGS-84 (World Geodetic System 1984) as the datum, using the range-Doppler model and the SRTM (Shuttle Radar Topography Mission). The Radarsat-2 has the same pixel spacing as RapidEye orthorectified product 3A that led to the assumption of a priori geometrically registered and superimposed RapidEye and Radarsat-2 data.
PPT Slide
Lager Image
Used RapidEye and Radarsat-2 (a) RapidEye (NIR-G-B color composite), (b) the Radarsat-2 (Pauli R-G-B combination) in study area 1, and (c) RapidEye (R-G-B color composite) and (d) Radarsat-2 (Pauli R-G-B combination) in study area 2
- 3.2 Results and discussion
The FVI was compared with the NDVI and the RVI to present a general understanding of the FVI, NDVI, and RVI from the RapidEye and Radarsat-2 data. To explore the results, the histograms of FVI were compared with those of the NDVI and the RVI. The histograms of Study Areas 1 and 2 are shown in Fig. 3 . The FVI histograms are similar to the NDVI histograms, but slightly differ. In Study Area 1, the frequencies of lower FVI values were increased because the vegetation which has high NDVI values and low RVI values shifted to lower FVI values. It could be inferred that this vegetation is a rice crop, which has a smaller canopy structure than the forest. In Study Area 2, the frequencies of lower FVI values were also increased because of the volume scattering difference in the vegetation with high NDVI values. In the visual analysis, the FVI result of Study Area 1 showed greater color variation and dynamic range than results of the NDVI and the RVI, so as those FVI results of Study Area 2. The RVI results of two Study Areas demonstrated some speckle noise despite the application of speckle filtering. In Study Area 2, the RVI results showed false alarms in urban areas with POA shifts. In Fig. 4(f) , the dashed black circle refers to urban areas, but the RVI depicts these areas as vegetation areas.
PPT Slide
Lager Image
Histograms of the FVI, NDVI, and RVI in the two study areas
PPT Slide
Lager Image
Results of the FVI, NDVI, and RVI in two study areas
We also compared the subset images of Study Area 1. In Fig. 5 , the area mainly covers paddy rice in the left portion and forests in the right portion. Thus, it fluctuates with the canopy structure, and varying amount of the biomass. In Fig. 5(b) , the NDVI fails to present information on the canopy structure and the types of vegetation. Also, the RVI could present the volume scattering difference but suffers from severe speckle noise as shown in Fig. 5(c) . In Fig. 5(a) , the FVI effectively represents the difference between the rice crop and the forest covers. Because the paddy rice is currently in its reproductive phase, the NDVI fails to separate these dense rice crops and forest areas. The density plot is shown in Fig. 5(d) using all pixels of the subset area. The density plot produces a two-dimensional plot of two variables, wherein the color is used to provide information on the frequency with respect to the two variables. With the NDVI, less than 0.7, the FVI and NDVI vary almost linearly. In an NDVI of more than 0.7, the NDVI fails to represent the difference between rice paddy and forest. However, in this area, the FVI dynamically changes because of the canopy disparity between the rice crops and the forests. In other words, the FVI not only provides information on the canopy structure, but also provides information of greenness and vigor.
PPT Slide
Lager Image
Subset imagery of study area 1. (a) FVI, (b) NDVI, (c) RVI, and (d) density plot of the FVI and the NDVI
Those different VIs provide different information on vegetation classes. Therefore, the separability of classes is further examined. Firstly, the separability of rice crops and forests in Study Area 1 is used to evaluate the sensitivity of different vegetation types. Secondly, in the POA shifts and the POA non-shifts urban areas with forest in the Study Area 2, the separability is analyzed to evaluate the aim of VI to enhance the vegetation class. Each class is made up of 2,000 pixels and picked up through the visual inspection. The ND (Normalized Distance) between two classes was used as a simple measurement of separability, which is defined as;
PPT Slide
Lager Image
whereas μ 1 and μ 2 are the mean values, σ 1 and σ 2 and are the standard deviations of each class. A high ND value indicates the pair class has high separability. The Fig. 6 shows a comparison of the calculated ND between the rice crop and forest pairs of Study Area 1, and between the forest and urban pairs of Study Area 2. In the rice crop and forest pairs, the FVI shows high separability due to the RVI contribution. The NDVI shows low separability, since it is insensitive to the dense vegetation condition and the vegetation structure parameter. In the POA shifted urban and forest pairs, the NDVI referred the high separability, but the separability of the FVI was similar to that of the NDVI. The RVI has low separability in the POA shifted urban and forest pairs, which caused unwanted false alarms. In other words, the RVI can represent urban areas as vegetated areas, while the FVI shows the highest value with the POA non-shifted urban and forest pairs. The visual analysis and ND values indicate that the proposed FVI provides more information on greenness, vigor, and other structural parameters of vegetation than only the NDVI and RVI.
PPT Slide
Lager Image
ND values (a) in study area 1 and (b) in study area 2
4. Conclusion
Those remote sensing VIs, which and are beneficial for the assessment of the biomass, vegetation water content, vegetation health, and crop production, are widely used. The successful applications of those VIs require information on the greenness, vigor, and canopy structure of vegetation. To address this issue, this paper presents a novel FVI. To characterize and measure the type, amount, and condition of the vegetation, an FVI was developed using the NDVI and the RVI. The experiment resulted in that an FVI is more advantageous than the NDVI and the RVI. To evaluate an FVI measurement of the vegetation types, we compared the ND values of various classes using two datasets: RapidEye and Radarsat-2. The proposed FVI was sensitive to the vegetation type and structure in the saturated NDVI areas, such as the forest and the rice crops. The main reason for this is that the geometric information of RVI was incorporated into the NDVI. Also, we found that the FVI remained unaffected by the POA shifts. However, the backscattered microwave is affected by the sensor configurations, such as the frequency, polarization, and incidence angle. Therefore, the further studies, using L- and X-bands data, will be addressed in our future work to improve the FVI, which is expected to provide the possible complementary information and contribute to the diverse vegetation monitoring.
Acknowledgements
This research was supported by an NRF (National Research Foundation) grant funded by the South Korean government (NRF-2014R1A1A1001995).
References
Gao B. (1996) NDWI-a normalized difference water index for remote sensing of vegetation liquid water from space Remote Sensing of Environment 58 (3) 257 - 266
Huete A. (1988) A soil-adjusted vegetation index (SAVI) Remote Sensing of Environment 25 (3) 295 - 309
Iribe K. , Sato M. (2007) Analysis of polarization orientation angle shifts by artificial structures IEEE Transactions on Geoscience and Remote Sensing 45 (11) 3417 - 3425
Jackson R. , Huete A. (1991) Interpreting vegetation indices Preventive Veterinary Medicine 11 (3) 185 - 200
Kim Y. , Van Zyl J. (2009) A time-series approach to estimate soil moisture using polarimetric radar data IEEE Transactions on Geoscience and Remote Sensing 47 (8) 2519 - 2527
Kim Y. , Jackson T. , Bindlish R. , Lee H. , Hong S. (2012) Radar vegetation index for estimating the vegetation water content of rice and soybean IEEE Geoscience and Remote Sensing Letters 9 (4) 564 - 568
Koppe W. , Gnyp M. , Hutt C. , Yao Y. , Miao Y. , Chen X. , Bareth G. (2013) Rice monitoring with multitemporal and dual-polarimetric TerraSAR-X data International Journal of Applied Earth Observation and Geoinformation 21 568 - 576
Kuenzer C. , Knauer K. (2013) Remote sensing of rice crop areas International Journal of Remote Sensing 34 (6) 2101 - 2139
Lee J.S. , Grunes M.R. , Mango S.A. (1991) Speckle reduction in multipolarization, multifrequency SAR imagery IEEE Transactions on Geoscience and Remote Sensing 29 (4) 535 - 544
Matthew M.W. , Adler-Golden S.M. , Berk A. , Richtsmeier S.C. , Levine R.Y. , Bernstein L.S. , Acharya P.K. , Anderson G.P. , Felde G.W. , Hoke M.P. , Ratkowski A. , Burke H.H. , Kaiser R.D. , Miller D.P. (2000) Status of atmospheric correction using a MODTRAN4-based algorithm SPIE Proceedings 4049 199 - 207
McColl K. , Entekhabi D. , Piles M. (2014) Uncertainy analysis of soil moisture and vegetation indices using aquarius scatterometer observations IEEE Transactions on Geoscience and Remote Sensing 52 (7) 4259 - 4271
Pettorelli N. , Vik J.O. , Mysterud A. , Gaillard J.M. , Tucker C.J. , Stenseth N.C. (2005) Using the satellite-derived NDVI to assess ecological responses to environmental change Trends in Ecology & Evolution 20 (9) 503 - 510
Shi J. , Jackson T. , Tao J. , Du J. , Bindlish R. , Lu L. , Chen K.S. (2008) Microwave vegetation indices for short vegetation covers from satellite passive microwave sensors AMSR-E Remote Sensing of Environment 112 (12) 4285 - 4300
Zhang L. , Furumi S. , Muramatsu K. , Fujiwara N. , Daigo M. , Zhang L. (2007) A new vegetation index based on the universal pattern International Journal of Remote Sensing 21 (1) 107 - 124
Zhang X. (2015) Reconstruction of a complete global time series of daily vegetation index trajectory from long-term AVHRR data Remote Sensing of Environment 156 457 - 472