Self Improvement

Seasonal and topographic effects on growing stock volume estimates from JERS-1 backscatter in Siberian forests

Description
Seasonal and topographic effects on growing stock volume estimates from JERS-1 backscatter in Siberian forests
Published
of 12
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
Share
Transcript
    Seasonal dynamics and stem volume retrieval in boreal forests using JERS-1 backscatter Maurizio Santoro ∗ a , Jan Askne **  b , Leif Eriksson * a , Christiane Schmullius * a , Andreas Wiesmann *** c , Johan E. S. Fransson **** d   a  Friedrich-Schiller-University Jena, Institute of Geography, Department of Geoinformatics;  b  Department of Radio and Space Science, Chalmers University of Technology; c  GAMMA Remote Sensing Research and Consulting AG; d  Swedish University of Agricultural Sciences, Department of Forest Resource Management and Geomatics ABSTRACT The paper analyses seasonal effects on L-band backscatter in boreal forests and the implications for stem volume retrieval (JERS-1 mission). As test sites, the estate of Kättböle, Sweden, and two compartments in Bolshe-Murtinsky, Siberia, were considered. The in-situ measured stem volumes ranged from 5 to 350 m 3 /ha in Kättböle and to 400 m 3 /ha in Bolshe-Murtinsky, at stand level. For each site nine SAR images were available. Forest backscatter strongly depended on seasonal conditions. With respect to other seasons, in frozen conditions the dynamic range was smaller and the forest backscatter at least 3 dB lower. When precipitation occurred, the backscatter showed saturation. In Kättböle, no saturation was found in images acquired at dry/unfrozen conditions. By means of a semi-empirical model, a regression between stem volume and backscatter was performed. Stem volume was then retrieved for an independent set of backscatter measurements. Images acquired at dry/unfrozen conditions showed a relative RMS error of around 30 % for the images acquired over Kättböle. At both sites the retrieval error was higher for other weather conditions, around 50%. When dry/unfrozen conditions occurred, multi-temporal combination of stem volume estimates showed the smallest error (22%). Hence, for boreal forest monitoring L-band images acquired at dry/unfrozen conditions should  be used. Keywords : JERS-1 SAR, backscatter, stem volume, Water Cloud Model, retrieval, weather, penetration depth. 1. INTRODUCTION Boreal forests are one of the largest biomes on Earth and are located in remote and scarcely inhabited areas, to which the access is limited for a long period of the year. They have a crucial role both for the economy of several nations and the Earth climate, thus raising global mapping issues 1 . The need of periodical monitoring of large and remote areas makes spaceborne remote sensing more appealing than the traditional sources of information, i.e. local surveys. Spaceborne SAR can provide images rather frequently, at any weather condition and independently from solar illumination, thus making such sensor suitable for forest monitoring at northern latitudes. The use of spaceborne SAR imagery to retrieve biophysical properties of boreal forests has been widely investigated 2-10 . Considering only the bands available on space platforms (X-, C- and L-band), the backscatter coefficient of forests has a weak dependence on stem volume, generally showing seasonal changes and saturation. Saturation depends on wavelength, incidence angle and type of forest, already occurring at a level typical of rather sparse forests, especially ∗  maurizio.santoro@geogr.uni-jena.de; phone +49 (0)3641 948885; fax +49 (0)3641 948882; http://omsjena.geogr.uni-jena.de; Friedrich-Schiller-University Jena, Institute of Geography, Department of Geoinformatics, Remote Sensing Unit, Loebdergraben 32, 07743 Jena, Germany; **  askne@rss.chalmers.se; phone +46 (0)31 772 1843; fax: +46 (0)31 772 1884; http://www.rss.chalmers.se/WWW_rsg/; Department of Radio and Space Science, Chalmers University of Technology, SE-412 96 Göteborg, Sweden; ***  wiesmann@gamma-rs.ch; phone +41 (0)31 9517005; fax +41 (0)31 9517008; http://www.gamma-rs.ch/; GAMMA Remote Sensing Research and Consulting AG, Thunstrasse 130, CH-3074 Muri / Bern, Switzerland; ****  Johan.Fransson@resgeom.slu.se; phone +46 (0)90 7866554; fax: +46 (0)90 778116; http://www.resgeom.slu.se/default.cfm; Swedish University of Agricultural Sciences, Department of Forest Resource Management and Geomatics, Remote Sensing Laboratory, SE-901 83 Umeå, Sweden.   for higher frequency 2, 11 . At L-band, the effect of climate conditions at acquisition has been reported in 4-6, 8, 12 . In 6, 7, 13, 14  it has been shown that the backscatter in boreal forests saturates between 150 and 225 m 3 /ha, thus being more suitable than X- and C- band for stem volume retrieval. The availability of an extensive dataset of JERS images and ground-truth data from two areas located in the boreal belt (Scandinavia and Siberia) allowed to carry out a thorough analysis of the seasonal changes of L-band backscatter in  boreal forests and the implications for stem volume retrieval. This is performed using a semi-empirical model relating the backscatter to stem volume. In the first part of the paper, the relationship between seasonal dynamics and model  parameters is analysed. In the second part, the estimation of stem volume from backscatter is described. The model is trained on a set of measurements and then tested on validation datasets to retrieve stem volume. Several stem volume estimates are finally combined in a multi-temporal approach. 2. GROUND-TRUTH DATA As test sites a forest estate near Kättböle (59 °  59’ N, 17 °  07’ E) in Sweden and two forest compartments in the Siberian territory of Bolshe-Murtinsky (91 °  30’ N, 57 °  14’ E; 92 °  08’ N, 56 °  54’ E) were considered. Both test sites have been  previously investigated using ERS SAR interferometric data 9, 10, 15 . The forest estate of Kättböle covers 550 ha, with relatively flat topography. Boreal coniferous species such as Scots pine and Norway spruce dominate even though some  broad-leaf tree, such as birch, are also present. The two Siberian forest compartments are part of the Bolshe-Murtinsky territory, located north of Krasnoyarsk along the river Yenisey in central Siberia. This is one of the 13 territories that were used in the SIBERIA project (SAR Imaging for Boreal Ecology and Radar Interferometry Applications) 9 . According to the numbering introduced within the SIBERIA project, the two compartments will be referred as “Bolshe-2” and “Bolshe-4”. The major tree species are spruce, fir and birch with a percentage of pine, larch, cedar and aspen. The two compartments cover 262 km 2  and 246 km 2 respectively, with varying topography. The distance between the centres of the two compartments is around 42 km. An extensive dataset of forest parameters, reported at stand level, was available for this study. The ground-truth data available consisted of forest stand boundary maps in digital form and field measurements of several forest parameters, including stem volume and tree height. The digital forest masks were available at 12.5 by 12.5 m pixel size for Kättböle and at 50 by 50 m for Bolshe-Murtinsky. In Kättböle an inventory was conducted in large part during 1995 using plots systematically distributed over the entire estate. Some measurements of randomly distributed plots were performed in 1996 as a complement 15 . The derived stem volumes were empirically corrected to the year of image acquisition using the site index. In addition, the area-fill factor was measured in 18 stands in Kättböle. This parameter represents the fraction of ground covered by tree crowns from the radar’s perspective 16 . For both compartments in Siberia an extensive GIS-based forest database was available. The data srcinates from the regular forest surveys performed by the Russian foresters. For Bolshe-Murtinsky the database was updated in 1998. No correction for the year of acquisition of the images could be performed. Based on information from the Russian forest inventory standards, stem volume accuracy is between 15% and 20%, depending on the age of the forest 9 . However, exact figures for the two compartments were not available. To aid interpretation, weather statistics in form of temperature, precipitation, relative humidity, snow cover and depth, and wind speed were used. For Kättböle one to seven daily measurements were obtained from three weather stations operated by the Swedish Meteorological and Hydrological Institute (SMHI) located near the site. For Bolshe-Murtinsky daily averages from a weather station located at around 50 km from Bolshe-4 were available. 3. SATELLITE IMAGERY   Imagery acquired by the HH polarised L-band (23.5 cm) SAR mounted on the Japanese satellite JERS-1 at a nominal incidence angle of 39° was used. For both test sites we had available 9 images acquired between April 1997 and August 1998 over Kättböle and between 1994 and 1998 over Bolshe-Murtinsky (Table 1). All images were in Single Look Complex (SLC) format. Nevertheless, since this work was carried out at stand level, averaging on a statistically sufficient number of pixels had the effect of reducing the noise in the backscatter measurements.    Kättböle Bolshe-Murtinsky 15 April 1997 T ≈ 1°, SD: 9 cm Precipitation 6 January 1994 T ≈ -17°, SD: 39 cm, Snowfall 29 May 1997 T ≈ 7°, Little rainfall 19 February 1994 T ≈ -17°, SD: 45 cm, Snowfall 12 July 1997 T ≈ 18° 14 October 1996 T ≈ -4° 8 October 1997 T ≈ 12°, Rainfall 27 November 1996 T ≈ -21°, SD: 25 cm, Snowfall 4 January 1998 T ≈ 2°, SD: 5 cm Melting snow on ground, Precipitation 10 January 1997 T ≈ -23°, SD: 45 cm 17 February 1998 T ≈ -2°, SD: 25 cm, Precipitation 23 February 1997 T ≈ -16°, SD: 53 cm 2 April 1998 T ≈ 1°, SD: 0 cm, Precipitation before acquisition (when T>0°) 8 April 1997 T ≈ 10°, Probable snow melt 16 May 1998 T ≈ 19° 22 June 1998 T ≈ 22°, Rain before and during acquisition 12 August 1998 T ≈ 21° 5 August 1998 T ≈ 19°, Rain before and during acquisition Table 1. Acquisition dates and weather conditions at acquisition. Temperature (T) is reported in degrees Celsius. For Bolshe-Murtinsky a mean value for the day is reported. Snow depth is indicated as SD. 3.1 Image processing Processing and radiometric calibration from JERS-1 level 0 to SLC format was done by Gamma Remote Sensing 17  and  NASDA, respectively for the Bolshe-Murtinsky and the Kättböle images. The radiometric calibration accounted for JERS sensitivity gain control, automatic gain control, and corrected for the JERS range antenna pattern. In addition the data were filtered for radio frequency interference. In order to compare radiometric calibration, raw data from one acquisition over Kättböle (15 April 1997) was processed to SLC by Gamma. A scaling factor for the radiometry was determined for the Kättböle site. Geocoding of SAR images was performed in order to match with the digital forest masks. For Kättböle a digital elevation model (DEM) produced by the Swedish National Land Survey with a pixel size of 50 by 50 m was resampled to 12.5 by 12.5 m pixel size using linear interpolation. The DIAPASON software by CNES was used for geocoding. For Bolshe-Murtinsky a DEM produced from interferometric ERS data 9  was used for images acquired before 1998. Those acquired in 1998 were geocoded to the global DEM “gtopo30” because of the lack of a DEM at the time of processing 17 . Images acquired before 1998 were geocoded to a 25 by 25 m pixel size 18  and then downsampled to 50 by 50 m to match the forest mask. Those acquired in 1998 had been processed directly to a 50 by 50 m pixel size. 3.2 Forest backscatter computation Using the digital forest stand map, forest stands were localised in the geocoded SAR images. In both cases stands with stem volume below 5 m 3 /ha were not considered. In order to reduce border effects and localisation errors on the  backscatter measurements, the stands were decreased in size by removing a two pixels wide zone along the perimeter of each stand. For each stand, mean backscatter was computed. Radiometric correction for the topography was not considered to be relevant. The backscatter of stands with a marked topography (i.e. standard deviation of height greater than 5 m) did not show a clear difference when compared to other stands having similar stem volume. In Kättböle only stands larger than 2 ha after shrinking were considered (more than 128 pixels). In total, 42 stands were taken into account for analysis, with an area ranging between 2 and 14 ha. Stem volume was distributed between 16 and 344 m 3 /ha, with average volume of 144 m 3 /ha, standard deviation of 76 m 3 /ha and a standard error of about 18%. In Bolshe-Murtinsky only those stands that had more than 32 pixels after shrinking were considered (i.e. at least 8 ha). This requirement strongly reduced the amount of stands but it did not change significantly the distribution of stem volume, as shown in Table 2. Figure 1 reports a comparison between the distributions of stem volume in the three test areas. The two Siberian sites had a higher mean volume than Kättböle, being distributed around much higher volumes.    Figure 1. Distribution of stem volume in the three test sites. Bolshe-2 Bolshe-4 Database Used Database Used Stands 1147 147 496 161 Min [m 3 /ha] 5 5 5 15 Max [m 3 /ha] 470 400 470 400 Mean [m 3 /ha] 230.4 219.6 175.7 161.3 Std. Dev. [m 3 /ha] 106.7 115.6 104.3 116 Table 2. Statistics for the two Siberian test sites. “Database” refers to the data provided in the GIS-based database. “Used” refers to those stands used in the study (i.e. area > 8 ha and stem volume > 5 m 3 /ha). 4. METHODOLOGY L-band backscatter from forests depends not only on the forest structure (density of foliage, branches and trunks, moisture content of trees and ground, soil roughness) but it is also related to the SAR imaging system (i.e., incidence angle, spatial resolution, polarization). The wavelength of L-band (23.5 cm) is such that in forested areas the wave  partly penetrates through the upper layers of the canopy, interacts with the main branches and gives rise to double- bounces between trunk and ground. 4.1 Modelling For this analysis, we assumed that the tree-ground double bounce is less important than the direct scattering from ground and vegetation. Previous investigations have showed the possibility of modelling L-band forest backscatter, considering these two contributions only 6, 14, 19 . Although not fully correct, the assumption can be motivated considering two characteristics of boreal forests. The rough topography of the ground can significantly deflect the double-bounce away from the radar while the thick and dense canopy can strongly attenuate the incoming wave. Support to the assumptions could be found in penetration depth values reported in 20, 21  (4 to 10 m) and in 8, 22  where the effect of a rough ground on the double-bounce strength at L-band has been discussed. In a similar manner to the Water Cloud Model for vegetation 23  we consider a simple model based on radiative transfer through a horizontal scattering and attenuating layer with gaps 19 : ( ) ( ) treeoveg treeo gr o gr o for   T T   −++−= 11  σ σ η σ η σ  . (1) The forest backscatter, o for  σ  , is considered as the sum of direct scattering from the ground through the canopy gaps, ground scattering attenuated by the canopy and direct scattering from the forest canopy. In order to take into account the   gaps in the canopy, each term is weighted by the area-fill factor, η  . For a complete description of the model see 10, 19 . In (1) o gr  σ  and oveg  σ   represent the backscatter from the ground and the vegetation layers respectively, while T  tree is the two-way transmissivity through the tree canopy. This expresses how much the incoming energy gets attenuated when it  passes through the tree canopy. The two-way tree transmissivity factor can be written as h e  α  − , where α   is the two-way attenuation per meter through the tree canopy and h is the attenuating layer thickness, which we assume to be the same as the tree height. The two-way attenuation per meter can be converted into penetration depth, δ  , by using the following equation: 110 10ln2 −     ⋅= α  δ  . (2) Equation (1) can be rearranged in order to highlight the scattering components from the ground and the vegetation: ( )  for oveg  for o gr o for   T T   −+= 1 σ σ σ  , (3) where ( )  h for   eT   α  η η   − +−= 1  (4) represents the two-way transmissivity through the whole forest canopy. Compared to X- and C-band where the attenuation through the tree canopy, α  , is very high, the longer wavelength of L-band implies larger penetration of the waves in the forest canopy, so that the term h e  α  − should not be neglected. In 7  the forest transmissivity has been expressed as V  e  β  − , where V   is the stem volume and  β   is an empirically defined coefficient. Hence, the forest backscatter in (3) can be expressed as a function of stem volume and (4) rewritten as (6): ( ) V oveg V o gr o for   ee  β  β  σ σ σ   −− −+= 1  (5) ( ) hV  ee  α  β  η   −− −−= 11 . (6) In (5) there are three unknown parameters: the ground and vegetation backscatter coefficients (  o gr  σ  and oveg  σ   ) and the forest transmissivity coefficient,  β  . For each test site, the three parameters were estimated using backscatter measurements from either all the stands (for the seasonal dynamics analysis) or part of them, (for the retrieval). Once  β   has been estimated, corresponding values of α   can be estimated using (6) if area-fill factor measurements are available. 4.2 Model-based stem volume retrieval By inverting (5), it is possible to express stem volume as a function of backscatter:     −−−= o gr oveg o for oveg  V  σ σ σ σ  β  ln1 , (7) and therefore estimate stem volume having available a set of backscatter measurements o for  σ  , independent from those used for training the model. Since the total amount of stands was different in Kättböle and in Bolshe-Murtinsky, the training and the test sets were defined in different manners. Nevertheless, we tried to have as training set a set of values that could uniformly represent the range of stem volumes typical of the test site.
Search
Similar documents
View more...
Tags
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks