Science Service System

Summary of Proposal LAN0101

TitleTesting usefulness of SAR for monitoring widespread insect defoliation in a pine forest
Investigator Solberg, Svein - Norwegian forest research institute, forest ecology
Team Members
Solberg, Anne - Norwegian Computing Centre,
SummaryThe objective of the project is to test the ability of TerraSAR x-band data for mapping of foliar mass in forest trees, as a means of forest health monitoring. We would like to have repeated surveys during the season in order to follow the development of the damage. We will develop models that relates X-band TerraSAR data to the ground truth data, being standard LAI measueements as well as up-scaling of these by air-borne laser scanning (LiDAR). In addition, we will do polarimetric radar response measurements on branch samples affected by the damage in a laboratory environment. A success of the project could be a breakthrough for the use of satellite data for a general forest health monitoring and mapping of forest damage.
Detailed reportTesting TerraSAR-X for forest disturbance mapping Report from TerraSAR-X science data application LAN_0101 Svein Solberg Dan Weydahl Rasmus Astrup Päivi Lyytikainen-Sarenmaa Harri Kartinen Introduction A widespread and severe insect attack was expected in the Palokangas forest area in Karelia, Finland in the summer 2008. The insect, a pine sawfly (Diprion pini) had had attacks in the area for a number of years already, and egg galleries in the pine needles were observed. A project was established with field inventory, airborne laser scanning and satellite SAR acquisition, including a number of TerraSAR-X images. The data were acquired both before and after the expected insect attack. The idea was to test the ability of SAR to detect the insect defoliation, - as a possible tool for forest health monitoring. Unfortunately, an extraordinary weather event ruined the insect attack: In the first week of June a heavy snowfall was seen in the area. This was the week of egg hatching, and most of the sensitive and tiny larvae were apparently killed. This report is an attempt to utilize the data in the best possible way, although the beautiful study we planned was severely affected. Rather than focussing on temporal changes, which were only minor during the summer, we here search for relationships between defoliation and SAR data in space, i.e. geographical patterns. As a representation of the amount of foliage in the forest we apply effective Leaf Area Index, LAIe. This is here defined as one half the total surface area of needles and also branches and tree stems, divided by the ground area above which these areas are found. LAIe is also somewhat correlated to other forest variables such as stem volume and biomass, at least when trees are healthy. The aim of this study is to evaluate the use of X-band SAR backscatter and coherence for mapping of LAIe. As it is well known that both backscatter intensity and coherence saturate at low levels of vegetation amount, it was a particular aim to test if these two SAR variables could be used to separate clear cuts from forested areas. Materials and methods Study area The test area was the Palokangas forest, Karelia, Finland. The forest area is intensively managed and harvested, and was clearly patchy, in the sense that the area could be divided in homogeneous stands with abrupt borders between, typically varying between clear-cuts, young stands and old stands. Digital terrain model (DTM) The DTM was geoid elevation derived from the national Finnish N60 database in the ETRS-TM35FIN projection. The elevation data were in dm, and was changed to m. The geoid elevations were changed to the WGS84 ellipsoid elevations by adding 16.512 m. This 16.512 m was the height difference between the geoid and the ellipsoid as measured at the location of the corner reflector, and it was found to be valid in entire Palokangas study area within a few centimeters. The projection was changed from ETRS-TM35FIN to UTM35N, as they are almost identical. Ground control point We out a trihedral radar reflector towards the orbital point for the planned TerraSAR-X acquisition. We measured the position of the reflector with differential GPS (dGPS). The position in UTM35 was northing 6 978 057.727 m, easting 699 199.095 m, and the elevation of the ground was 192.8 m above the WGS84 ellipsoid. Field inventory We laid out a set of 30 inventory plots where we carried out measurements of effective leaf area index (LAIe) with the LICOR LAI-2000 Plant Canopy Analyzer, and standard tree measurements from which we calculated stem volume and biomass. The position of the plots was measured with dGPS. Airborne laser scanning The entire area was covered by an airborne laser scanning (ALS) both at the beginning and the end of the summer. By combining the field inventory and the ALS data, we generated a wall-to-wall reference data set on LAIe to serve as a ground truth. The ALS based modelling of LAIe followed the method as described by Solberg et al. (2006), where ALS penetration rate is converted to LAIe following the Beer-Lambert law. Based on the reference data we carried out an automatic segmentation of the area into 609 polygons representing forest stands, after excluding peatland and lakes and rivers. This exclusion was done manually based on one land cover map and a Landsat satellite image. The LAIe varied considerably making this a valuable study area. Mean LAIe for the 609 forest stands varied from zero to 3.85, with a mean value of 1.07. SAR data The area was covered by a series of TerraSAR-X acquisitions in the summer. This included high-resolution spotlight (4), spotlight (10) and stripmap (3) modes (one stripmap scene was cancelled due to maintenance). This makes 3 time-series of acquisitions covering the same area and acquired from the same orbit position. The 10 spotlight images received main focus, and was a series of images taken regularly with 11 days interval. Here we only used the VV polarization. We tested the relationship between backscatter and LAIe and between coherence and LAIe. The backscatter intensity data were not speckle filtered in a conventional way, however, instead we carried out a temporal multi-looking of the series of 10 spotlight images. Coherence was derived for all images 11 days apart from each other, i.e. 9 pairs of spotlight images. Prior to the multi-temporal filtering and the coherence estimation the images were co-registered. This was done in 3 steps, first to the nearest pixel by using a central window over the images, then to one 10th of a pixel by a grid of smaller windows, and then finally to one 100th of a pixel with a final grid of 400 points. A threshold of minimum cross correlation was set to 0.2, and windows having a value below this was discarded. For the remaining windows a range and an azimuth shift were calculated, and an overall shift of an image was estimated. The backscatter intensity images were geo-coded using the DTM and the GCP, while the coherence images were geo-coded only with the DTM. Statistical analyses All reference data and SAR data were resampled to a 10x10 m spatial resolution. However, all statistical analyses were based on mean values for the forest stands: The forest stand polygon file was overlaid on both the reference data on LAIe, and also on the geocoded SAR backscatter and coherence data. Mean values of the 609 stands were then extracted, and statistical analyses carried out on this file. Results Backscatter intensity The VV backscatter intensity remained fairly stable with a mean digital number value around 110 throughout the summer. It tended to increase with increasing LAIe. This relationship was curvilinear with a saturation at a low LAIe level. The relationship was weak, with a Spearman rank order correlation coefficient up to 0.34, - however, mostly statistically significant. The relationship also varied from scene to scene, and in three cases it disappeared. The strongest relationship was found for spotlight image no 3. Coherence The coherence within the pairs varied from zero to 0.75, and there was also a temporal variation from pair to pair during the summer. For the forest stands the mean value varied from zero to 0.45. The relationship between LAIe and coherence was curvilinear and negative. At the forest stand level the Spearman rank order correlation was stronger than with backscatter intensity, with correlation coefficients up to 0.78. The correlation varied somewhat from pair to pair, or more specifically it disappeared for pair no 6. The relationship was strongest for high-coherence pairs. By inverting the coherence, the relationship with LAIe turned linear and linear regression analyses were carried out. The coefficient of determination (R2) varied from about zero in pair no 6 up to 0.58 in pair no 9. There is no saturation effect here, however, the residual variance increases with increasing LAIe. Temporal change There were some changes in LAIe during summer, mainly a slight decrease. However, some stands were clear cut, and their LAIe value decreased by 100%. It was clear that these stands had a corresponding strong increase in coherence. The effect of mode The results for the two other modes; high resolution spotlight and for stripmap, were similar and of the same magnitude as for spotlight. Discussion Both backscatter intensity and coherence were sensitive to small LAIe variations in low-LAIe forest stands. In general coherence was more suitable than backscatter intensity for forest monitoring. The relationship was stronger, and by inverting coherence there was no saturation. Treeless areas appeared as bright. This means that coherence could be used for detection of clear cuts, including monitoring of clear cuts. This might be a particularly valuable supplement to another use of SAR image pairs; interferometric height (InSAR height). It has been demonstrated in a number of studies that InSAR height is strongly related to stem volume, biomass and tree height. There might be a limitation with the use of InSAR height for forests with low biomass. Firstly, InSAR height is derived by subtracting two DEMs, i.e. the canopy surface model from an X- or C-band sensor, and the terrain model. Both these models will have a certain inaccuracy, and the inaccuracy will be larger for their difference. This means that w might get relatively large errors in low biomass stands when using InSAR height. In some cases InSAR height would even turn out with negative values, where regression models would predict negative stem volume, biomass or tree heights. To avoid problems in such cases, coherence might serve as a valuable supplement to clearly identify low biomass stands and clear cuts. The results here were equally good with stripmap, which means that the regular Tandem-X data acquisition for the global DEMs would provide suitable data for this. In addition, the Tandem-X would have considerably higher coherence than in this study, which means that the relationships between coherence and LAIe or variables such as biomass would be both stronger and more stable over time. A lack of coherence can be a result of three types of de-correlation; temporal, spatial and noise: In this study the temporal de-correlation is presumable the dominating type, as is generally found in repeat-pass interferomtery, and it depends largely on the volume or height of the forest. For the future Tandem-X data, the de-correlation would mainly be determined by the spatial de-correlation, while no temporal de-correlation would be present. The noise part might be generally low, partly because co-polarization will be used with high signal-to-noise ratios. Hence, in the regular Tandem-X acquisitions the spatial de-correlation will dominate, and it will depend largely on volume scattering which varies with variables such as LAIe, biomass, stem volume and tree height.

Back to list of proposals

© DLR 2004-2016