Estimation of Forest Variables using Radargrammetry on TerraSAR-X Data in Combination with a High Resolution DEM
Persson, Henrik1; Fransson, Johan E.S.1; Santoro, Maurizio2; Wegmueller, Urs2
1Swedish university of agricultural sciences, SWEDEN; 2Gamma Remote Sensing, SWITZERLAND
In forest management it is of great interest to have accurate information about forest variables in order to make reliable decisions for short and long term planning of silviculture treatments. In Sweden, about 23 million ha are covered by managed forest, representing a major natural resource. To cover large areas, satellite remote sensing can play a vital role in mapping the state of the forest in an efficient way. Although a lot of research studies have been performed in forest remote sensing using satellite data, there is still a need for new techniques and methods to be developed to improve existing estimates of forest parameters. Radargrammetry is a fairly known technique, but its potential for estimating forest variables has only been explored to a limited extent. Lately, new satellite Synthetic Aperture Radar (SAR) sensors like TerraSAR-X and COSMO-SkyMed in combination with access to high resolution Digital Elevation Models (DEMs) have opened the possibility to achieve promising results of estimating forest height and biomass with radargrammetry (Perko et. al., Remote Sensing 2011; Vastaranta et. al., IGARSS 2012). For tree height (Vastaranta et al. 2012) showed a relative RMSE of 12.2% with radargrammetry while the corresponding figure was 8.1% using ALS data, in a similar study. Radargrammetry has the advantage that a full time series of images can be exploited to estimate forest parameters, thus allowing substantial improvement with respect to one or few estimates.
Objective The objective of this study was to investigate the possibilities of estimating tree height and forest biomass, at stand level, using repeat-pass X-band radargrammetric SAR 3D data from the TerraSAR-X mission together with an Airborne Laser Scanning (ALS) DEM and in situ data. The short repeat-pass cycle of the TerraSAR-X and Tandem-X missions (11 days) contribute to that a large number of images per unit area over a limited time period can be available, which increase the potential in developing an efficient monitoring tool.
Materials The study is carried out at two Swedish test sites, Krycklan and Remningstorp, located in northern Sweden (64°07° N, 19.55° E) and southern Sweden (58°30' N, 13°40' E), respectively, where extensive remotely sensed and in situ data sets have been collected through the years. The test sites are managed for timber production. Krycklan has a more hilly topography while Remningstorp has relatively flat terrain. The forests are mainly dominated by Norway Spruce (Picea abies), Scots Pine (Pinus sylvestris) and Birch species (Betula spp.).
At the test sites, high-resolution TerraSAR-X and ALS 3D data, and in situ data at both plot level (10 m and 40 m radius) and stand level are available. Through processing of the TerraSAR-X data a Digital Surface Model (DSM) of the forest canopy was derived. By combining the DSM and an ALS DEM, the forest canopy height above ground was calculated to produce a Canopy Height Model (CHM). From the point cloud, canopy height (e.g., height percentiles) and density metrics (e.g., proportion of points in the vegetation) were derived at plot level and used in the model development.
Methods When choosing appropriate SAR images to work with, different incidence angles and combinations of time spans were considered and tested for the processing. Different filters to improve the matching requirements were tested and in agreement with (Perko et. al. 2011), a classical Lee-filter turned out to perform well. The filtered images were matched with Fast Fourier Transform (FFT). Different post-processing schemes (combinations of median filtering, interpolation window sizes, histogram matching and Signal-To-Noise thresholds) are currently examined to further improve the final CHM. Averaging of the full time series of images is also to be evaluated.
Different metrics (height, standard deviation, height percentiles) were then extracted from the CHM and models relating the addressed forest variables to the extracted metrics were fitted using linear regression analysis, where the independent variables (metrics) were selected based on regression model fit statistics and studies of residual plots.
The evaluation at stand level was done in two ways. The first validation was made with ALS comparisons and for the second validation, field plots were used and compared. The results were evaluated at stand level in terms of Root Mean Square Error (RMSE) and bias (in percentage of the surveyed mean).
Expected results The estimation accuracy is expected to be comparable to traditional forest inventories based on aerial photo-interpretation or subjective field surveys. It is also anticipated that the accuracy will be similar to those obtained with methods based on photogrammetric matching of digital aerial images for tree height and stem volume estimation
Discussion The outcome of this study is important, since it will contribute to provide results to evaluate the performance of the X-band radargrammetry concept for high resolution sensors. If successful, the proposed method has the potential to retrieve forest tree height and biomass with high-resolution and accuracy from space in areas where a high-resolution DEM of the ground topography is available. It is, therefore, expected to be a cost-effective way of mapping and updating forest variables.
Conclusion The obtained results are foreseen to be relevant for updating outdated inventory stand registers in support of forest management planning, implying that 3D data from TerraSAR-X have potential for operational use in forestry.