An evaluation of the JPL TOPSAR for extracting tree heights

In this paper, the accuracy of the digital elevation model (DEM) generated by the Jet Propulsion Laboratory (JPL) TOPSAR for extracting canopy height is evaluated. For this purpose, an experiment using C-band TOPSAR at the Michigan Forest Test Site (MFTS) in Michigan’s Upper Peninsula was conducted. Nearly 25 forest stands were chosen in MFTS, which included a variety of tree types, tree heights, and densities. For these stands, extensive ground data were also collected. The most important and difficult-to-characterize ground truth parameter was the forest ground level data, which is required for extracting the height of the scattering phase center from the interferometric SAR(INSAR) DEM. To accomplish this, differential global positioning system (GPS) measurements were done to accurately ( 5 cm) characterize the elevation of: 1) a grid of points over the forest floor of each stand and 2) numerous ground control points (GCPs) over unvegetated areas. Significant discrepancies between GPS and TOPSAR DEM and between the two TOPSAR DEMs of the same area were observed. The discrepancies are attributed to uncompensated aircraft roll and multipath. An algorithm is developed to remove the residual errors in roll angle using elevation data from: 1) 100-m resolution U.S. Geological Survey DEM and 2) the GPS-measured GCPs. With this algorithm, the uncertainties are reduced to within 3 m. Still, comparison between the corrected TOPSAR DEMs shows an average periodic height discrepancy along the cross-track direction of about 5 m. Simulation results show that this might have been caused by multipath from an object near one of the INSAR antennas. Careful examination of the coherence image and the backscatter image also show such periodic patterns. Recommendations are provided for the extraction of the best estimate of the scattering phase center height, and a model is provided to estimate actual tree height. It is accurate to within 1 m or 10% for the red pine test stands used here.