A new approach automatically produces a hierarchical set of image segmentations for detailed analysis of forest data.
This lidar data fusion approach is based on associating samples from sparse lidar data with groups of region objects determined by a unique image segmentation approach, HSeg (Hierar - chical Segmentation). This segmentation approach, which was previously developed by co-innovator James Tilton, is ideal for this application because HSeg automatically produces a hierarchical set of image segmentations, i.e., a set of several image segmentations of the same image at different levels of detail in which the segmentations at coarser levels of detail can be produced from simple merges of regions at finer levels of detail. This enables a simple approach for selecting an appropriate level of segmentation detail. HSeg also automatically classifies the spatially continuous region objects into region classes, through a tight intertwining of region growing segmentation, which produces spatially connected region objects, with non-adjacent region object aggregation, which groups sets of region objects together into region classes. No other practical, operational image segmentation approach has this tight integration of region growing, object finding with non-adjacent region aggregation. HSeg produces image segmentations with high spatial fidelity — enabled by the tight intertwining of region growing segmentation with non-adjacent region object aggregation.
Also, Hseg controls the importance of spatially adjacent region merging relative to spatially non-adjacent region merging (aggregation) through the Swght parameter, which can vary from 0.0 to 1.0. At Swght = 0.0 no non-adjacent region object aggregation is performed, and with Swght = 1.0, equal weighting is given to spatially adjacent and spatially non-adjacent region merging.
The initial tests were performed using Landsat TM data transformed into brightness, greenness, and wetness tasseled cap features in an area where there is “wall-to-wall” lidar data from the Laser Vegetation Imaging Sensor, LVIS. This Landsat TM data was collected on Sept. 5, 2007 from over central Maine, and the LVIS data was collected in August 2009. The spaceborne lidar (SSL) data was simulated by selecting tracks out of the LVIS data at an appropriate density. The LVIS data served as the ground reference data for the evaluation of the results.
Later, additional tests were performed using UAV SAR data with three polarizations: HH, HV, and VV. The original SAR data was at 5-meter pixel resolution, but 6×6 blocks of pixels were averaged of this data to produce 30- meter pixel resolution SAR data to better compare the results with the tests with the 30-meter transformed Landsat TM data.
This work was done by James Tilton, Bruce Cook, and Paul Montesano for Goddard Space Flight Center. GSC-16535-1