Link zu WSL Hauptseite Swiss Federal Institute for Forest, Snow and Landscape Research WSL
 
Duration: 2011 -

Shape matching in (non-rigid) natural environments

Shape matching is an important principle in image processing to convert unordered 2D or 3D data into contiguous data structures. In rigid or man-made environments shape matching can be defined as correspondence problem between reference and unknown shapes. The matching error corre-sponds to the summed up distance between corresponding points along the matched shapes (Belongie, Malik 2002).

Many remote sensing datasets like digital surface models (DSM) or point clouds from terrestrial laser scanning contain high resolution data with a regular spatial distribution, but lack any feature infor-mation. For applications in forest, snow and landscape research, high level features like shrub, single trees, tree crowns or forest stand are requested. Contrary to many applications for feature extraction of rigid objects like roof structures, buildings or roads, matching algorithms for vegetation structures cannot depend on robust edge information. Often are edges not even obvious for the professional visual observer, e.g. the delineation of tree crowns by aerial image interpretation.

CHM from ADS80 and LiDAR CIR ADS80 Matched tree crown candidates
Figure 1: CHM from ADS80 and LiDAR, 1m GSD (Click to enlarge image)
Figure 2: CIR ADS80, 25cm GSD (Click to enlarge image) Figure 3: Matched tree crown candidates (circles) with CIR ADS80. Blue region contains buildings with low NDVI (Click to enlarge image)

Finding tree crowns allows to model forest stands without precise knowledge about the true position of single tree stems. The pixel-wise difference in a canopy height model (CHM) does not yield robust estimates of tree positions and the often used curvature function (fourth-order polynomial within a 3x3 window) yields too many artifacts in urban areas and scattered forest stands. Therefore a specific shape matching approach for tree crowns has been developed, which tries to overcome the limited robustness of functional models. In a customized window, slope values from the center along 8 directions will be evaluated (yellow pattern in Figure 1). If the spatial evaluation identifies a tree crown candidate, the weighted spectral histogram verifies that hue, saturation and lightness are within feasible limits (yellow pattern in Figure 2).

Grouped tree crowns Enlarged view
Figure 4: Grouped tree crowns between 2008 and 2011 (Click to enlarge image) Figure 5: Enlarged view of Figure 4 (Click to enlarge image)

Using a DSM with 1m ground resolution and color-infrared images (CIR ADS80) with 0.25cm allows to extract tree crown candidates (Figure 3). Height values below 3m are ignored (typically shrub) and values with low NDVI are reclassified as artificial objects like buildings (e.g. blue region in Figure 3).

Grouping all single tree crowns allows to automatically derive a forest mask covering whole Switzer-land (Figure 4). The red part is shown in detail with Figure 5. The red delineation in Figure 5 sur-rounds grouped tree crowns found with the presented shape matching algorithm.

Reference

  • Belongie S., Malik J. & Puzicha J., 2002. Shape Matching and Object Recognition Using Shape Contexts, IEEE Transactions On Pattern Analysis And Machine Intelligence, Vol. 24, APRIL 2002: 509-522

Publication

Contact

Keywords tree crown, forest stand, shape matching, object detection, aerial image

 

Spacer