Segmentation is a common Data Science problem that involves locating objects or boundaries in images. In addition to being useful on its own, good segmentation is often important for classification of objects within images.

Summary: Segmentation

Inputs: Hyperspectral images, RGB images, and LiDAR point cloud data.
Output: Shapefiles containing estimates of the position and shape of individual tree crowns (ITC data).
Training Files: ITC_OSBS_plot.shp, OSBS_plot_hyper.tiff, hyper_bands.csv, OSBS_plot_chm.tiff, ptcloud_plot.csv, rgb_plot.tiff.
Test Files: hyper_plot.tiff, chm_plot.tiff, ptcloud_plot.csv, rgb_plot.tiff
Submission File: itc_subm_plot.shp (including all 4 sub files; .shp, .shx, .dbf, .prj)

3.1 Segmentation in the ecology domain

Identifying the position and size of individual trees from remote sensing is useful for understanding forest structure and an important first step in species classification. It is also a complex version of the common image segmentation task because trees often overlap each other and look similar, and because the available data is heterogeneous, involving many bands, multiple resolutions, and point cloud height data. Participants will use Hyperspectral, LiDAR and high resolution RGB photographs to locate and segment individual crowns in multiple plots.

3.2 Training Data

Input for the training phase will be data for a subset of plots at Ordway Swisher and will include shapefiles with polygons for each individual tree crown (ITC) along with associated hyperspectral, LiDAR, and RGB data. LiDAR data provides information on the spatial variation in canopy height that may allow partitioning of crowns of neighboring trees with similar spectral signatures. Hyperspectral data allows development of spectral signatures to identify object categories (e.g. by assigning spectrally-similar categories to the same cluster). RGB photographs provide finer resolution information (0.25x0.25 m as opposed to 1.0x1.0 m for the hyperspectral data), which may be helpful to separate trees that are close to one another and to refine boundary placement. Using the remote sensing data, participants will build models to estimate crown boundaries in plots where ITC polygons are not available. The ITC data can be considered as ground truth, since polygons of individual crowns were manually delineated in the field by field experts and drawn directly on remote sensing images. The input data consists of two types of data: a shape file containing ITC data for 70% of the plots and remote sensing data for all plots, including one of the following for each plot: Hyperspectral (geotiff format), LIDAR (geotiff and csv), and RGB (geotiff). File naming will be of the form itc_train_plot.shp, hyper_plot.tiff, chm_plot.tiff, ptcloud_plot.csv, rgb_plot.tiff where plot indications the numeric code of the plot of each plot.

3.3 Test Data

Data provided for testing will be the hyperspectral, LiDAR, and RGB data described above for the plots not provided in the training dataset. ITC data will not be provided for the test set, as the goal is for participants to predict its contents from the remote sensing data.

3.4 Submission Data

The submission files will include one shape file (including all 4 subfiles) for each test plot named as itc_subm_plot.shp.

3.5 Performance Metrics

Participants’ performance on each plot will be calculated as the mean pairwise Jaccard Coefficient, J(A,B), calculated for each predicted polygon with the corresponding ground truth polygon. When there is ambiguity with regard to which submitted polygon corresponds to which ITC, we will choose the mapping that gives the best score to the participant.

Equation describing Jaccard Coefficient

The score for this task is the average of the plot level scores, which are themselves the average scores of the polygons within each plot. This method is simple, does not require assignment of predicted crowns to specific ITCs by the participants, and has good continuity since the Jaccard coefficient has good continuity. The output for the crown delineation algorithm is expected to produce polygons with no overlap, but if there are cases of overlapping crowns in the submitted polygons, they will be disambiguated by disregarding the intersection in the numerator of the Jaccard Coefficient. Thus, it is in the participant’s best interest to remove areas of intersection by dividing those areas among the overlapping crowns.