Stand Segmentation

Summary

CAFRI’s general approach to date for mapping and modeling has been primarily pixel-based. That is, we organize the lanscape into gridded units (usally 30m), summarize remotely sensed data (e.g. LiDAR point cluods) within these units, and push individual predictions for these units. This is mostly done to conform with others and align with existing data products like Landsat and its derivative products. However, operating at the pixel level offers another benefit – we can treat the landscape as a continous surface and don’t have to worry about discrete boundaries or edges. It’s a lot simpler this way.

Despite the convenience of the pixel, we know they aren’t all that meaningful ecologically, and they don’t reflect the way we think about the landscape in patches or stands for forests in particular. Grouping pixels also coarsens’s the resolution, offering fewer units to operate on, which may save us computation time and storage space (who knows though, polygon datasets are tricky…). In particular, as we look to extend our initial predictions forward in time under different scenarios pixels don’t make a ton of sense since existing simulators like United States Forest Service FVS operate on stands as individual units.

So the goal of this project is to create spatial objects from our pixel-based timeseries rasters, where each object contains pixels with similar AGB trajectories. This process is commonly referred to as segmentation in most cases is performed on optical imagery. We are currently in an exploratory phase, where we are trying to identify the best:

  • open source algorithms in python or R
  • data to feed into segmentation algorithms
  • reference data to tune or evaluate said algorithms and datasets

Finding a best approach from the many options will be one large portion of this project. Another question we hope to tackle is how objects and segmentation hyperparameters might vary over different landscapes, though we are starting in working forests as we have the best reference data here.

People

Some papers

  • Clinton, Nicholas, et al. “Accuracy assessment measures for object-based image segmentation goodness.” Photogramm. Eng. Remote Sens 76.3 (2010): 289-299. doi: 10.14358/PERS.76.3.289
  • Nowosad, Jakub, and Tomasz F. Stepinski. “Extended SLIC superpixels algorithm for applications to non-imagery geospatial rasters.” International Journal of Applied Earth Observation and Geoinformation 112 (2022): 102935. doi: 10.1016/j.jag.2022.102935
  • Costa, Hugo, Giles M. Foody, and Doreen S. Boyd. “Supervised methods of image segmentation accuracy assessment in land cover mapping.” Remote Sensing of Environment 205 (2018): 338-351. doi: 10.1016/j.rse.2017.11.024
  • Gómez, Cristina, Joanne C. White, and Michael A. Wulder. “Characterizing the state and processes of change in a dynamic forest environment using hierarchical spatio-temporal segmentation.” Remote Sensing of Environment 115.7 (2011): 1665-1679. doi: 10.1016/j.rse.2011.02.025
  • Costa, Wanderson Santos, et al. “Spatio-temporal segmentation applied to optical remote sensing image time series.” IEEE Geoscience and Remote Sensing Letters 15.8 (2018): 1299-1303. doi: 10.1109/LGRS.2018.2831914