Concepts & Methods

The technical tasks we are focused on in forest mapping involve delineating and classifying the conditions within forest stands, which are considered as two distinct steps.

The term “stand” is generally used to refer a forested area ranging in size from a few acres to more than one hundred acres that can be distinguished from adjacent forested areas based on species composition, tree size, density, spatial arrangement, productivity, and/or management history and accessibility.

Generalizing the Stand

In the context of object detection and delineation, the concept of a “stand” can also be generalized to refer to any region of interest that can be distinguished from adjacent regions using available input data. Any objects on the landscape that that can be represented with polygons/masks can be the “targets” of segmentation workflows like the ones we adopt. Examples include identifying areas with forest composition or structure that provides suitable habitat for a particular wildlife species, or areas that have been affected by natural of human disturbances with varying levels of intensity such as harvests, pathogens, or fire.

Pixels vs. Stands

There has been a significant amount of research and product development related to generating pixel-scale predictions of forest attributes. Our project departs from this approach by recognizing the importance of aggregating the landscape into practical management units which still remain the basis for most forest conservation and management planning purposes.

The fundamental premise of the stand mapping aspect of this project is that human-drawn stand boundaries provide a good place to start for teaching machines to recognize and attempt to replicate how human managers delineate forest conditions for practical uses.

Stand Delineation Approach

We are currently working on two main segmentation approaches:

• Multi-stage segmentation pipelines linking together image preprocessing, over-segmentation, region-merging, and boundary post-processing using tools primarily drawn from from scikit-image
• Instance segmentation using a region-based convolutional neural network, Mask-RCNN, adapted from the PyTorch implementation

A significant part of this effort involves the construction of a benchmarking dataset that includes several layers of features as well as the targets which include bounding boxes and masks distinguishing major land cover types (water, field, forest, barren, impervious) and the distinct instances of each cover type.