Recent technological advancements in uncrewed aerial vehicles (“UAVs” or “drones”), computer vision, and image processing have opened a new era in which forests can be mapped at the individual tree level across broad extents using imagery from low-cost consumer drones. The Open Forest Observatory (OFO), based at the University of California Davis, has developed a set of open-source software tools that make it easy for forest scientists and practitioners to employ advanced image and geospatial data processing algorithms to detect individual trees in drone imagery, map them geospatially, classify their health status and species using computer vision (assuming sufficient training data), and evaluate the accuracy of the predictions against co-located ground-based inventory data.

Existing computer vision models trained for tree detection and segmentation take a variety of approaches to handling the geospatial nature of drone-derived orthomosaics, making it difficult to switch between models and compare them. For example, large geospatial orthomosaics must be split into small "chips" for input to a computer vision model, and the resulting chip-level predictions must be reassembled, deduplicated in overlap areas, and reassigned geospatial reference information. The Tree Detection Framework provides this "geospatial boilerplate" functionality via a standardized interface in which the user specifies chip resolution, dimensions, and stride the same way regardless of which computer vision backend they use. It also implements a popular geometric algorithm for detecting treetops as local maxima in a canopy height model, via the same interface, enabling easy and rigorous intercomparison. It is simple to add support for additional models and algorithms.

When classifying tree species from drone imagery using computer vision models, the most common approach is to use the orthomosaic, which provides a single top-down view of each tree. This approach ignores the wealth of information in the raw drone images, which are highly overlapping and therefore provide numerous distinct views of each tree from different angles. However, drone images are not geospatial data products, and there is no direct way to translate the locations of tree crowns in raw drone images into precise geospatial polygons. Geograypher fills this gap. It leverages the fact that when you know the precise position and orientation of the drone camera, along with the camera's lens model, each pixel in a raw drone image can be translated into a 3D geospatial ray. Combined with the 3D mesh model derived from photogrammetry, these rays can be translated into precise geospatial points. Geograypher employs these concepts, drawing on computer graphics principles, to render geospatial information onto raw drone images and project information from those images into geospatial coordinates.

Using Tree Detection Framework, we tested a variety of computer vision models and a geometric algorithm at over 200 forested sites across California where we had drone imagery and co-located geospatial ground reference forest inventory data for validation. For each model or algorithm, we tested a range of parameterizations (e.g., image resolution). We found the optimal algorithm and parameterization depended on local stand conditions, suggesting that coarse preliminary knowledge of the conditions at a mapping site can substantially improve prediction accuracy. By lumping sites into coarse stand structure classes and selecting the optimal tree detection model for each site based on its structure class, we obtained median overstory tree detection recall and precision over 80%.

Using Geograypher, we produced a set of computer vision model training images from 86 plots (5,714 trees) that had drone imagery and co-located geospatial ground reference forest inventory data for validation. This represented a “multi-view” dataset, leveraging the highly overlapping drone imagery to obtain multiple views of each tree from different perspectives. When predicting to an independent dataset of 26 plots (1,896 trees) representative of the same domain, the model predicted tree species with 85% accuracy. At the same sites, a conventional orthomosaic-based approach (only one top-down view per tree) achieved accuracy of 73%. Additional experiments revealed that even for predictions in a single sub-region of our domain, a model trained on domain-wide data performed better than one trained on sub-region specific data. Similarly, when only coarse taxonomic classifications are required (e.g., pine, fir), greatest accuracy is achieved via post-inference lumping, rather than training a separate coarse-class model.

Overall, our software tools and data processing workflow represent an open, reproducible, rigorous, best-practice approach for forest mapping and species classification capable of achieving accuracies sufficient for practitioner deployment.