Geospatial tree species prediction using multi-view drone imagery and computer vision

Amritha Pallavoor , David Russell , Derek Jon Nies Young

Imagery from uncrewed aerial vehicles (“drones”) is an increasingly popular modality for understanding forests at a large scale due to its relatively low cost and ability to be collected on demand. Computational tools to identify the location of individual trees and their species can inform forest management efforts, such as prioritizing regions to thin to reduce wildfire risk. Previous work on species prediction has relied on imagery with a single top-down view produced from an orthomosaic, which contains image artifacts and limited information about the side views of the tree. In this work, we used the high resolution raw drone imagery to train machine learning models, and leverage its highly overlapping nature to capture multiple views of each tree for species prediction. We precisely geolocated individual drone images and estimated 3D geospatial models of each stand using photogrammetry. These photogrammetry products were used to derive Canopy Height Models (CHM) to detect tree crowns and identify individual treetops through a geometry-based local maxima algorithm. Then, we spatially aligned the drone detected trees to the field data, matched trees between the modalities, and assigned field-observed species to the detected trees. Using Geograypher, we cropped each individual tree crown from every raw image it appears in and used this dataset to train an EfficientNetV2 species prediction model. For unseen data, we detected trees, generated species predictions independently across image views, and assigned each tree the most frequently predicted species class. In this workflow, we used data collected from 107 field plots comprising 9,238 matched trees. The multi-view drone image prediction workflow achieved a macro-averaged recall of 79% compared to the baseline of 63% for the orthomosaic approach. Our results demonstrate that leveraging the inherent multi-view nature of raw drone images and aggregating predictions at the individual tree level improves species classification accuracy across a geospatially diverse set of conifer forests in the western U.S. This workflow enables scalable tree-level species mapping and can support downstream applications in forest inventory and management.