The paper introduces a novel approach to utilizing aerial imagery for Earth science and natural resource management, specifically targeting the limitations of traditional methods that rely on synthesized top-down “orthomosaic” images. These conventional methods often lack vertical information and may include processing artifacts, which can hinder accurate predictions, such as tree species classification.
The proposed method generates predictions directly from raw aerial images and accurately maps these predictions into geospatial coordinates using semantic meshes. This approach allows analysts to utilize the highest quality data, capture information about the vertical aspects of objects, and leverage multiple viewpoints of each location for enhanced robustness. The method is released as a user-friendly open-source toolkit, making it accessible for broader use.
The effectiveness of this multiview approach is demonstrated on a new benchmark dataset comprising four forest sites in the western United States. This dataset includes drone images, photogrammetry results, predicted tree locations, and species classification data derived from manual surveys. The results show a significant improvement in classification accuracy, from 53% to 75%, compared to the orthomosaic baseline on a challenging cross-site tree species classification task.
This advancement highlights the value of using raw images and multiple perspectives for more accurate and robust predictions in aerial imagery analysis. By addressing the limitations of orthomosaic images and providing a comprehensive toolkit, this method offers a significant contribution to the fields of Earth science and natural resource management. The open-source nature of the toolkit ensures that it can be widely adopted and further developed by the research community, fostering continued innovation and improvement in aerial imagery analysis.
