Explore the temporal evolution of the burn site through interactive 3D reconstructions captured at monthly intervals. Click on any model below to load and interact with the 3D scene. Please use a device with a GPU and ensure that your browser has access to the GPU. Otherwise, the models and browser may be very slow. Each model is roughly 1GB in size. The 3D models of each month are available upon request.
Land management and stewardship teams continue to lack the tools to capture 3D spatiotemporal insights of the ecosystems they oversee. For wildfire management at the wildland-urban interface, teams face challenges in capturing vegetation growth over time after a fuel reduction program and connecting seasonal changes to the vegetation distribution across the treated area. Current approaches rely on triangle meshes or point clouds generated from photogrammetry or LiDAR surveys on drones or hiked traverses. However, the difficulties in optimizing these meshes lead to large triangles that inadequately approximate the bulk vegetation shape, and the point cloud data is often too sparse for local plant-scale understanding.
To address this gap, we extend recent machine learning-based computer graphics techniques in 3D Gaussian Splatting (3DGS) to reconstruct scenes at the wildland-urban interface from handheld imagery (i.e., a typical tablet or smartphone camera) with sufficient detail to identify species, capture individual leaves, recover plant stature, and disambiguate overhanging plant individuals. We develop a new method to match the 3DGS reconstruction of these scenes across months, associating plant growth across seasons interactively in 3D. To achieve the centimeter-level matching, we adapt the Umeyama algorithm and the iterative closest point algorithm from point cloud maps to the environment-specific 3DGS scene, leveraging the probabilistic interpretation of the 3D Gaussian data structure and robustly handling visual and geometric changes associated with vegetation phenology over time.
We have applied our method to recent pile burns at Stanford's Jasper Ridge 'Ootchamin 'Ooyakma Biological Preserve at monthly intervals. We demonstrate differences in ecological response where some piles featured the unexpected return of a rare and threatened bushmallow, and others remained more barren. This pile burn microcosm implicates the need for plant-level 3D spatiotemporal models to understand ecosystem recovery to fire mitigation practices.
Session B11A - Advancements in Wildland Fire Science, Management, and Engagement: Integrating Earth Observation Technologies and Collaborative Development I
Oral Presentation
@inproceedings{Neamati_2025,
title = {Tracking Recovery: Temporally-Matched 3D Gaussian Splatting of Ecosystems after Prescribed Burns},
author = {Daniel Neamati and Grace Gao},
year = {2025},
booktitle = {American Geophysical Union (AGU) Annual Meeting},
address = {New Orleans, LA},
month = {December},
publisher = {AGU},
doi = {10.22541/essoar.176918805.57741912/v1}
}This work was supported by the TomKat Center for Sustainable Energy Graduate Fellowship for translational research. The work was conducted at the Jasper Ridge 'Ootchamin 'Ooyakma Biological Preserve with guidance and support from the staff, especially Dr. Adrianna Hernández, Dr. Sheena Sidhu, and Trevor Hebert. We would also like to thank the students who helped with the data collection: Evan Twarog, Lyla Dong, Phil Roberge, Bek Malikov, Madison Hunt, Jerry Wang, Jack Goler, Athena Kolli, and Rudy Mohapatra. We would like to thank the rest of the NAV Lab for their helpful discussions and feedback.