Satellite detection of crown-scale tree mortality and survival from California wildfires
Dan Dixon, University of California Davis, djdixon@ucdavis.edu (Presenter)
Yunzhe Zhu, University of California Davis, ypfzhu@ucdavis.edu
Yufang Jin, University of California, Davis, yujin@ucdavis.edu
This research proposes a general framework to detect crown-scale tree mortality and survival with PlanetScope monthly time series inputs (pre and post forest disturbance) at 3-m spatial resolution. Included in the framework is 1) a Spatio-Temporal Convolutional Neural Network (ST-CNN) designed to extract spatio-spectral-temporal features unique to tree crown mortality/survival from PlanetScope time series and 2) a semi-automatic crown label generation workflow which uses pre-fire lidar (automatic) and post-fire aerial imagery (manual) to build a wall-to-wall database of tree-level mortality and survival outcomes. We applied the framework to detect tree mortality and survival following 15 wildfires in California from 2018-2021 using a database of 1,000 128 x 128 pixel scenes (3 x 3-m) containing tree and shrub crowns labeled as survival or mortality. Applied to an independent test dataset, the accuracy assessment suggests performance was high and stable in the Sierra Nevada and North Coast and Mountains ecoregions on tree survival (user's acc. = 76-79%; producer's acc. = 90-92%) and tree mortality (user's = 82-84%; prod's = 80-82%) and moderate performance in the Central Foothills and South Coast and Mountains (Tree survival: user's = 72-82%; prod's = 83-85%; Tree mortality: user's = 67-87%, prod's = 58-66%). An accuracy assessment by tree height shows performance was mostly stable when predicting on trees taller than 10-m, yet reduced performance between 5-m to 10-m. A second check on observed and predicted data resampled to 32-m found minimal bias for both tree mortality and survival prediction. Lastly, we demonstrate the scalability of the ST-CNN for state-wide application on all large 2020 wildfires in California (n=42; ~1.6 M ha burn area) to report crown-scale tree mortality and survival outcomes. These higher resolution crown-scale data are urgently needed to monitor, rehabilitate, and enhance forest resiliency alongside shifting disturbance regimes including wildfire.
Poster: Poster_Dixon_3-41_64_35.pptx
Associated Project(s):
Poster Location ID: 3-41
Presentation Type: Poster
Session: Poster Session 3
Session Date: Thu (May 11) 3:00-5:00 PM
CCE Program: LCLUC