Satellite detection of canopy-scale tree mortality and survival from California wildfires with spatio-temporal deep learning

Abstract

Mapping forest disturbances is paramount to carbon monitoring, estimating environmental drivers, and developing strategies to enhance forest resilience. Existing forest change products from Landsat and Sentinel-2 have improved our understanding of large-scale disturbance patterns; however, their relatively coarse spatial resolution (10 to 30-m) leads to the use of mixed pixels which constrains their application for detecting heterogeneous survival or mortality outcomes occurring at the level of individual trees or canopies. PlanetScope multispectral imagery at 3-m and near-daily frequency offers new capabilities to detect and monitor diverse tree mortality patterns following disturbance across landscapes. This research proposes a framework to detect canopy-scale (3 × 3-m) tree/shrub mortality and survival using PlanetScope monthly time series. A 3D Spatio-Temporal Convolutional Neural Network (ST-CNN) deep learning model was designed to fully utilize the spatial context and the temporal change unique to canopy survival and mortality from the PlanetScope time series. As a crucial component for training robust and scalable deep learning models, a large set of labels was collected via a semi-automatic workflow by combining pre-disturbance lidar crown segmentation and post-disturbance aerial imagery interpretation. We applied the framework to detect canopy-scale mortality and survival following 15 large wildfires in California from 2018–2021. We sampled 1,176 384 × 384-m scenes from burned areas with pre-fire aerial lidar, containing >1.8M tree and shrub canopy polygons labeled as dead or alive following wildfire. Evaluated with an independent testing dataset, the optimized ST-CNN model detects heterogeneous patterns representing survival and mortality outcomes at 3-m resolution which accurately align with observed/labeled data. Tree mortality detection accuracy was high and stable in the Sierra Nevada and North Coast Mountains ecoregions (user’s = 83%–86%; producer’s = 81%–82%), but decreased slightly within the sparser Central Foothills and South Coast and Mountains (user’s = 77%–81%; producer’s = 58%–61%) often due to confusion between shrub and tree mortality. Producer’s accuracy of tree mortality and survival increased with canopy height and remained stable (>75%) on canopies taller than 11-m. Further, a sensitivity analysis demonstrates the performance benefits of using spatial and/or temporal convolutions in the ST-CNN architecture for model prediction. Lastly, we demonstrate the scalability of the ST-CNN for regional-scale application on all large 2020 wildfires in California (∼1.6 Mha burn area). The wall-to-wall post-fire maps showed an overall 3-m tree mortality rate of 58.8%, ranging from 32% to 94% among individual fires. The trained ST-CNN provides an ecologically detailed estimation of the pre-disturbance forest composition (trees, shrubs, non-woody) and their outcomes (survival or mortality) post-disturbance. These data will improve higher resolution monitoring and assessment of forest disturbance impacts, allow for better understanding of forest vulnerability, and support forest management strategies and actions.