Widespread tree mortality events occur in temperate forests during periods of severe drought. Severe droughts like the 2012-2016 California drought have become more frequent over the last several decades, threatening forests no longer aligned with their current climate conditions. To better predict what will happen to these forests in the future, we need an improved understanding of the response of forests to severe drought and the capability to predict tree mortality risk at regional scales. The increasing frequency of flight campaigns by organizations like the National Ecological Observatory Network (NEON) provides an opportunity to generate detailed maps of individual tree mortality to support risk analyses at the scale of the individual tree and support satellite-based estimates of tree mortality over a broader domain.

In the first chapter of this dissertation, my coauthors and I used lidar and multispectral surface reflectance from the NEON airborne observation platform to map individual tree mortality over a 160 km2 area during and after the 2012-2016 drought for two sites in California’s Sierra National Forest. We derived tree locations and crown perimeters from the lidar point clouds and used surface reflectance and changes in crown perimeters between 2013 and 2017 to map 2017 tree mortality for more than one million trees. We found that cumulative tree mortality after the drought could be as high as 50%. In addition, we found that the subsequent effects of wildfire after the drought can be severe, with the Blue Fire of 2021 killing almost the trees within its perimeter. While tree mortality at low elevations appeared to saturate near 50%, cumulative tree mortality at higher elevations can be considerably lower (25%), with elevated rates of mortality continuing in the years after the drought subsides.

In Chapter 2, we used the data set from Chapter 1 to retrospectively investigate biophysical drivers of tree mortality risk from a severe drought at the two sites in the Sierra Nevada using a machine learning method called extreme gradient boosting. Our classification models of tree mortality performed better on the lower elevation site (74% accuracy on the validation and held-out test data sets), which experienced high mortality (50%) during the drought. The most significant driving variables we explored at this site were tree height, distance to rivers, and canopy cover fraction. We found that our models trained on data from one study area did not perform well at the other, highlighting the importance of developing tree mortality benchmarking data sets, which encompass a broad domain for training predictive models of tree mortality.

In Chapter 3, we aimed to create a model to estimate tree mortality fraction over a broader domain. We linked the individual tree mortality data set from Chapter 1 to the Landsat time series using one-dimensional convolutional neural networks. The R^2 values for the relationship between the mortality fraction observations and convolutional neural network predictions was 0.44 for the entire data set, including pixels with no trees, and 0.57 when we filtered for pixels with at least four trees. Our model enables the expansion of tree mortality estimates to broader spatial domains, which may help uncover fundamental interactions among biophysical drivers of tree mortality needed to generalize process-based tree mortality models at regional scales. This is important because these models are used to predict the biosphere's response to current and future climate to help predict future concentrations of atmospheric carbon.

Our approaches and datasets provide a means to estimate tree mortality and predict tree mortality risk at the scale of individual trees across our study domain in the Southern Sierra Nevada to broader scales across California. These analyses may improve our understanding of forest dynamics after severe drought and subsequent wildfire to improve projections of forest structure and carbon cycling in the Anthropocene.