Southern Sierra Nevada Fractional Land Cover 2014 - 2018

This dataset is designed to capture landscape changes brought on by a severe tree mortality episode in California's southern Sierra Nevada range, by mapping multiple land cover types in a continuous way at annual time-steps throughout the mortality event. By mapping multiple land cover types we can see not only the occurrence of tree mortality explicitly, but also associated landscape changes, such as reduced live forest cover and increased bare ground cover. Fractional cover maps provide a continuous measure of the percentage of a land cover type present within each pixel across a landscape. We modeled fractional land cover in the southern Sierra Nevada at annual time steps (2014-2018) using a two-step approach. First, we created high resolution (5x5-meter pixel) categorical models of land cover for a ~2000 km2 area by training random forest classification models on RapidEye (Planet Labs Inc.) satellite spectral data for each year (McGregor et al, 2021 https://doi.org/10.7267/j9602723w) . Training data for this step were created by visually interpreting random plots over high resolution NAIP aerial imagery. Land cover classes included live forest, tree mortality, shrub, herbaceous, and bare ground. Class definitions generally follow California Wildlife Habitat Relationships (CWHR) categories. CWHR tree dominated habitats were combined to represent a single “live forest” category where trees were alive and “tree mortality” where dead. Shrub dominated habitats, namely Mixed and Montane Chaparral, were included in the “shrub” class. Perennial grasslands and wet meadow habitats were included in the “herbaceous” class, and non-vegetated habitats along with annual grasslands made up the “bare ground” class. We interpreted 75-125 samples per land cover class, per year. Model accuracy was evaluated using a 30% holdout of training data per class. To construct fractional cover models, we created new training data representing land cover proportions by overlaying 30x30-meter square plots (matching Landsat pixel resolution) with the 5-meter categorical maps. Proportional values for each class (excluding herbaceous) were calculated based on the 36 categorical pixels nested within each 30m plot. These new training data were then used to model the relationship between proportional cover values and Landsat 8 derived spectral covariates using a random forest regression model. See metadata file for accuracy assessment details. The output pixel values represent estimates of the percentage of a land cover type within each pixel. Pixel values should be multiplied by 0.0001 to convert from integer to decimal values representing proportions (0.0 - 1.0). Tree mortality was not modeled in 2014 as it was scarce in the study area at that time, and no fractional cover maps were created for herbaceous cover, leaving a total of 19 individual fractional cover maps. Tree mortality estimates represent the proportion of a pixel covered by dead trees, and includes all standing dead trees regardless of time since mortality.