Assessing Coniferous Forest Cover Change and Associated Uncertainty in a Subbasin of the Great Salt Lake Watershed: A Stochastic Approach Using Landsat Imagery and Random Forest Models

We present a stochastic method for classifying high-elevation coniferous forest coverage that includes an uncertainty estimate using Landsat images. We evaluate trends in coniferous coverage from 1986 to 2024 in a sub-basin of the Great Salt Lake basin in the western United States This work was completed before the recent release of the extended National Land Cover Database (NLCD) data, so we use the 9 years of NLCD data previously available over the period from 2001 to 2021 for training and validation. We perform 100 draws of 5130 data points each using stratified sampling from the paired NLCD and Landsat data to generate 100 Random Forest Models. Even though extended NLCD data are available, our model is unique as it is trained on high elevation dense coniferous stands and does not classify wester pinyon (Pinus edulis) or Utah juniper (Juniperus osteosperma) shrub trees as “coniferous”. We apply these models, implemented in Google Earth Engine, to the nearly 40-year Landsat dataset to stochastically classify coniferous forest extent to support trend analysis with uncertainty. Model accuracy for most years is better than 94%, comparable to published NLCD accuracy, though several years had significantly worse results. Coniferous area standard deviations for any given year ranged from 0.379% to 1.17% for 100 realizations. A linear fit from 1985 to 2024 shows an increase of 65% in coniferous coverage over 38 years, though there is variation around the trend. The method can be adapted for other specialized land cover categories and sensors, facilitating long-term environmental monitoring and management while providing uncertainty estimates. The findings support ongoing research forest management impacts on snowpack and water infiltration, as increased coniferous coverage of dense fir and spruce increases interception and sublimation, decreasing infiltration and runoff. NLCD data cannot easily be used for this work in the west, as the pinyon (Pinus edulis) and juniper (Juniperus osteosperma) forests are classified as coniferous, but have much lower impact on interception and sublimation.