A Machine Learning and Data Fusion Approach for Classifying High-Resolution Operational Land Imager Spectral and Vegetation Dynamic Data in Support of Habitat Mapping in the Transboundary Region of the Sonoran and Mojave Deserts.

Existing geospatial data and tools are inadequate for producing high-resolution vegetation maps of the Sonoran and Mojave Deserts due to their complexity and the discontinuity in data availability across the United States-Mexico border. Vegetation distribution and dynamics are critical to guiding land management and conservation decisions throughout the region, particularly in Bird Conservation Region (BCR) 33, The Sonoran and Mojave Deserts. This research will produce the first high-resolution vegetation community map of BCR 33 by prototyping new methods for desert vegetation classification through machine learning and openly sourced data. The proposed methodology will be prototyped at the Santa Rita Experimental Range site, located within the Sonoran Desert of Arizona. NSF National Ecological Observatory Network-Airborne Observation Platform 1 m hyperspectral and 10 cm RGB images, field data, and ground-level images will be used to train, calibrate, and validate a machine-learning land cover classification algorithm. Training data includes Landsat 8 Operational Land Imager spectral and time-series data from 2013-2020 augmented with phenology metrics and ancillary field data. The initial model will be based on the Random Forest classification method and will generate a fine-scale land cover distribution map with a special focus on riparian vegetation communities which are critically important for desert wildlife and are often misclassified in existing vegetation community maps. This early prototyping effort will serve as a proof of concept for the machine learning and data fusion methods that will be used for generating the high-resolution dynamic vegetation community map for the whole BCR 33 region.