Classifying Land Cover across the Pacific Ocean using Random Forests

Coastal marine environments have repeatedly been shown to be impacted by adjacent, onshore anthropogenic development. Cities and agricultural lands leak turbid waters, rich with heavy metals and fertilizer, into their surrounding environments, causing eutrophication and overall species mortality. Accurate land cover data is necessary to study these impacts on coastal systems, which is particularly time-critical given the combined stressors of anthropogenic development and sea-level rise. Unfortunately, for world-wide land cover data that is publicly available, we have found extremely low rates of accuracy within the Pacific region (e.g., numerous urban areas on islands misclassified as savannahs). While more accurate land cover data - quality controlled by humans and made available by NOAA Digital coast - within the Pacific do exist, they are regionally limited to the U.S. States and Territories. To address this limitation and recognition for the need for topical land cover data for the entire Pacific, we have created our own Pacific-focused land cover dataset. Specifically, we have reduced latitudinal bias while still expanding in spatiotemporal resolution by using the US Pacific islands NOAA Digital Coast layers in concert with imagery compiled from LANDSAT, MODIS, and Sentinel data to train, test, and validate a Random Forest Classifier within Google Earth Engine. Our land cover classifications are at a 30-m resolution across the tropical and subtropical Pacific regions and have a yearly temporal resolution beginning 2000 to present day. We present this data, our steps in creating it, and development patterns that we have observed.