Abstract:

Harmful algal blooms (HABs) in the Chesapeake Bay are challenging to characterize from satellite observations because of optically complex waters, mixed phytoplankton communities, variable environmental conditions, and limited coincident field observations. This NERTO project developed a reproducible framework integrating historical phytoplankton observations, NOAA satellite-derived algal bloom products, geospatial analysis, ecological information, and machine-learning methods to support future AI-assisted discrimination of diatom- and dinoflagellate-dominated events. Field observations were matched with georeferenced satellite imagery and evaluated through a structured quality-control workflow addressing mixed pixels, land contamination, clouds, missing data, non-detect conditions, saturation, and coordinate inconsistencies. Of 31,531 field-to-satellite records evaluated, 860were initially classified as valid, while neighboring-pixel analysis identified additional potentially recoverable observations. A preliminary Mini U-Net segmentation model demonstrated the feasibility of the approach but also revealed limitations associated with automatically generated labels, motivating development of an interactive expert-review tool for event validation, masking, and labeling. Ecological analyses additionally explored seasonal phenology and environmental variables as potential features for future multimodal models