Long-term Evaluation of Three Satellite Ocean Color Algorithms for Identifying Harmful Algal Blooms (Karenia Brevis) Along the West Coast of Florida: A Matchup Assessment

We present a simple algorithm to identify Karenia brevis blooms in the Gulf of Mexico along the west coast of Florida in satellite imagery. It is based on an empirical analysis of collocated matchups of satellite and in situ measurements. The results of this Empirical Approach is compared to those of a Bio-optical Technique - taken from the published literature - and the Operational Method currently implemented by the NOAA Harmful Algal Bloom Forecasting System for K. brevis blooms. These three algorithms are evaluated using a multi-year MODIS data set (from July, 2002 to October, 2006) and a long-term in situ database. Matchup pairs, consisting of remotely-sensed ocean color parameters and near-coincident field measurements of K. brevis concentration, are used to assess the accuracy of the algorithms. Fair evaluation of the algorithms was only possible in the central west Florida shelf (i.e. between 25.75°N and 28.25°N) during the boreal Summer and Fall months (i.e. July to December) due to the availability of valid cloud-free matchups. Even though the predictive values of the three algorithms are similar, the statistical measure of success in red tide identification (defined as cell counts in excess of 1.5 × 10(4) cells L(-1)) varied considerably (sensitivity-Empirical: 86%; Bio-optical: 77%; Operational: 26%), as did their effectiveness in identifying non-bloom cases (specificity-Empirical: 53%; Bio-optical: 65%; Operational: 84%). As the Operational Method had an elevated frequency of false-negative cases (i.e. presented low accuracy in detecting known red tides), and because of the considerable overlap between the optical characteristics of the red tide and non-bloom population, only the other two algorithms underwent a procedure for further inspecting possible detection improvements. Both optimized versions of the Empirical and Bio-optical algorithms performed similarly, being equally specific and sensitive (~70% for both) and showing low levels of uncertainties (i.e. few cases of false-negatives and false-positives: ~30%)-improved positive predictive values (~60%) were also observed along with good negative predictive values (~80%).