Neural network-based retrieval of phytoplankton abundance from remotely-sensed ocean radiance

An ocean color inversion model is presented for a wide variety of oceanic and coastal waters. The model is based on neural networks trained with realistic synthetic datasets. The model presented here retrieves chlorophyll a concentration as a proxy for phytoplankton abundance. One advantage to the model presented here is that inversion is not limited only to phytoplankton abundance, but rather the same methods could be used to retrieve absorption due to colored dissolved matter, absorption or backscatter due to non-algal particles, or particle beam attenuation. Initial application of the model on MODIS satellite data is presented here; however, the method is applicable to airborne or satellite remote sensing. Results indicate that the neural network inversion performs very well compared to the common OC4 empirical [chl a] algorithm, offering reduction of mean absolute deviation of error and root mean square error by factors of roughly 16 and 18 times, respectively.