| Bio-optical properties of the ocean have been broadly investigated using coastal zone color scanner (CZCS) and Sea-viewing Wide Field-of-view Sensor (SeaWiFS) imagery, from which numerous globally and regionally optimized bio-optical (chlorophyll a) algorithms have emerged. The situation for optically complex inland waters, particularly large lakes, is quite different. In this case, similar regionally optimized bio-optical algorithms do not yet exist. As a first step in developing regional bio-optical algorithms for the Keweenaw region of Lake Superior, we tested ten published marine algorithms (nine empirical and one semi-analytical) using optical data and discrete water samples collected during 1998--2000. The semi-analytical approach better approximated chlorophyll a concentrations in the Keweenaw region of Lake Superior compared to empirical algorithms. We concluded that optimization of the semi-analytical algorithm, using observed regional optical and chlorophyll a data, would improve our results for the Keweenaw region of Lake Superior. In the second step of this study, we found the semi-analytical algorithm closely approximated the measured chlorophyll a concentrations using measured absorption coefficients of phytoplankton at 675 nm (aϕ(675)). However, the expressions involved in deriving aϕ(675), along with regionally optimized parameters, did not return meaningful values of aϕ(675) for the Keweenaw region of Lake Superior. One problem is that the particulate backscattering, an important inherent optical property, is poorly understood for Lake Superior. More samples of particulate backscattering spectra need to be collected in order to estimate aϕ(675) values accurately from remote sensing reflectance. Instead of developing a new conventional algorithm in the third step of this study, we employed an artificial neural network technique to estimate chlorophyll a concentrations in the Keweenaw region of Lake Superior. Six artificial neural networks were constructed and trained using combinations of remote sensing reflectance (Rrs) at 412, 443, 490, 510, 555, 665 and 685 nm. The best training result was obtained from the neural network using combination of Rrs values at 555, 665, and 675 nm as inputs. The prediction error of the resulting neural network was below 30%. |
