Using a dataset consisting of 9000 reflectance spectra simulated using HYDROLIGHT 5 for a broad range of observable natural water conditions, we have developed three neural networks (NNs) working in parallel to model the inverse problem for both oceanic and coastal waters. These NNs are used to relate the water leaving remote sensing reflectance (Rrs) at available MODIS visible wavelengths (412, 443, 488, 531, 547 and 667nm) to the phytoplankton (aph), non-phytoplankton particulate (adm), dissolved (ag) absorption and particulate backscattering (bbp) coefficients at 443nm. These reflectance derived parameters (aph(443), adm(443), ag(443), bbp(443)) are then combined with the measured reflectance values and used as input to a fourth NN, (IOP NN [Chl]), to derive chlorophyll concentration ([Chl]). Unlike NNs previously developed by us that were trained on a synthetic dataset and then tested on the NASA Bio-Optical Marine Algorithm Dataset (NOMAD), the (IOP NN [Chl]) network was both trained and tested solely on NOMAD. Although the inherent optical properties (IOP) can be derived from the optical signal through their direct relation to the Rrs, the relationship of [Chl] to IOP varies with location and season, and is therefore difficult to model globally. In order to demonstrate that the inclusion of derived IOP estimates along with radiance measurements can improve the retrieval of [Chl], we construct a neural network that is trained to derive [Chl] from reflectance measurements only We also compare our [Chl] product to that obtained from the current OC3 algorithm implemented by NASA’s Moderate Resolution Imaging Spectroradiometer (MODIS). Finally, we apply our algorithm to MODIS data and present and analyze the global seasonal variability for all three parameters.