Retrieval of Surface Ozone from UV-MFRSR Irradiances using Deep Learning

High concentration of surface ozone is harmful to humans and plants. USDA UV-B Monitoring and Research Program (UVMRP) uses Ultraviolet (UV) version of Multi-Filter Rotating Shadowband Radiometer (UV-MFRSR) to measure direct, diffuse, and total irradiances every three minutes at seven UV channels (i.e. 300, 305, 311, 317, 325, 332, and 368 nm channels with 2 nm full width at half maximum). Based on the wavelength dependency of aerosol optical depths, there have been plenty of literatures exploring retrieval methods of total column ozone from UV-MFRSR measurements. However, few has explored the retrieval of surface ozone. The total column ozone is the integral of the multiplication of ozone concentration (varying by height and time) and cross section (varying by wavelength and temperature) over height. Because of the distinctive values of ozone cross section in the UV region, the irradiances at seven UV channels have the potential to resolve the ozone concentration at multiple vertical layers. If the UV irradiances at multiple time points are considered together, the uncertainty or the vertical resolution of ozone concentrations can be further improved. In this study, the surface ozone amounts at the UVMRP station located at Billings, Oklahoma are estimated from the adjacent (i.e. within 200 miles) US Environmental Protection Agency (EPA) surface ozone observations using the spatial analysis technique. Then, the (direct normal) irradiances of UVMRP at one or more time points as inputs and the corresponding estimated surface ozone from EPA as outputs are fed into a pre-trained (dense) deep neural network (DNN) to explore the hidden non-linear relationship between them. This process could improve our understanding of their physical/mathematical relationship. Finally, the optimized DNN is tested with the preserved 5% of the dataset, which are not used during training, to verify the relationship.