Место издания:SARATOVSKIY ISTOCHNICK Saratov, Russia
Первая страница:71
Последняя страница:72
Аннотация:Natural dissolved organic matter (DOM) is the main (~ 95%) form of organic carbon in the oceans, and, therefore, it is the important component of marine ecosystems and the key chain in the carbon cycle. The production of plankton and algae, and organic matter from land (terrigenous) are the main sources of DOM in the ocean. The basis of terrigenous organic matter is humic substances (HS). They have a significant effect on the color of water, and, on the one hand, they can limit photosynthesis by absorbing light in the blue region, and on the other, HS limit the negative effects of UV radiation on plankton absorbing the UVB (280-320 nm) and the UVA (320-400 nm) bands of sunlight. Simultaneously the quantitative assessment of HS is very important for remote sensing (for example estimating chlorophyll concentrations using satellite imagery) in the visible range due to the effect on the optical properties of water. Rapid optical techniques are best suited for the HS quantitative determination. It is known that the DOM content has a good with absorption coefficient at 350 nm and the UV portable lidars (fluorescence) are used for the evaluation of DOC content in the surface layer. However, the fraction of the HS in DOM is not commonly estimated. It seems appropriate to use multivariate calibration models for the determination of the total HS content due to their diversity resulting in complex shape of fluorescence spectra. The Latin hypercube provides a low correlation of factors in design even for a small number of experiments. This allowed us to use a compact training set of 35 samples for 5 factors. Although, we measured EEM fluorescence spectra,initially we constructed PLS calibration models based on fluorescence spectra at typical excitation wavelengths of lidars and the most common excitation wavelength in the OpenFluor database of natural fluorophores. Thereafter we have applied sparse PLS for the full EEM spectra. We selected the best models (1-4 principle components) in terms of prediction error by leave-one-out crossvalidation. The models also provide physically significant loadings. The good predictive ability of the obtained calibration models was proved by a good correlation of the predicted HS contents with the dissolved organic carbon content in an independent set of Arctic determined by the combustion method.