Open AccessUsing the remote sensing methods is the most promising for day-to-day control of oil pollution. The laser-induced fluorescence method provides efficient detection and classification of oil pollutions. To monitor oil pollutions on the earth surface is more complicated than on the water one because of lower fluorescence intensity and interfering fluorescence of natural objects available on the earth surface.
Properties of oil pollution classifiers depend largely on the number and positions of spectral bands of fluorescence registration. Reducing the number of spectral bands allows us to diminish computation complexity and cost of equipment. In some cases the reduction increases classification accuracy. The number of spectral bands can be reduced through increasing their width.
The paper presents mathematical modeling of oil pollution detection and classification. The experimentally obtained fluorescence spectra of oil pollutions on different substrates were used as input data. The k-nearest neighbors algorithm was used to detect and classify oil pollutions. Cross validation was applied in mathematical modeling.
The mathematical modeling results have shown that for oil pollutions detection using over 8 spectral bands (band width less than 50 nm) a classification error rate does not depend on the further increasing number of the spectral bands.
As to the type classification of oil pollutions (4 classes), an increasing width of the spectral bands up to 60 nm (the number of spectral bands reduced up 7) does not lead to a significantly decreasing overall classification accuracy.
In the case of the sort classification of oil pollutions (8 classes) a local maximum of the overall accuracy has been observed at 25-30 nm width of the spectral band (14-16 spectral bands). The spectral resolution improvement (increasing the number of bands) does give an essentially increasing accuracy.
The paper has shown that to detect and classify oil pollutions on the earth surface there is no need to acquire fluorescence with high spectral resolution and numerous spectral bands.