Calibration transfer between SWNIR Si PDA instruments

Silicon photodiode array (PDA) detectors find use within spectrometers targeted to high speed, low cost applications, and applications that require use of the Herschel region of the spectrum. It is obvious that matching of instrument optics must minimise the software correction required in the calibration transfer process, however, in practice photodiode spectrometers vary in both photometric response and wavelength accuracy.

The photometric response of four Zeiss MMS1 spectrometers (pixel step of ca. 3.3 nm, FWHM 10 nm) was assessed using a stabilised halogen light source with combinations of neutral density filters and bandpass spectral filters. Wavelength assignments of five Zeiss MMS1 spectrometers were assessed using a Mercury-Argon calibration source and compared to the “manufacturer assigned” wavelengths. New pixel wavelength assignments were generated for the 695 – 1014 nm range.

Interactance spectra were acquired of 88 apples, with two spectra acquired from each of two sides of the fruit, using three MMS1 units. Subsets of the spectra were used for calibration (53 fruit), validation (20 fruit) and piecewise direct transfer (PDS) (15 fruit). PLS regression models for fruit TSS were developed on each unit and were iteratively tested against the unaltered and PDS transformed spectra of the other units, using the original and new wavelength calibration, and a artificial shift in the wavelength scale.

The photometric response of the units varied by wavelength, however a high correlation was evident between the photometric responses of corresponding pixels in different PDAs. Given this linear response, absorbance pre-treatment should correct for photometric response differences between instruments.

An artificial shift in the wavelength scale of as little as 0.5 nm produced an increase in RMSEP of 1.84 and a bias of -2.19. The third order polynomial coefficients provided in the Zeiss specification sheet for pixel to wavelength conversion across the whole spectrum in an error of up to 3.2 nm or up to 2.3 nm for the wavelength range used in the PLS regressions on TSS. When using the newly assigned wavelengths, the performance of the models for each unit improved, for both unchanged and PDS transformed spectra, in terms of SEP. The SEP improved as much as 0.24 SEP for the unchanged spectra. In every case, the new wavelengths and PDS combination performed better (RMSEP, SEP and Bias), than PDS with the original wavelengths. The bias evident in the transferred models may be dealt with using simple slope and bias correction.