REGRESSION-ANALYSIS OF MICROWAVE-SPECTRA FOR TEMPERATURE-COMPENSATED AND DENSITY-INDEPENDENT DETERMINATION OF WHEAT MOISTURE-CONTENT

Partial least-squares regression (PLSR) was used to generate wheat moi sture content predictive models from eight-frequency microwave attenua tion (A) and phase (P) spectra in the 10.36 to 18.0 GHz range, as obta ined by a free-space technique with a 10.4 cm thick sample. Spectra (n = 379) were measured for a set of grain samples that had been treated to span the agriculturally practical ranges of moisture content (M) ( 10.6 to 19.2% g/g(wet)), temperature (K) (-1 to 42 degrees C), and hul k density (D) (0.72 to 0.88 g/mL). The sample-property space formed by M, K, and D was used to prune redundant samples and select representa tive subsets for calibration (n = 279), cross-validation (n = 40 segme nts), and testing (n = 31). Twelve model types are reported and vary f rom attenuation or phase alone to the combination of attenuation, phas e, temperature, and density (i.e., APKD). For optimization of each PLS R model, the raw spectral, temperature, and density data were preproce ssed with variable ratios, mathematical transformations, and/or variab le scaling. The lowest moisture prediction errors were for temperature - and density-corrected models with variables AKD or APKD; these produ ced root-mean-square cross-validation and prediction errors (RMSECV an d RMSEP) of 0.19 to 0.20% in moisture content units. The more practica l unifrequency models, APK at 15.2 GHz, and AK at 18.0 GHz, yielded RM SECV values of 0.21% and 0.35%, respectively. Addition of temperature to dielectric data always substantially reduced the model error. Howev er, the multiplicative effect of density is well corrected by using th e ratio A/P, or partly corrected by using the features in the attenuat ion spectra. Data trends suggest that dual-frequency PK models might b enefit from a wider frequency range, and unifrequency AK models might be better at frequencies higher than 18.0 GHz. The results presented m ake it possible to evaluate a wide variety of instrumental configurati ons that might be proposed to suit particular engineering criteria suc h as measurement accuracy, range of operating conditions, and hardware complexity.