Prediction of Nitrate Concentration of Two Bioenergy Feedstock Grasses through Using a Spectroradiometer.

Rapidly and accurately monitoring crop nitrate concentration is critical for both plant nutrient management and animal health, but can be difficult. Traditional methods are laborious and require destructive plant samplings followed by chemical analyses, thus, alternative methods are warranted. The objective of this research is to design a rapid nitrate assessment and evaluation method on two native warm-season grass species, including Alamo Switchgrass (Panicum virgatm L.) and Cheyenne Indiangrass [Sorghastrum nutans (L.) Nash.] using both remote sensing-based precise agronomic instruments and regression-based mathematical models. Both grass species were planted in a greenhouse under controlled growing conditions and fertilized with urea at 0 kg/ha (control), 65 kg/ha (low) 130 kg/ha (medium), and 260 kg/ha (high) of N. Plant height and leaf chlorophyll data were recorded weekly. All spectral data were recorded using a hyperspectroradiometer and then converted to ASCII format for regression analysis. We implemented computational algorithms using Matlab software based on a Generalized Regression Neural Network model to recognize the spectral pattern differences and predict the nitrate concentration across two grass species. Classification models were successfully constructed and validated. Both plant height and foliar SPAD readings were strongly affected by species (P < 0.05), N treatment (P < 0.05), and harvesting date (P < 0.05). No interactions were detected in plant height responses. Both two-way and three-way interactions were found in foliar SPAD readings, however, these interactions appeared due to differences in magnitude effect only. The prediction models accurately estimated botanical nitrate concentration even at its low level (R² = 0.71, RMSE = 0.84). Future studies are warranted to generate a broader range of botanical nitrate concentration, which could greatly increase the applicability and robustness of our prediction models.