Hyperspectral indices optimization algorithms for estimating canopy nitrogen concentration in potato (Solanum tuberosum L.)

Abstract Many empirical models based on hyperspectral indices (HIs) have been developed to estimate nitrogen (N) status of crops. However, most of the researches by far focused on the identification of sensitive bands of HIs, and have not identified the importance of formula formats to achieve their best performance. The current study aimed to investigate the response of band optimization and formula formats to canopy N concentration (CNC) of potato (Solanum tuberosum L.) plants, and to verify the performance of HIs through optimized algorithms based on a multi-site and -year study. Three field experiments involving different potato cultivars with 3–6 N rates were conducted from 2014 to 2016 in Inner Mongolia, China. The band optimization HIs were first tested using a simulated dataset by the PROSAIL model and validation dataset from farmers’ fields. Results showed that the optimized HIs generally had more robust performances for CNC prediction than the published indices. The optimized HIs explained 56%–74% of the variations in potato CNC in contrast with 3%–53% variation of the published HIs. Compared with published HIs, band optimization could significantly improve the performance of HIs by 16%–71%. The choice of the formula formats affected the explanatory power of the optimized HIs by 3%–18%. Our study found that both the performance of HIs and the position of the sensitive bands were greatly influenced by formula formats. The results from the evaluation of noise equivalent that independent from farmers’ field and PROSAIL model showed that the best performance was from Opt-CCCI. The central sensitive bands of Opt-CCCI were found at 600, 582, 650 nm. Opt-mRER also exhibited well in noise equivalent and independent validation from farmers’ field, while they could not be verified with the PROSAIL model because of the absence of the wavebands from 340 to 400 nm. Optimization for HIs indicates that there will be a great potential to improve the use of hyperspectral sensing for the estimation of field croṕs CNC.