Hyperspectral imagery to monitor crop nutrient status within and across growing seasons
Nanfeng Liu,Philip A. Townsend,Mack R. Naber,Paul C. Bethke,William B. Hills,Yi Wang
DOI: https://doi.org/10.1016/j.rse.2021.112303
IF: 13.5
2021-03-01
Remote Sensing of Environment
Abstract:<p>Imaging spectroscopy provides the opportunity to monitor nutrient status of vegetation. In crops, prior studies have generally been limited in scope, either to a small wavelength range (e.g., 400–1300 nm), a small number of crop cultivars, a single growth stage or single growing season. Methods that are not time- or site-specific are needed to use imaging spectroscopy for routine monitoring of crop status. Using data from four cultivars of potatoes (<em>Solanum tuberosum</em> L), three growth stages and two growing seasons, we demonstrate the capacity of full-range (400–2350 nm) imaging spectroscopy to quantify nutrient status (petiole nitrate, whole leaf and vine total nitrogen) and predict tuber yield in potatoes across cultivars, growth stages and growing seasons. We specifically tested the capabilities of: (1) ordinary least-squares regression (OLSR) using traditional hyperspectral vegetation indices (VIs); (2) partial least-squares regression (PLSR) using full spectrum (400–2350 nm), VNIR- (visible-to-near infrared: 400–1300 nm) or SWIR-only (shortwave infrared: 1400–2350 nm) wavelengths; (3) predictive models developed for one potato type or planting season on withheld data from a different type or season. Our results show that OLSR models produced poor predictions with data from all dates pooled together (validation <em>R</em><sup>2</sup> < 0.01). Single-date OLSR models performed better (<em>R</em><sup>2</sup> = 0.20–0.60, relative RMSE = 15–30%). PLSR models performed well and were comparable using different spectral regions (full-spectrum, VNIR-only and SWIR-only), with validation <em>R</em><sup>2</sup> = 0.68–0.82 and RRMSE = 12–25%. Testing across potato types, models produced reliable predictions (<em>R</em><sup>2</sup> = 0.45–0.75, RRMSE = 13–30%), but with some bias. Cross-season models had validation <em>R</em><sup>2</sup> = 0.46–0.75 and RRMSE = 17–100%, with a more significant bias than the cross-potato type models. To achieve models that are generalizable and robust, we recommend: (1) obtaining ground measurements that capture the full range of plant growth conditions and developmental stages, and (2) ensuring that image processing approaches minimize spectral discrepancies among dates.</p>
environmental sciences,imaging science & photographic technology,remote sensing