Hyperspectral imagery to monitor crop nutrient status within and across growing seasons

Authors: NANFENG, LIU - University Of Wisconsin; WANG, YI - University Of Wisconsin; NABER, MACK - University Of Wisconsin; Bethke, Paul; HILLS, WILLIAM - University Of Wisconsin; TOWNSEND, PHILIP - University Of Wisconsin

Submitted to: Remote Sensing of Environment
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 1/9/2021
Publication Date: 1/20/2021
Citation: Nanfeng, L., Wang, Y., Naber, M.R., Bethke, P.C., Hills, W.B., Townsend, P.A. 2021. Hyperspectral imagery to monitor crop nutrient status within and across growing seasons. Remote Sensing of Environment. 255. Article 112303. https://doi.org/10.1016/j.rse.2021.112303.
DOI: https://doi.org/10.1016/j.rse.2021.112303

Interpretive Summary: Timely monitoring of plant nutrient status is critical to making decisions about when to fertilize potatoes. Plant tissue analysis is used currently to diagnose potato nitrogen status. This method is time consuming and does not readily capture variation in nitrogen status that occurs across a commercial field. We investigated the potential of aerial remote sensing to predict potato foliar nitrogen status, as well as potato yield. Field data and images from an airborne sensor were collected from experimental field plots. Across cultivars, predictions based on imaging data were moderately-to-strongly correlated (R2=0.45-0.75) with field observations. Cross-season models were less well correlated with the measured data. To achieve more robust models may require additional data collection and new data processing approaches that minimize technical differences between collection dates. These findings demonstrate that aerial imaging is a promising technique for monitoring potato nitrogen status, and that additional experimentation is required to improve the accuracy of the results.

Technical Abstract: Timely monitoring of plant N status is critical to N fertilization decisions, thereby affecting the production and quality of potatoes (Solanum tuberosum L.). Traditional plant tissue analysis used to diagnose potato N status is time consuming and cannot characterize the spatial variability within a field. For this reason, we investigated the potential of aerial hyperspectral remote sensing for predicting potato foliar N concentrations (including petiole NO3-N, whole leaf and vine %N), as well as tuber yield. Specifically, we wanted to answer the following questions: (1) how do traditional hyperspectral VIs (Vegetation Indices) perform in modelling crop variables of potatoes? (2) does VNIR- (Visible-to-Near infrared: 400-1300 nm) or SWIR-only (Shortwave infrared: 1400-2500 nm) spectrum have a predictive ability comparable to full-spectrum (400-2500 nm)? (3) are predicting models transferable across different potato cultivars and planting seasons? Field data and airborne Hyspex (Norsk Elektro Optikk, Norway) images were collected at an experimental field for four potato cultivars and two growth stages of two planting seasons. OLSR (Ordinary Least Square Regression, VIs used as predictors) and PLSR (Partial Least Square Regression, reflectance at all spectral bands used as predictors) were performed to calibrate crop variables to hyperspectral data. Our results showed that OLSR models generally produced poor predictions with data from all dates pooled together (validation R2 <0.01). Results from single-date OLSR models showed that dry matter VIs using the wavelengths within the SWIR exhibited a predictive ability comparable to chlorophyll and water VIs using the wavelengths within the VNIR (R2=0.20-0.60, relative RMSE=15-30%). The performance of PLSR models using different spectral regions (VNIR-only, SWIR-only and full-spectrum) was comparable, with a validation R2=0.68-0.82 and RRMSE=12-25%. Across cultivars, models produced predictions moderately-to-strongly correlated (R2=0.45-0.75, RRMSE=13-30%) with observations, but with an apparent bias. Cross-season models had a validation R2=0.45-0.75 and RRMSE=17-100%, with a more significant bias than the cross-cultivar models. To achieve a success of generalizable/robust models, we suggest to: (1) obtain ground measurements that capture a dynamic range of plant conditions and growth stages, and (2) used new processing approaches to minimize the spectral discrepancy among dates.