Using hyperspectral remote sensing to develop new insights into precision nitrogen management for potato production

• This study explored vegetation indices (VIs) calculated from 400 to 2500 nm hyperspectral images collected over multiple crop growth stages during three growing seasons (2018, 2019, and 2020) on two potato varieties to predict crop nitrogen (N) status and tuber yield via machine learning models. • Some VIs (such as TCARI, WI, NDTI, PRI, and SRWI) consistently exhibited high correlations with potato traits across various years and combinations of years. • The Xtreme Gradient Boosting models required both VIs and information about agronomic practices as inputs to consistently produce reliable predictions of leaf total N and final tuber yield across individual years and combinations of years. • Yield gains in response to N fertilization could be predicted for most imaging collection dates across various potato crop growth stages over three years. However, the N status of in-season crops in response to supplemental N application exhibited high year-to-year variations. • This study establishes a foundation for research on precision N management in potatoes by examining the interactions among plant growth, N fertilization, and the plant’s hyperspectral signatures. Hyperspectral remote sensing has demonstrated significant potential for monitoring nutrient status and predicting productivity of potato crops. This study utilized hyperspectral vegetation indices (VIs), calculated from imaging data collected over multiple crop growth stages during three growing seasons (2018, 2019, and 2020) on two varieties, to predict crop nitrogen (N) status and final yield using the Xtreme Gradient Boosting (XGB) model. It was found that several VIs (such as TCARI, WI, NDTI, PRI, and SRWI) exhibited consistently high correlations with the target traits across different growing seasons. It also revealed that incorporating agronomic factors, including N application rates, varieties, and crop growth stages as additional inputs, could greatly enhance the XGB model’s predictive accuracy across individual years and combinations of years. The Coefficient of Determination ( R² ) was 0.303-0.922 for petiole nitrate-N, 0.510-0.918 for whole leaf total N, and 0.581-0.834 for tuber yield. Furthermore, the most novel analyses of this study predicted changes in in-season N status and final yield due to fertilization using models developed from spectral signatures of non-fertilized and fertilized plots, respectively. The results indicated that yield gains in response to N fertilization could be predicted for most imaging dates over the three years; however, the in-season crop N status response to supplemental N showed high year-to-year variations. This study offers new insights (about timing and amount of N fertilizer application) into precision N management of the N-demanding potatoes through advanced hyperspectral imaging and machine learning techniques, contributing to the goal of minimizing unnecessary N fertilizer use and its adverse impact on the environment.