Unmanned Aerial System (UAS) Imaging for Vineyard Nutrition: Comparing typical multispectral imagery with optimized band selection from hyperspectral imagery

The USA annual grape production value exceeds $6 billion across as many as 1 million acres (>400,000 ha). Grapevines require both macro- and micro-nutrients for growth and fruit production. Inappropriate application of fertilizers to meet these nutrient requirements could result in widespread eutrophication through excessive nitrogen and phosphorous runoff, while inadequate fertilization could lead to reduced grape quantity and quality. It is in this optimization context that unmanned aerial systems (UAS) have come to the fore as an efficient method to acquire and map field-level data for precision nutrient applications. Hyperspectral imagery (HSI) offers high spectral dimensionality to determine the optimized radiometric relationships with nutrients but is more expensive - in terms of acquisition and operation - than typical multispectral (MS) camera systems. We operated an UAS equipped with both a MicaSense RedEgde-M (5-band: blue, green, red, red edge, and near infrared) MS camera and a Headwall Nano-Hyperspec (272 bands: 400-1000 nm) HSI sensor over a Concord vineyard block in Portland, NY, USA at the onset of fruit ripening in 2020. Additional flights were flown over the same site during bloom and ripening in 2021. Coincident leaf blade samples were collected from 100 individual vines for laboratory analysis of nitrogen content (per dry weight). We directly compared the accuracy of two grapevine nitrogen models, the first a regression based on a typical 5-band MS camera, and a second based on five HSI bands, selected via an ensemble feature selection model, incorporating widely used base rankers (support vector, random forest, sequential feature selection, etc.). Early results from the 2020 data achieve a leave-one-out cross-validation RMSE of less than 0.2% nitrogen for samples ranging between 2.4-3.6% nitrogen. However, the optimized bands from HSI imagery provide limited improvement to the same sample nitrogen content predicted with MS imagery. We will explore this comparison further by incorporating 2021 imagery and nitrogen samples to assess multiple growth stages and seasons.