New satellite-based sampling protocols for grapevine nutrient monitoring

Abstract

Context and purpose of the study – Extension specialists often recommend nutrient monitoring through leaf blade or petiole sampling twice a season for each vineyard block. However, due to the time and labor required to collect a large, random sample, many growers complete the task infrequently or incorrectly. Readily available remote sensing images capture the vineyard variability at both spatial and temporal scales, which can capture canopy and soil variability and be used to guide growers to representative sampling locations.

Material and methods – Mean composites of Sentinel-1 Synthetic Aperture Radar (SAR) images as a proxy of soil characteristics and Sentinel-2 Normalized Difference Vegetation Index (NDVI) as a proxy of canopy characteristics were clustered into three clusters (low-medium-high variability zones) using the Kmeans++ algorithm. Two spatial sampling protocols: (i) Grower Path (GP) (ii) NDVI+SAR3 and one standard Random20 (R20) protocol, were tested against the full block nutrient concentration (control of the study). R20 was a computer-generated random sample of 20 locations in each vineyard block. GP consisted of three sampling locations which were the centroid of the low-medium-high variability zones. NDVI+SAR3 was one location sampling grid (30mx30m) calculated using the mean absolute distance between each pixel and its cluster centroid. Field-specific sampling trials were conducted at bloom and veraison in the vineyards of Western New York and the Finger Lakes region in 2021 and 2022. Both macro (N, P, K, Ca, Mg) and micro-nutrients (Al, B, Cu, S, Fe, Mn, Na, Zn) were analyzed. All pixels were sampled for two blocks of cultivars –  Riesling and Concord. The mean absolute percentage error (MAPE) was calculated for each block, comparing GP, NDVI+SAR3, and R20 with overall nutrient concentration.

Results – R20 explained overall nutrient variation with approximately <1% MAPE for macro and micronutrients at bloom and veraison in both years. In comparison, GP had higher error rates for macro (3.6%) and micro-nutrients (8.9%) at bloom and similar with 3.8% and 9.4% error at veraison. At bloom, GP captured variability of important macronutrients like N, P, and K with 4.2%, 6.9% and 1.0% error rates. Micro-nutrients like Cu and B had higher errors of 9.2% and 6.8%, respectively. At veraison, these error rates were approximately the same for macronutrients but much larger for micro-nutrients. NDVI+SAR3 exhibited lower errors compared to GP and slightly higher errors compared to R20. The MAPE for N, P, K and Mg for macronutrients was between 1-3% at bloom and veraison. For micronutrients, like Cu and B, the MAPE was 2%-3% at bloom, almost doubling at veraison (6%). The errors were marginally higher at veraison than bloom across all sampling protocols, with a difference of <0.5% for macro-nutrients and <2% for micro-nutrients using R20 and NDVI+SAR3. Further exploration should exploit narrow-band remote sensing images for the block’s different size, climate, soil and topography. Future work should use R20 nutrient concentrations to compare with spatial sampling protocols as it captures the vineyard variability adequately.