Study site.

The study site was located at the University of Florida Institute of Food and Agricultural Sciences Tropical Research and Education Center in Homestead, FL (25.5°N, 80.5°W). This location has a subtropical climate with an average annual air temperature of 24 °C and an average annual rainfall of 1572 mm (Florida Automated Weather Network 2020, 2021, 2022). The soil is classified as Krome gravelly loam soil (loamy-skeletal, carbonatic, hyperthermic lithic udorthents), which is intensely cultivated and relatively shallow (<23 cm deep) (Noble et al. 1996; US Department of Agriculture, Natural Resources Conservation Service 1991). In all trial years (2020, 2021, and 2022), experiments were performed in different areas of the same field to avoid legacy N effects. Background soil chemical and physical properties were determined annually before fertilization, except in 2020 (Table 1), when the soil was analyzed 4 months after experiment termination, in a nearby uncultivated area, due to pandemic-related setbacks.

Table 1. Background soil chemical and physical properties across site years (0 to 15 cm depth).

View Fullscreen

Nitrogen rate experiments.

Nitrogen rate experiments for flower hemp production were conducted to determine whether plant size and color differences resulting from six N fertilizer rates were detectable by drone-based aerial imaging and corresponded to harvested aboveground biomass. Initially, two varieties, ‘Wife’ and ‘Maverick’, were included, but the study focused on ‘Wife’, because it was more reliably detected by drone-based aerial imaging due to its larger plant size and smaller plant density. ‘Maverick’ is a dwarf autoflower variety that did not develop robust canopies for detection. ‘Wife’ is a high-cannabinoid, photoperiod-sensitive variety that was vegetatively propagated and grown in a greenhouse for approximately 5 weeks. Rooted propagules were transplanted into the field on 26 May 2020, 7 May 2021, and 20 Jun 2022 and were grown for 3 to 4 months. The target plant density was approximately 2990 plants/ha. Each experiment consisted of 24 experimental units arranged into a randomized complete block design with four blocks. Each unit (28 m²) was planted with six plants and treated with one of six N rates: 0, 56, 112, 168, 224, and 280 kg·ha⁻¹ N, applied in three split applications.

In 2020, all units received 56 kg·ha⁻¹ N monthly until each unit reached its designated N rate. By the third application, the units assigned to treatments of 224 and 280 kg·ha⁻¹ N received 112 and 168 kg·ha⁻¹ N, respectively, to reach their total assigned N rate. The first application was applied before transplanting using 44N–0P–0K environmentally smart nitrogen polymer-coated urea (Agrium U.S. Inc., Loveland, CO, USA), and the subsequent two applications were topdressed around the base of each plant (1 m²) as 46N–0P–0K prilled urea (TradeMark Nitrogen Corp., Tampa, FL, USA). In 2021 and 2022, N treatments were applied similarly but with one N source (as prilled urea; 46N–0P–0K). In 2021, the split applications were 60% pre-plant, 20% at the vegetative stage, and 20% at flowering and seed formation stage (Mediavilla et al. 1998), whereas in 2022, the split applications of N fertilizer were 30%, 30%, and 40%, respectively, at the same growth stages. In 2020 and 2021, 67 kg·ha⁻¹ P₂O₅ (as triple superphosphate; 0N–46P–0K) and 336 kg·ha⁻¹ K₂O (as granular sulfate of potash; 0N–0P–50K) were incorporated into the soil before transplanting. In 2022, the same phosphorus and potassium fertilizer sources were used, although application rates were reduced by half due to sufficient background soil levels. In the absence of rainfall, drip irrigation supplied 12.7 mm of water column weekly. Weeds were managed mechanically, manually, and with biodegradable paper mulches.

Harvest procedure.

Flower harvest was determined when at least 50% of floral trichomes’ resin transitioned to an amber color. Harvest occurred on 16 Sep 2020, 21 Sep 2021, and 20 Sep 2022. The main stem was cut at the soil surface to harvest aboveground fresh biomass for each plant, which was then weighed and recorded. In 2020, subsamples were taken to estimate dry matter, whereas in 2021 and 2022, each plant was oven-dried and hand-processed into separate flower, leaf, and stem fractions.

Image collection with drone.

We compared the harvested aboveground biomass from the field trials with spectral data collected about 1 month before harvest annually to identify differences in canopy size and reflectance resulting from applied nitrogen rates. Hemp farmers are required to determine harvest date no more than 1 month before harvest so that regulatory testing may be completed to approve the lot for harvest. Furthermore, the timing of the spectral data collection reflected this timing for harvest decision-making and allowed the hemp to reach harvestable floral maturity. Any plants that died during this period were excluded from the comparison.

Raw drone images were captured by a Sequoia multispectral sensor mounted onto a Parrot Bluegrass Fields drone (Parrot Drone SAS, Paris, France) after an inertial measurement unit and spectral calibration were performed. The drone had an autopilot system that allowed for autonomous navigation of an established flight plan over the field (Fig. 1). The Sequoia multispectral sensor, equipped with a 16-megapixel RGB camera and four multispectral sensors, captured canopy reflectance and drone images in the visible (green, red, and red edge) and NIR spectra with the following band specifications (Table 2). A sunshine sensor (Parrot Drone SAS) located on top of the drone continuously captured incoming light and adjusted for light variability that could alter reflectance data and image processing.

View Figure

• Download as Powerpoint
• Download Figure

Table 2. Sequoia multispectral sensor band specifications.

View Fullscreen

Drone flights and vegetation sensing were conducted between 10:00 AM and 1:00 PM EST to minimize shadows in collected drone imagery. The multispectral camera captured high-resolution images (2.5 cm pixel⁻¹) at a rate of one picture per second with a 70% side and front overlap at an altitude of 61 m. The drone operated at a speed of 5 to 7 m·s⁻¹ to cover the area of interest (2.2 ha) with a flight time of approximately 10 min.

Image processing and data products.

Acquired drone imagery was processed with Pix4DCapture software (Pix4D S.A., Prilly, Switzerland). Raw images were aligned and stitched together to create a seamless four-band (green, red, red edge, and NIR) stack orthomosaic with ground control points and a digital elevation model. The GPS locations for the plants were extracted from the flight images, and the final orthomosaic was clipped to the field boundary extent and uploaded to an online server. Plant canopies were outlined and extracted by supervised classification of the digital elevation model after automated spectral selection with artificial intelligence (Stanford et al. 2024). By distinguishing distinct canopy heights from surrounding vegetation and soil, we refined the AI extraction with manual corrections to address any errors related to canopy area overestimation, underestimation, and misidentification.

The NDVI and the GNDVI data products were evaluated for each pixel of the image using standard NDVI and GNDVI equations (Eqs. [1] and [2]) with results extracted for each canopy:$$\text{NDVI}=\frac{\text{NIR}-\text{Red}}{\text{NIR}+\text{Red}}$$ [1] $$\text{GNDVI}=\frac{\text{NIR}-\text{Green}}{\text{NIR}+\text{Green}}$$ [2] where “Red” denotes the red channel, 620 to 700 nm; “NIR” denotes the near-IR channel, 750 to 830 nm; and “Green” denotes the green channel, 510 to 590 nm. The ratio between the difference of the numerator and the sum of the denominator results in a VI value ranging from −1.0 to 1.0. The VI value informs vegetation conditions, such as canopy coverage, biomass, chlorophyll content, and plant health. Lower VI values are associated with nonvegetative surfaces or unhealthy, stressed, and sparse plantings, while higher values are associated with moderate to healthy and dense vegetation (Jackson and Huete 1991; Walthall et al. 2003).

Analysis.

The average plant aboveground biomass, floral yield, canopy area, NDVI, and GNDVI around harvest were evaluated for each unit with a 95% confidence interval to express variation. Statistical analyses were performed using a two-way analysis of variance [ANOVA, “aov()”] followed by a Duncan post-hoc test [“duncan.test()”; agricolae version 1.3-3] for each parameter using R studio software, version 4.0.2 (RStudio, Inc., Boston, MA, USA). The two-way ANOVA assessed both the main effect of nitrogen rate and the interaction effect of nitrogen rate and year for each variable. Significant ANOVA results (P < 0.05) were followed by a Duncan post-hoc test to identify differences between nitrogen rate groups and determine their consistency across years. To further explore the relationship between biomass, canopy area, and VIs, pairwise correlations were conducted using the linear model [“lm()”]. The analysis reported the coefficient of determination (r²) to indicate the strength and direction of their relationship.

Summary of results.

Hemp spectral characteristics were distinguishable from surrounding weedy vegetation and bare ground, whereby hemp was found to have a higher NIR and red edge reflectance (Fig. 1A). Through field evaluation and aerial imaging, we found that increasing N application rates up to 224 kg·ha⁻¹ positively increased values for aboveground biomass, floral yield, canopy area, and vegetation indices (Fig. 2). We also observed noticeable interactive effects by years driven by a high level of weed intensity in 2021 and a delayed planting in suboptimal conditions in 2022 (Figs. 3 and 4). Analysis is presented both in aggregate and by year to thoroughly evaluate trends and deviations. Although the statistical differences showed year-to-year variability, our aggregate results also indicate an overall pattern of plant production and health to N rate. We observed a positive linear correlation between biomass and canopy area across all experiment years (Fig. 5A–C) and also a strong positive correlation between log-transformed biomass and spectral indices (NDVI and GNDVI) over successive study years (Fig. 6A–F).

View Figure

• Download as Powerpoint
• Download Figure

View Figure

• Download as Powerpoint
• Download Figure

View Figure

• Download as Powerpoint
• Download Figure

View Figure

• Download as Powerpoint
• Download Figure

View Figure

• Download as Powerpoint
• Download Figure

Aboveground biomass and floral yield.

Results from the two-way ANOVA for all study years (2020 to 2022) found that both the main effects of N rate and the interaction effects of N rate and year were significant (P < 0.001) for aboveground dry biomass (biomass yield) and floral yield (Figs. 2A–B and 3A–F). Overall, maximum mean biomass yield (1252 kg·ha⁻¹) was obtained from the nitrogen application rate of 224 kg·ha⁻¹ N, which was significantly higher than all other treatment rates (Fig. 2A). Hemp definitively benefits from moderate to high rates of N fertilizer indicated by the minimum mean biomass yield (308 kg·ha⁻¹) at 0 kg·ha⁻¹ N without a significant difference at 56 kg·ha⁻¹ N (Fig. 2A). Maximum floral yield (482 kg·ha⁻¹) was obtained from applying 224 kg·ha⁻¹ N but was not significantly different from 112, 168, and 280 kg·ha⁻¹ N (Fig. 2B). The minimum floral yield (129 kg·ha⁻¹) resulted from 0 kg·ha⁻¹ N and did not significantly differ from 56 kg·ha⁻¹ N.

Evaluated across years, our results identified an N rate recommendation for floral hemp production between 112 to 224 kg·ha⁻¹ N. Similar N rate recommendations were established from global studies of hemp crops (Atoloye et al. 2022; Caplan et al. 2017a, 2017b; Papastylianou et al. 2018; Vera et al. 2010; Yang et al. 2021), despite the unique gravelly and calcareous condition of our soil. At nitrogen rates above the recommended range (i.e., 280 kg·ha⁻¹ N), biomass and floral yield were reduced by 21% and 17%, respectively, compared with the 224 kg·ha⁻¹ N rate. Similar findings were reported by Anderson et al. (2021) and Saloner et al. (2019), who observed that higher application rates of N fertilizer resulted in reduced plant growth, biomass accumulation, and harvestable yield in C. sativa crops. These results could be explained by nitrogen toxicity, which results in cell damage from increased intracellular pH, nitrate accumulation, salinity, and oxidative stress (Saloner and Bernstein 2020). Cannabis crops have been shown to express symptoms of N toxicity as unusually dark green foliage with downward curling leaves progressing into leaf necrosis and increased plant mortality (Anderson et al. 2021; Saloner and Bernstein 2020). The expression of phytotoxicity (i.e., dark green foliage) in hemp is an observable characteristic for understanding crop health from aerial imaging.

For results analyzed separately by year, the N fertilizer rate significantly affected biomass yield in 2020 and 2021 (P < 0.001; Fig. 3A and B), similar to the combined-year analysis, whereas no significant effect was found for 2022 (P = 0.598). In 2022, the highest mean biomass yield (1101 kg·ha⁻¹) was observed at 112 kg·ha⁻¹ N, and the lowest mean biomass yield (608 kg·ha⁻¹) at 280 kg·ha⁻¹ N (Fig. 3C). Differences between floral yields obtained from applying N at the different rates were significant (P < 0.001) for 2020 and 2021, similar to the combined-year analysis, but not significant for 2022 (Fig. 3D–F). Current nutrient recommendations for floral hemp production in Florida suggest application rates of 112 to 168 kg·ha⁻¹ N (Brym et al. 2024). These rates vary with respect to hemp variety and production target whereby the most vigorous and productive of varieties observed correspond to the higher range of the N rate recommendation. Conversely, poorer performing years that might experience establishment, weed, and disease challenges as were observed in 2021 and 2022 might not respond effectively to higher N rates. A standard nutrient management best practice for Florida is to split N applications to multiple application events during the year, allowing for adaptive timing and rate of fertilizer to match crop development and performance.

Canopy area.

As detected by drone-based aerial imaging, canopy area was observed to positively increase with N fertilization (Figs. 2C and 3G–I). Nitrogen rate, year, and the rate × year interaction significantly affected canopy area for combined-year analysis, and 2020 independently showed significant post-hoc groupings (P < 0.001). Combining years, the largest mean canopy area (0.51 m²) occurred with 224 kg·ha⁻¹ N but only differed significantly from canopy areas at 56 and 0 kg·ha⁻¹ N (Fig. 2C). Similarly, in 2020, lower fertilizer rates (0 to 56 kg·ha⁻¹ N) were statistically different from canopy area values at moderate to high fertilizer rates (>112 kg·ha⁻¹ N), whereas in 2021, the 280 kg·ha⁻¹ N rate resulted in an average canopy area that was statistically greater than canopy area at 0 and 168 kg·ha⁻¹ N yet not statistically different from average canopy area values at 56, 112, and 168 kg·ha⁻¹ N (P < 0.01). In 2022, no significant differences were observed between the fertilizer groupings (P = 0.847). In 2021, canopy area was significantly higher at 280 kg·ha⁻¹ N when compared with 0 and oddly with 168 kg·ha⁻¹ N (Fig. 3H). In 2022, canopy area was not significantly affected by fertilizer rate, although an increasing trend was observed with intermediate rates (112 to 224 kg·ha⁻¹ N).

Plant canopies appeared to be less sensitive to changes in N application rates than aboveground biomass, with increases at 56 and 112 kg·ha⁻¹ (Fig. 2A and C). While our study did not report data from the vegetative growth stage, similar N rate studies have shown that leaf and branch counts, as well as canopy coverage, positively responded to increasing N fertilization up to moderate levels (90 to 100 kg·ha⁻¹) during vegetative growth (Campiglia et al. 2017; Caplan et al. 2017b; Farnisa et al. 2023; Tang et al. 2017), which correlated to increases in growth and yield during flowering (Caplan et al. 2017a; Tang et al. 2018). In addition, canopy development, crop growth, and yield were negatively correlated to N fertilizer rates at or below 60 kg·ha⁻¹ N. The observed analytical results between plant biomass and canopy area suggest varying sensitivity of plant size and behavior to limited N rates (Barker and Pilbeam 2006; Saloner and Bernstein 2020; Zhang et al. 2007). For example, we observed a significant decrease in aboveground biomass and some instances of plant mortality at 280 kg·ha⁻¹ N, but no statistical difference was observed in canopy area. Although canopy area can serve as an indicator of plant size and growth, it may lack the practical resolution to distinguish aspects of nutrition and production status.

Vegetative indices.

NDVI and GNDVI calculated from canopy reflectance increased with increasing N rate (P < 0.001; Fig. 2D–E). The results were significant when aggregated and with an interaction by year (Figs. 2D–E and 4). NDVI values increased from 0.36 to 0.58 as N rates increased from 0 to 224 kg·ha⁻¹ N, whereas GNDVI values increased from 0.11 to 0.27. The highest mean values for NDVI (0.58) and GNDVI (0.27) were obtained at 224 kg·ha⁻¹ N, similar to observed biomass and floral yield. NDVI at 224 kg·ha⁻¹ N was statistically higher than those of all other N rates aside from the 168 kg·ha⁻¹ N rate. GNDVI at 224 kg·ha⁻¹ N was also a statistical maximum. At 280 kg·ha⁻¹ N, both VIs exhibited significant decreases in reflectance values when compared with 224 kg·ha⁻¹ N, with NDVI and GNDVI reflectance values decreasing by 19% and 22%, respectively (Fig. 2D–E). The lowest NDVI (0.36) and GNDVI (0.11) values were observed at the 0 kg·ha⁻¹ N rate. These results suggest that NDVI and GNDVI from aerial images are useful indicators for plant N status; however, the practical results were not consistent year to year.

When the results were analyzed for each year, N fertilization positively influenced VIs in 2020 and 2021, but not in 2022 (P < 0.01; Fig. 4). In 2021, fewer groupings limited the interpretation of VI. Canopies of hemp plants were extracted from aerial images to evaluate the VI, which were challenged by weed competition in 2021 and late season planting in 2022, likely affecting the outcome. Overall patterns in the results remain, such as the maximum mean value for NDVI observed at 168 and 224 kg·ha⁻¹ N in 2020 (0.61) and 2021 (0.69) and the minimum mean reflectance values under no fertilization in 2020 (0.40) and 2021 (0.25). The highest GNDVI values occurred between 168 and 224 kg·ha⁻¹ N in 2020 (0.37) and 2021 (0.25). In 2022, both VIs obtained their greatest values at 224 kg·ha⁻¹ N followed by a reduction in reflectance values at 280 kg·ha⁻¹ N, although these results were not statistically significant.

NDVI is used to monitor and assess crop health and nutrition, where an NDVI value of 0.58, as observed in our study, is associated with moderately healthy plants with dense canopies and good fertility (Nichiporovich and Radevich 2012). Conversely, NDVI values below 0.40 are associated with unhealthy, stressed, and sparse vegetation with nutrient deficiency. GNDVI is sensitive to chlorophyll content in vegetation associated with an increase in photosynthetic activity and crop biomass (Candiago et al. 2015; Cherlinka 2019; Li et al. 2008; Viña et al. 2011). An increase in N rate has been shown to affect chlorophyll content in Cannabis plants, where those treated with inadequate N resulted in N stress and lower chlorophyll content. Both indices are viable candidates for crop status determination and likely both obtainable by a multispectral camera. From our study, NDVI was more effective than GNDVI in detecting differences in N rate treatments, and the year 2020 was more discernable than the other years. Dense canopies of Cannabis, which were well-fertilized and watered, were found to be more spectrally discriminable than sparse canopies (Daughtry and Walthall 1998). Conversely, high weed pressure and smaller-sized Cannabis plants were a challenge to separate by drone imagery (Walthall et al. 2006). Well-maintained hemp crops make strong candidates for evaluation by aerial imaging with detailed in-season information possible to assist management decisions such as split N rate applications.

Plant biomass correlation with aerial imaging.

Plant biomass correlates well (r² > 0.65) with features generated by aerial imaging (i.e., canopy area, NDVI, GNDVI) evaluated independently for each year studied (Figs. 5 and 6). When evaluated across years, canopy area maintained a moderate relationship with plant biomass (r² = 0.557, y = 1412x + 317), whereas NDVI and GNDVI resulted in weak relationships with plant biomass (r² = 0.214, log10(y) = 2.45x +5.26; r² = 0.397, y = 2704x + 311). NDVI was the strongest correlation with aboveground biomass at harvest evaluated independently for each year (r² > 0.80), whereas GNVDI also showed strong correlation (r² > 0.76). This is consistent with studies reporting that NDVI was best suited for estimating aboveground biomass of wheat under different N fertilizer rates (Erdle et al. 2011; Geipel et al. 2016; Viña and Gitelson 2005), although an obvious operational preference would be a model that performs effectively across years instead of a model adapted for each year. These results indicate that plant identification, selection, and quantification from aerial imaging is a viable tool for determining hemp growth and production with the important consideration of specific site and year (Daughtry and Walthall 1998; Hall et al. 2014) (Fig. 1B). The observations from our study are well associated with changes in plant performance due to fertilizer application that can be further inferred from an assessment with aerial imaging. This study indicates that a strong correlation between aboveground biomass and NDVI can be adapted to different plants, environmental conditions, levels of weed intensity, and plant-level response to fertility.

The relationship of aboveground biomass and canopy area were best fit across years with a linear regression (Fig. 5). Some variability in estimated regression slope was observed between years with closer slope estimates for 2020 and 2022 than for 2021. In 2021, the plants were exposed to higher weed pressure and overall smaller at harvest than in 2020. In 2022, plants were also smaller due to later planting. The relationship between aboveground biomass and canopy area in 2021 was characterized with a shallower slope and showed slight improvements in model performance when aboveground biomass was log-transformed (linear, r² = 0.822; log, r² = 0.890). Features of plant size may be expected to follow scaling laws (Deng et al. 2012; Niklas 2004); for example, offering the opportunity to link remotely sensed canopy area to aboveground biomass. Growth in plant size occurs in multiple dimensions leading to nonlinear (i.e., exponential or allometric) patterns. Canopy area is a two-dimensional feature, whereas aboveground biomass is three-dimensional. Plants in the study were effectively estimated with a linear regression indicating that plants were relatively similar in size, although some signs of the nonlinear relationship were detected in 2021.

The linear regression models with NDVI were best fit when aboveground biomass was log-transformed, illustrative of the expected nonlinear scaling of aboveground biomass (Niklas 2004). Consistent with the relationship between aboveground biomass and canopy area, the highest correlation with NDVI was observed in 2020. However, for GNDVI, 2021 and 2022 were slightly higher correlations than 2020. Correlation of aboveground biomass with GNDVI was improved in 2020 and 2022 when aboveground biomass was modeled as linear and not log-transformed. These correlation estimates generated by the study appear strong, although additional calibration of observed features and predicted values must account for variability in hemp variety, field characteristics, and weed pressure.