Introduction
Water hyacinth, known as one of the world’s worst weeds, is arguably the most intensively managed invasive plant species in Florida with management costs exceeding US$3.6 million per year (FWC 2024; Hiatt et al. Reference Hiatt, Serbessoff-King, Lieurance, Gordon and Flory2009; Holm, Reference Holm1977; Langeland et al. Reference Langeland, Fishel and Gettys2014). Native to South America, this free-floating aquatic plant was introduced to North America as an ornamental in 1884 and quickly became problematic (Wunderlich Reference Wunderlich1962). Water hyacinth populations can double in size in as little as 6 d via vegetative reproduction of ramets (Degaga Reference Degaga2018). Ramets fragment from mother plants and spread readily through water currents, wind, boating activities, and intentional movement (Degaga Reference Degaga2018). Water hyacinth forms dense mats across the water surface that limit access and navigation, block and damage infrastructure such as bridges and flood control structures, provide habitat to disease vectors, decrease water quality, and reduce biodiversity (Holm, Reference Holm1977; Villamagna and Murphy Reference Villamagna and Murphy2010).
Since the 1970s, water hyacinth has primarily been managed proactively to keep population levels as low as possible by frequent (daily to weekly) deployment of boat-based applicators who search for and treat incipient plant populations with aquatic herbicides. Foliar applications of diquat and 2,4-D have been the commercial standard for water hyacinth management for decades; however, other herbicides such as carfentrazone, florpyrauxifen-benzyl, glyphosate, and penoxsulam can also provide control and are applied based on site-specific management needs (Enloe et al. Reference Enloe, Sperry, Leary and Ferrell2022; Gettys Reference Gettys2014; Mudge and Netherland Reference Mudge and Netherland2014b). Diquat is a fast-acting herbicide and highly effective across a wide range of conditions (Kyser et al. Reference Kyser, Madsen, Miskella and O’Brien2021; Wersal and Madsen Reference Wersal and Madsen2012). Auxin-mimic herbicides such as 2,4-D and florpyrauxifen-benzyl induce death by stimulating uncontrolled growth, with 2,4-D showing results in days, whereas florpyrauxifen-benzyl is slower to achieve the same level of control (Hildebrand, Reference Hildebrand1946; Mudge et al. Reference Mudge, Turnage and Netherland2021). Amino acid synthesis inhibitors such as glyphosate and penoxsulam result in slow symptom development that progresses over several weeks (Mudge and Netherland Reference Mudge and Netherland2014b; Wersal and Madsen Reference Wersal and Madsen2010).
Scientists and technicians who manage water hyacinth prefer to use fast-acting herbicides, including diquat and 2,4-D, for their quick, visible effects, and because treated areas can be easily identified within hours to a day after application (Mudge and Netherland Reference Mudge and Netherland2014a). However, this rapid damage can sometimes lead to public concern about herbicide use (Heinzman et al. Reference Heinzmann, Savchenko, Prince and Leary2024). Public alarm is lessened when slower-acting herbicides such as florpyrauxifen-benzyl, glyphosate, and penoxsulam are applied due to the inconspicuous symptoms the herbicides cause after treatment. Acetolactate synthesis (ALS) inhibitors are also generally more selective toward emergent native plants, which many resource managers find desirable (Mudge and Netherland Reference Mudge and Netherland2014b). However, ALS-inhibitor resistance is prominent among many terrestrial weed species, and some water hyacinth populations in Florida are suspected to have a reduced sensitivity to ALS inhibitors (Brown et al. Reference Brown, Prince and Sperry2024; Heap Reference Heap2014).
In large-scale herbicide treatments, efficacy can be variable due to plant growth stage, nondetected plants that do not receive treatment, environmental conditions, human error, or population susceptibility (Ganie et al. Reference Ganie, Kaur, Jha, Kumar and Jhala2018; Madsen Reference Madsen1999). This commonly leads to refuge plants remaining after treatment, thereby sustaining weed populations for regrowth and reinfestation (Cacho et al. Reference Cacho, Spring, Pheloung and Hester2006). To mitigate this, management efforts should include frequent surveillance to evaluate herbicide efficacy and determine follow-up treatments to prevent refuge populations from becoming large infestations. Herbicide efficacy evaluations are traditionally conducted through visual ratings based on phytotoxicity symptoms. Phytotoxicity refers to the symptomology that plants exhibit in response to herbicide injury, such as chlorosis and necrosis. Although these ratings are subjective, they can provide adequate accuracy and necessary numerical data for statistical analysis of herbicide efficacy by researchers. However, visual phytotoxicity assessments have their limitations under field conditions. A commonly used survey method for monitoring is the line–point intercept survey, which involves recording observations at equally spaced points along transects distributed throughout the water body (Madsen Reference Madsen1999). Some survey areas may be inaccessible by boat or be large enough that frequent monitoring is a significant drain on resources (Jakubauskas et al. Reference Jakubauskas, Peterson, Campbell, Campbell and Penny2002). The high growth rate and mobility of water hyacinth populations also contribute to the frequency of monitoring required, adding to the cost and resources allotted to management (Jakubauskas et al. Reference Jakubauskas, Peterson, Campbell, Campbell and Penny2002).
Remote sensing technology can be a critical tool for streamlining the monitoring process of herbicide efficacy, thus significantly reducing the cost, time, and resources required compared to reliance on traditional visual monitoring (Jakubauskas et al. Reference Jakubauskas, Peterson, Campbell, Campbell and Penny2002). While low-resolution satellite imagery (e.g., Sentinel 2; Landsat 8) has been used to map water hyacinth and predict injury, its spatial resolution is too low to map water hyacinth at the area coverages maintained by a proactive management regimen (Dube et al. Reference Dube, Mutanga, Sibanda, Bangamwabo and Shoko2017; Pádua et al. Reference Pádua, Antão-Geraldes, Sousa, Rodrigues, Oliveira, Santos, Miguens and Castro2022; Robles et al. Reference Robles, Madsen and Wersal2010; Rodríguez-Garlito et al. Reference Rodríguez-Garlito, Paz-Gallardo and Plaza2023).
As an alternative to satellite imagery, an unmanned aerial system (UAS) equipped with optical cameras and automated flight planning can quickly cover large areas and obtain high-resolution visually interpretable information (Cummings et al. Reference Cummings, McKee, Kulkarni and Markandey2017; Müllerová Reference Müllerová2019). Many natural-area managers use more affordable red, green, and blue (RGB) sensors and onboard navigation sensors to directly georeference the captured images to fit their practical needs (Dronova et al. Reference Dronova, Kislik, Dinh and Kelly2021; Kior et al. Reference Kior, Yudina, Zolin, Sukhov and Sukhova2024). Curran et al. (Reference Curran, Cox, Robinson, Robertson, Strom and Stahl2020) found that UAS-gathered surveys using onboard navigation systems were more spatially accurate, faster, and more efficient than manual line point-intercept surveys.
The RGB bands of an inexpensive digital camera mounted to a UAS can allow visualization of herbicide symptomology in plants (Kior et al. Reference Kior, Yudina, Zolin, Sukhov and Sukhova2024). Changes in plant physiology qualitatively change light spectra due to the absorption of light in the visible range by photosynthetic pigments, water, and the internal structures of leaves (Kior et al. Reference Kior, Yudina, Zolin, Sukhov and Sukhova2024). For example, herbicides that affect photosynthetic activity can result in changes in reflectance in the red spectral range, which can be detected by cameras (Kior et al. Reference Kior, Yudina, Zolin, Sukhov and Sukhova2024). Kior et al. (Reference Kior, Yudina, Zolin, Sukhov and Sukhova2024) reported that RGB spectral bands can estimate plant biomass and chlorophyll content with high efficiency. These bands can be used in various calculations to generate vegetation indices (VIs), which are designed to estimate key aspects of plant health. These indices have been shown to correlate with chlorophyll content, herbicide-induced injury, and biomass in previous studies (Abrantes et al. Reference Abrantes, Queiroz, Lucio, Mendes Júnior, Kuplich, Brendemeier and Merotto Júnior2021; Liu et al. Reference Liu, Hatou, Aihara, Kurose, Akiyama, Kohno, Lu and Omasa2021; Lussem et al. Reference Lussem, Bolten, Gnyp, Jasper and Bareth2018). While several studies in row cropping systems have been carried out to correlate VIs from inexpensive RGB cameras with plant health, to our knowledge, no studies have applied this methodology to monitor aquatic invasive plant management activities. Aerial monitoring of herbicide injury to aquatic invasive plants could significantly improve management efforts by reducing fieldwork demands and providing timely insights for making management decisions. Given the success of RGB VIs in assessing herbicide impact on terrestrial plants, we propose that water hyacinth injury can also be effectively monitored using this approach. Therefore, the objective of this study is to develop models for predicting herbicide efficacy on water hyacinth in response to six different herbicides using VIs derived from RGB imagery captured by a UAS.
Materials and Methods
Growth and Treatment Parameters
Experiments were conducted at the University of Florida’s Center for Aquatic and Invasive Plants in Gainesville, Florida (29.72°N, 82.42°W), during the spring and summer of 2023. Plants were grown in 151-L white, high-density polyethylene mesocosms with a 56-cm diam and a 71-cm depth, spaced approximately 1 m apart (Figure 1). Each mesocosm contained well water amended with 0.08 g L−1 of water-soluble fertilizer (24-8-16, Miracle-Gro® All Purpose Plant Food; Scotts Company, Marysville, OH) and 0.01 g L−1 of chelated iron (Grow More Iron Chelate 10%; Grow More, Gardena CA). Mature water hyacinth plants 23 to 30 cm tall sourced from Rodman Reservoir (29.52°N, 81.88°W) were transferred to experimental units (five plants per mesocosm) and left to establish for 1 mo prior to herbicide application, at which time each mesocosm had 100% plant cover. Fertility was monitored using an electrical conductivity meter (GroLine Waterproof EC/TDS Tester; Hanna Instruments, Smithfield, RI) and fertilized with the same amount of fertilizer each time to maintain electrical conductivity measurements of 0.04 mS cm−1. Insect pests were managed as needed using carbaryl (Sevin SL; Bayer CropScience LLC, St. Louis, MO) and bifenthrin (UP-Star Gold Insecticide, UPL, Cary, NC). During the first run, the average temperature was 22.5 C, and the average humidity was 74.5%, with weather conditions ranging from sunny to scattered clouds. In the second run, the average temperature was 27.2 C, and the average humidity was 81%, with weather conditions ranging from sunny to strong thunderstorms (NCEI 2023). Mesocosm water quality reflected similar parameters typical of a Florida eutrophic lake.
Figure 1. The study site is at the University of Florida Center for Aquatic and Invasive Plants, 6 wk after the spring treatment, at 30 m above ground level (0.76 cm/px). Three-panel gray scale reflectance target is pictured in the center of the study.
Each mesocosm was randomly assigned to receive a treatment, and the study had a factorial arrangement of treatments plus a nontreated control and four replications. Factors included herbicide active ingredient (2,4-D, diquat, carfentrazone, florpyrauxifen-benzyl, glyphosate, and penoxsulam) and rate (typical field use rate and maximum labeled rate) (Table 1). Herbicides were applied using a CO2-pressurized backpack sprayer equipped with two XR11004 nozzles (TeeJet Technologies, Glendale Heights, IL) spaced 18 inches (45.72 cm) apart to achieve an effective swatch width of 36 inches (91.44 cm), ensuring uniform spray coverage. The herbicides were selected to demonstrate a range of modes of action and symptom development profiles commonly used for water hyacinth management, with application rates reflecting both standard field rates and maximum label rates (Madsen et al. Reference Madsen, Owens and Getsinger1995; Mudge et al. Reference Mudge, Turnage and Netherland2021; Wersal and Madsen Reference Wersal and Madsen2010, Reference Wersal and Madsen2012). Nozzle size was chosen to accurately deliver 935 L ha−1 of solution at the applicator’s walking speed while minimizing off-target drift. Calibration was checked before and after treatment to ensure consistency throughout the treatment. The first run of the experiment was initiated on April 14, 2023 (spring), and the second run on July 6, 2023 (summer).
Table 1. Herbicide treatments and application rates for water hyacinth control in spring and summer studies.
a A nonionic surfactant (Induce; Helena Chemical Company, Collierville, TN) was applied at 2.5 mL L−1. Florpyrauxifen-benzyl was applied with a methylated seed oil concentrate (Leci-Tech; Loveland Products, Inc., Loveland, CO) at 10 mL L−1.
b Manufacturer locations: Alligare LLC, Opelika, AL; Bayer CropScience LLC, St. Louis, MO; Sepro Corporation, Carmel, IN; Syngenta Crop Protection LLC, Greensboro, NC.
Data Collection
Efficacy was visually estimated weekly by the same person for 6 wk after treatment (WAT). Visually evaluated efficacy was based on phytotoxicity: growth, stunting, and visible damage compared with nontreated control plants based on a scale from 0% to 100% where 0% = healthy unaffected plants and 100% = complete death. Corresponding images were captured at noon during cloud-free periods using a DJI Mavic 2 Pro quadcopter (DJI, Shenzhen, China) equipped with a Hasselblad L1D-20c RGB camera (Hasselblad, Gothenburg, Sweden) featuring a 20-megapixel CMOS optical sensor. If weather reports indicated cloudy conditions at noon, images were taken in the next closest cloud-free period to noon. A single image was designed to encompass the entire study region due to the small study area and low flight altitude. The study design was completely randomized to ensure that distortions around the edge of the image did not disproportionately affect any specific treatment group. The sensor was positioned at a nadir over the center of the entire experiment at an altitude of 30 m above ground level, producing a ground sampling distance of 0.76 cm px−1. The camera has a 77-degree field of view, an aperture range of f/2.8 to f/11, a focal length of 35 mm, and an ISO range of 100 to 3,200. Each captured image was 5,472 × 3,648 pixels.
Image Calibration
To standardize RGB values across images, mean pixel values for each RGB band were extracted from a PhotoVision 24-inch One-Shot Digital Calibration Target three-panel grayscale reflectance target (PhotoVision Inc, Mint Hill, NC) placed at the center of the site using the histogram tool in ImageJ (U.S. National Institutes of Health, Bethesda, MD). Color curves in GIMP (Kylander Reference Kylander1999) were then used to adjust the tonal range and color balance by mapping input RGB values to reference values from the target manufacturer, and this process was applied to each image to account for variations in lighting conditions.
Image Processing
Using QGIS software (QGIS Development Team Reference Development Team2024), circular polygons with an area of approximately 0.25 m2 were created to delineate each mesocosm, isolating vegetation from the background. The Zonal Statistics tool was then used to extract the mean RGB pixel values within each polygon, with digital numbers ranging from 0 to 255 (where 255 represents the highest intensity and 0 represents the absence of that color).
Image Analysis
The extracted RGB values were used to compute VIs in RStudio (RStudio Team 2024) based on equations in Table 2. Selected VIs were chosen based on their demonstrated correlations with herbicide efficacy or crop yield in previous studies (Abrantes et al. Reference Abrantes, Queiroz, Lucio, Mendes Júnior, Kuplich, Brendemeier and Merotto Júnior2021; Liu et al. Reference Liu, Hatou, Aihara, Kurose, Akiyama, Kohno, Lu and Omasa2021; Lussem et al. Reference Lussem, Bolten, Gnyp, Jasper and Bareth2018).
Table 2. Vegetation Index names, references, and corresponding equations.
a R,G, and B correspond to the red, green, and blue bands of an image.
Data Analysis
Data analysis was performed with RStudio software (v.4.4.2; RStudio Team 2024). The following R packages were used: DHARMa (Hartig Reference Hartig2017), ggplot2 (Wickham Reference Wickham2016), rstatix (Kassambara Reference Kassambara2019), tidyverse (Wickham et al. Reference Wickham, Averick, Bryan, Chang, McGowan, François and Yutani2019), and multcomp (Hothorn et al. Reference Hothorn, Bretz, Westfall, Heiberger, Schuetzenmeister, Scheibe and Hothorn2016). Analysis of variance detected no difference in the interactions among rate, season, and treatment; therefore, data were pooled across these parameters to reflect a variety of rates and timings at which water hyacinth may be treated. Data were filtered to the 3 wk when peak efficacy of each herbicide was demonstrated (1 to 3 WAT for diquat, 2,4-D, and carfentrazone; 2 to 4 WAT for florpyrauxifen-benzyl and glyphosate; and 4 to 6 WAT for penoxsulam) as determined by prior studies (Madsen et al. Reference Madsen, Owens and Getsinger1995; Mudge et al. Reference Mudge, Turnage and Netherland2021; Wersal and Madsen Reference Wersal and Madsen2010, Reference Wersal and Madsen2012). Nontreated control plant data were also paired with the treated data for the corresponding monitoring weeks. The VIs were correlated with visually evaluated efficacy using Spearman’s correlation coefficient due to its robustness to outliers and ability to handle ranked data. The best VI for each herbicide was chosen by selecting the VI with the highest correlation. Additionally, the VI with the highest correlation with visually evaluated efficacy when all herbicide data were combined was chosen for analysis to create a combined model. Data were then subjected to a linear regression using a random selection of 80% of the data with visual efficacy as the response and the best vegetation index as the independent variable. The linear relationship between the observed and predicted visual efficacy values was then evaluated using the remaining 20% of the data to ensure model robustness. The decision to use linear regression was based on an initial visual inspection of scatter plots showing a linear relationship between the variables, as well as supportive R 2 values from various vegetation index models indicating that linear models adequately captured the underlying relationship.
Results and Discussion
Vegetation Indices for Herbicide Visually Evaluated Efficacy
Correlations between the VIs and visually evaluated efficacy were strong and negative across various herbicides and for the combined models (ρ < 0.0001) (Table 3). The VI with the strongest correlation for each herbicide to predict efficacy was selected. However, many of the VIs demonstrated similar levels of correlation, suggesting that multiple indices may be similarly effective in predicting visually evaluated efficacy. The carotenoid reflectance index (CRI) was selected for 2,4-D, diquat, and florpyrauxifen-benzyl; the visibly atmospheric resistance index (VARI) was selected for carfentrazone and penoxsulam; and the excess greenness minus redness index (EXGR) was selected for glyphosate. Since visually evaluated efficacy showed the strongest correlation with EXGR when all treatment data were aggregated, this VI was chosen to create a combined model (Tables 3 and 4). All linear models had significant negative relationships between the VI and visually evaluated efficacy (Figure 2) with R 2 values ranging between 0.47 and 0.75.
Table 3. Spearman’s correlation coefficients between visually evaluated efficacy and vegetation indices by herbicide.a–d
a Abbreviations: CRI, modified carotenoid reflectance index; EXGI, excess green index; EXGR, excess greenness minus red index; GLI, green leaf index; MGRVI, modified green red vegetation index; MPRI, modified photochemical reflectance index; RGBVI, red-green-blue (RGB) vegetation index; TGI, triangular greenness index; VARI, visible atmospherically resistant index.
b Vegetation indices were calculated from the red, green, and blue bands of the image according to calculations listed in Table 2.
c
All correlations were significant with
$\rho$
< 0.0001.
d Bold values indicate highest correlation for that herbicide.
Table 4. Equations for predicting visually evaluated efficacy when water hyacinth (Eichhornia crassipes) is affected by herbicide. a
a Abbreviations: EXGR, excess greenness minus red index; R, G, and B correspond to digital numbers from the red, green, and blue bands of a digital camera; VE, visually evaluated efficacy; WAT, weeks after treatment.
Figure 2. Linear relationship between the highest correlated vegetation indices (Table 2) with visually evaluated efficacy when water hyacinth is affected by herbicide at 1 to 3 wk after treatment (WAT) for diquat, 2,4-D, and carfentrazone; 2 to 5 WAT for florpyrauxifen-benzyl and glyphosate; and 3 to 6 WAT for penoxsulam (n = 57).
The CRI demonstrated the highest correlations with visually evaluated efficacy for half of the treatments, indicating its robustness as a predictor of herbicide efficacy against water hyacinth. This VI was developed for nondestructive total carotenoid estimation in agricultural contexts from the principles that healthy vegetation has high reflectance in the green band (Gitelson et al. Reference Gitelson, Kaufman, Stark and Rundquist2002). Gitelson et al. (Reference Gitelson, Kaufman, Stark and Rundquist2002) found that reciprocal reflectance in the range 510 nm to 550 nm was linearly related to the total pigment content in leaves. Abrantes et al. (Reference Abrantes, Queiroz, Lucio, Mendes Júnior, Kuplich, Brendemeier and Merotto Júnior2021) adapted this VI for assessing herbicide injury to soybeans with an RGB camera and found CRI to have significant relationships with visually evaluated efficacy of herbicide treatments. Of the VIs tested, the CRI was the only index that did not include the red band as part of the calculation. Water hyacinth does not produce high levels of anthocyanins (red pigment) in response to injury, which is another reason why excluding the red band may have been beneficial. Newete (Reference Newete2014) similarly found that a VI calculated using green and green-blue wavelengths (the photochemical reflectance index), was significantly correlated with water hyacinth stress even though it was not as robust as VIs that included the near infrared band.
The VARI, which was developed to estimate green vegetation fraction in wheat canopies with minimal sensitivity to atmospheric effects (Gitelson et al., Reference Gitelson, Kaufman, Stark and Rundquist2002), is one of the most widely used VIs in agriculture within the visible spectrum (Xue and Su Reference Xue and Su2017). Rampazzo et al. (2022) found that VARI measurements complemented in-field estimates of soybean injury across various herbicide treatments. In the current study, VARI demonstrated the highest correlations with visually evaluated efficacy for water hyacinth treated with carfentrazone and penoxsulam. Despite their differences in mode of action and symptom development timelines, water hyacinth treated with these herbicides showed lower levels of maximum control compared to all other herbicides used in this study, which may have been why the same VIs had the best results for both treatments (Figures 2 and 3). While penoxsulam can cause progressive injury up to 10 wk after treatment (Wersal and Madsen Reference Wersal and Madsen2010), this study was limited to 6 wk. Additionally, carfentrazone has a history of inconsistent control of water hyacinth (Wersal and Madsen Reference Wersal and Madsen2012). The peak symptomology from both herbicides was exhibited as chlorosis compared to necrosis exhibited by the other herbicide treatments used in this study.
Figure 3. Linear relationship between predicted and observed visually evaluated efficacy values (Table 2) when water hyacinth is affected by herbicide treatments 1 to 3 wk after treatment (WAT) for diquat, 2,4-D, and carfentrazone; 2 to 4 WAT for florpyrauxifen-benzyl and glyphosate; and 3 to 6 WAT for penoxsulam (n = 15).
The excess greenness index was developed by Woebbecke and Von Bargen (Reference Woebbecke and Von Bargen1995) for separating green plants from soil and residue for image analysis and has been widely cited (Gitelson et al. Reference Gitelson, Kaufman, Stark and Rundquist2002; Lamm et al. Reference Lamm, Slaughter and Giles2002; Mao et al. Reference Mao, Wang and Wang2003). However, Meyer et al. (Reference Meyer, Hindman, Jones and Mortensen2004) noted that a disproportionate amount of redness from the background of the image may reduce the accuracy of this index, so Meyer and Neto (Reference Meyer and Neto2008) developed the EXGR index to minimize this problem. Abrantes et al. (Reference Abrantes, Queiroz, Lucio, Mendes Júnior, Kuplich, Brendemeier and Merotto Júnior2021) found that EXGR could satisfactorily estimate herbicide damage and soybean-estimated yield loss from dicamba and 2,4-D. In our study, we found that EXGR had the highest correlation with the visual efficacy of water hyacinth in response to glyphosate, as well as the highest correlation with the aggregated dataset (Figures 2 and 4). Glyphosate has been shown to reduce anthocyanin production, which could have resulted in a more prominent drop in red color, thus showing a high response to this index (Hoagland Reference Hoagland1980). Additionally, all herbicides lead to a reduction in greenness over time, which this VI effectively captures, likely explaining why it performed the best when applied to the aggregated data set.
Figure 4. Left: Linear relationship between the highest correlated vegetation index (Table 2) with visually evaluated efficacy when water hyacinth is affected by herbicide for the aggregated data (n = 342). Right: A linear relationship between predicted and observed visually evaluated efficacy for the vegetation index that had the highest correlation with visually evaluated efficacy for the aggregated data (n = 90).
Predicting Visually Evaluated Efficacy
A perfect model would have a slope of 1, R 2 of 1, and RMSE of 0 (Figure 3). While all linear relationships between predicted and observed visual efficacy values had moderate to high R 2 between 0.42 and 0.81, equations reliably predicted visually evaluated efficacy only in the medium ranges, but they poorly predicted visually evaluated efficacy in the extreme ranges (25% < x > 90%) (Figure 2). The upper extreme range corresponds to necrotic plants that are approaching complete control. As water hyacinth dies, the release of nutrients into the water may promote algal blooms, while the increased space makes room for other vegetation, such as duckweed, to colonize the mesocosms (Clugston Reference Clugston1963). This problem was exacerbated by the fast-acting herbicides, such as diquat, used in our study, which had already resulted in high levels of injury before the first data acquisition date. Contamination from algae and duckweed may have increased the greenness in these cases and skewed the VI values higher. Nontreated mesocosms represented the lower extreme of visually evaluated efficacy, with values less than 25%. Biomass production in the untreated mesocosms often presents a level of visual stress in these mature water hyacinths due to the natural senescence of older leaves that were not being accounted for with the visually evaluated efficacy observations. Additionally, the presence of flowers and various leaf angles may also have limited predictability of low injury (Robles et al. Reference Robles, Madsen and Wersal2010). Rampazzo et al. (Reference Rampazzo, De Lima E Silva, Moraes, Albert, Alves, Jakelaitis and Tejada2022) found that VI estimates of injury appeared to be less sensitive to differentiating low levels of injury than a trained observer. Some herbicide symptoms such as the curling, twisting, and callus formation caused by auxin herbicides may be visible to an observer before chlorosis-induced color changes can be observed in imagery.
This study demonstrates the feasibility of using a low-cost UAS equipped with a digital camera to estimate the visually evaluated efficacy of water hyacinth treated with six different herbicides. The method developed in this study could be modified to visually estimate the effects of herbicide treatments, but also other emergent and floating vegetation, and it has the potential to aid the development of a cost-effective tool for routinely monitoring water hyacinth chemical management. Open water present in the mesocosms was included in the vegetation index calculations to mimic field conditions, where more water would be exposed as a treatment progresses. However, water clarity and turbidity, which vary by water body and are likely to differ from mesocosm conditions, could make these VIs less reliable as treatments progress and more water is exposed. Therefore, future research should aim to translate this controlled study into field conditions to validate the practical application of these findings. Future analysis should also focus on using other spectral calibration methods, such as empirical line calibration. Efforts should focus on testing imagery with bands beyond the visible spectrum and automating the GIS processing workflow to reduce turnaround time for follow-up treatment planning.
Practical Implications
Remote sensing may improve the effectiveness of a proactive management program. Quadcopters equipped with digital cameras are inexpensive and easy to use by natural area managers, and regular aerial surveys could more quickly and efficiently capture large areas of interest than traditional monitoring methods. Vegetation indices such as the CRI, VARI, and EXGR are strongly correlated with visually evaluated efficacy of water hyacinth and can be easily calculated from these aerial surveys in GIS. These VIs may be able to aid an image analyst in differentiating healthy and injured plants. This information could improve herbicide treatment monitoring by detecting missed water hyacinth populations or ineffective treatments for planning follow-up herbicide applications. The use of UAS imagery and VIs offers a promising way to monitor herbicide treatments in water hyacinth management. By reducing the need for intensive field monitoring and improving detection of treatment efficacy, these methods can enhance invasive species management strategies.
Acknowledgments
Mike Durham contributed to the study design. J.P. Keller conducted herbicide treatments, applied insecticides, and fertilizer. Isaac Clark, Aiden Clark, and Alex Freeland helped set up mesocosms. Simon Reilly provided guidance on statistical analysis. Emma Matcham, Dr. Amr-Abd-Elrahman, and Hannah Brown provided support during the manuscript writing process.
Funding
Partial funding for this project was provided by the Florida Wildlife Conservation Invasive Plant Management Section.
Competing Interests
The authors declare they have no competing interests.
Open access
