Using High-Resolution Multispectral Data to Evaluate In-Season Cotton Growth Parameters and End-of-the-Season Cotton Fiber Yield and Quality

Abstract

Estimating cotton fiber quality early in the season, or its field variability, is impractical due to limitations in current methods, and it has not been widely explored. Similarly, few studies have tried estimating the parameters contributing to in-season cotton yield using UAV-based sensors. Thus, this study aims to explore the potential of using UAV-based multispectral images to estimate important in-season parameters, such as intercepted photosynthetically active radiation (IPAR), cotton height, the number of mainstem nodes, leaf area index (LAI), and end-of-the-season yield and cotton fiber quality parameters. Research trials were carried out in 2018 and 2020 in two experimental fields. In both years, a randomized complete block design was used with three cotton cultivars (2018), three plant growth regulators (2020), and three different irrigation levels to promote variability (both years). Cotton growth parameters were collected throughout the season on the same dates as UAV flights. Yield and fiber quality data were collected during harvest. The VI-based models used in this study were mostly sensitive to differences in cotton growth and final yield but less sensitive in detecting variation in cotton fiber quality indicators, such as length, strength, and micronaire, early in the season. The best performing regression model among the three fiber quality indicators was achieved in 2020, using a combination of four VIs, which explained 68% of the micronaire variability at 71 DAP. Results from this study also showed that multispectral-based VIs can be applied as early as the squaring stage at around 44 DAP to estimate most cotton growth indicators and final lint yield. Multiple linear regression validation models for height using NDVI, GNDVI, and RDVI obtained an R² of 0.62, and for LAI using MSR and NDVI an R² of 0.60. For lint yield, the best regression model combined four VIs and explained 66% of the yield variability. The ability to capture the variability in important growth and yield parameters early in the season can provide useful insights on potential crop performance and aid in in-season decisions.

1. Introduction

Cotton is the most important source of natural fiber in the world and has a yearly impact of approximately USD 600 billion in the global economy [1]. The United States (U.S.) is the third biggest cotton producer worldwide. According to the latest report from the U.S. Department of Agriculture (USDA) in the 2022/2023 season, the U.S. total area planted with cotton reached 5.5 million hectares resulting in a production of 3.15 million tons [2]. The estimated area planted with cotton for the country in 2024 is expected to increase by nearly 8%, which is projected to result in a total production of 3.48 million tons.

In 2023, the state of Georgia contributed one half of a million tons of the total U.S. cotton production [3], occupying the position of second largest producer in the country behind the state of Texas. The average yield for the state in the same year was 1100 kg/ha, which was the fourth largest yield achieved in the state [4]. In terms of productivity, Georgia ranked above states such as Texas and Oklahoma but had a lower average yield than Mississippi and Arkansas.

Cotton yield is highly affected by varying weather conditions and external factors, such as precipitation and irrigation, sunlight, and temperature [5,6]. Water availability is a key limiting factor for growth, yield, and fiber quality [7]. Decreased yield due to water stress can be a concern even in humid regions, such as that seen in the state of Georgia [8]. The sunlight duration and the amount of light intercepted by the canopy can greatly affect cotton growth and development, which directly impacts cotton yield and quality [5].

Monteith (1972) has defined yield as the product of three main components: harvest index, cumulative intercepted photosynthetically active radiation (IPAR), and radiation use efficiency [9]. Overall, all components directly or indirectly depend on the crop canopy’s ability to intercept light and convert that light into dry matter. One of the determining factors for IPAR, for instance, is the canopy area development represented by the leaf area index (LAI) [10]. Both LAI and IPAR play an important role in determining the cotton final yield. In a study conducted in 2021, Ermanis et al. evaluated the impact of different yield parameters on cotton yield under varied irrigation levels and found that IPAR had a positive impact on yield [11]. Similarly, IPAR was pointed out as one of the biggest contributors to yield loss in late-planted cotton [12].

The correlation between plant growth and final yield has also been explored through the evaluation of spatial variability in plant height. Cotton yield tends to increase with increased plant stature, showing a positive linear relationship with height [13]. Although, in some instances, wherein excessive cotton plant growth occurs, a decrease in yield can be seen [14]. In the same study from 2022, the spatial comparison between predicted height and cotton yield was assessed and showed that the spatial variability in plant height during mid-season can be a great indicator of high- and low-yield areas in a field. Nonetheless, results also indicated that, despite the linear regression model explaining the general relationship between cotton height and yield, spatial variability should not be ignored. Errors in yield estimation using plant height were significantly higher in areas with low soil electrical conductivity, high sand content, and low soil water holding capacity [13].

Major research efforts have been dedicated to improving cotton yield and understanding its main driving factors. Cotton fiber quality parameters, such as fiber length and strength, and micronaire, are also of great interest to growers and the cotton industry [15]. The fiber quality is extremely variable depending on environmental factors, such as sunlight, temperature, water, and nutrient availability [15,16]. Such factors influence the composition and structure of these fibers during development and can be a determining factor if the fiber quality will be below or above a desired standard. An important stage in fiber development is the transition stage, in which the cotton fiber stops elongating and starts to synthesize a secondary cell wall [17]. An early transition stage results in higher fiber strength and can be achieved with elevated accumulated heat units (growing degree days; GDD). Measuring fiber quality parameters is important in determining the final quality of cotton yarn [18]. Fiber length is a key factor in determining the yarn’s spinning efficiency, fineness, and strength.

In addition to environmental conditions, cotton fiber quality is highly influenced by genetics. The interaction between environment and genetics causes great variability in fiber quality in the field, between plants, within the same plant, and even, within the same cotton boll, presenting a great challenge in improving overall final yarn quality [19]. Most current methods for fiber quality measurement rely on the use of a high-volume instruments (HVIs) for fiber bundle classification and an advanced fiber information system (AFIS) for single fiber testing [20]. Although accurate, the current fiber quality testing system is time-consuming, labor-intensive, and limited to a small scale [21]. These tests are also restricted to indoor facilities, after the cotton boll is harvested.

The use of UAV-based remote sensors to monitor crop responses has significantly increased over the years to address the labor and scale limitations of ground data collection. High-resolution RGB images have been used to detect and estimate open bolls in the field for yield prediction at the end of the season [22], to estimate cotton canopy parameters, such as height and LAI [23], requiring a dense point cloud generated for multiple images taken at different angles and altitudes, and to predict single boll weight, while the cotton was still in the field [24]. Feng et al. [25] used a combination of RGB, multispectral, and thermal imagery to estimate yield in cotton at a field scale. Results from this study found that a combination of these three sensors accurately predicted cotton yield at the flowering stage and at 9 days before harvest. Despite the increased use of high-resolution UAV data for cotton monitoring, most studies have focused on yield predictions. However, the yield prediction was focused on prediction models developed at later stages [25] or at the end of the season [22]. Additionally, cotton in-season parameters, such as LAI and height, predicted on the 2022 study [23] required the development of dense point clouds from multiple images taken at different angles and altitudes.

Differently from yield, the estimation of other end-of-the-season parameters, such as cotton fiber quality early in the season or its field variability, is impractical due to limitations in current methods and has not been widely explored. Similarly, a lower number of studies have tried estimating in-season cotton yield contributing parameters using UAV-based sensors [26]. One study example that attempted to estimate these less explored parameters is Pokhrel et al. (2023), which used RGB and multispectral images to estimate IPAR, radiation use efficiency (RUE), and harvest index (HI) in response to different nitrogen application rates [26]. Although high accuracy prediction models were developed, the study focused on generalized models for the entire season. Xu et al. [21] also utilized the combination of RGB and multispectral images to estimate field variability in fiber quality and observed high performance in predicting micronaire, upper-half mean fiber length, and uniformity.

UAV-based remote sensors show great potential in enabling the detection and estimation of in-season cotton parameters that can indicate how the crop will perform in terms of yield and fiber quality. However, existing work has not focused on studying these relationships at specific growth stages, nor has it developed models that can be used early in the growing season. Thus, this study addresses this gap by using UAV-based multispectral images to estimate important in-season parameters, such as IPAR, cotton height, the number of mainstem nodes (points on a plant stem where branches, leaves, or buds originate), LAI, and end-of-the-season yield and cotton fiber quality parameters at different crop growth stages and using a simple single sensor approach. The specific objectives are (1) to select the most relevant vegetation indices (VIs) for estimation models for cotton growth, yield, and quality parameters and (2) to identify the earliest stage during the cotton growth season that VIs can be used to accurately estimate these parameters. In this study, it is hypothesized that (1) VIs based on the combination of red and NIR reflectance can be more effective in detecting differences in cotton growth parameters due to the striking difference in reflectance signatures between the two bands, and that (2) using VI-based models can enable the estimation of cotton growth and yield parameters early in the season.

2. Materials and Methods

2.1. Study Site and Treatments

The research trials were carried out in 2018 and 2020 at the University of Georgia’s Stripling Irrigation Research Park (SIRP) located in Camilla, GA (Figure 1). Both fields have Lucy loamy sand soil and follow a cotton/peanut/corn crop rotation system commonly used in the state. In 2018, a 1 ha experimental field planted on May 2nd was divided into 54 plots that were 4 rows wide and 12.2 m long. The study followed a randomized complete block design in which three different cotton cultivars; PHY 330, PHY 490 (PhytoGen Cottonseed, Indianapolis, IN, USA) and ST 6182 (BASF, Ludwigshafen, Germany) planted at a 2.5 cm depth, were randomly assigned within three different irrigation levels (dryland, 100% of crop evapotranspiration (ET_c), and 125% ET_c), and replicated 6 times. The irrigation system used in 2018 was a linear move with overhead sprinklers (Valley Irrigation, Valley, NE, USA), equipped with variable rate irrigation technology (VRI).

In 2020, a 3.6 ha field was planted on May 13th. The field was divided into 27 plots that were 12.2 m long and 4 rows wide. A randomized complete block design with three irrigation levels, three plant growth regulator (PGR) levels, and three replications was used. The field was under three different small center pivot irrigation systems with each pivot being considered a block. The irrigation treatments were the same as used in 2018 with a dryland, 100% ET_c, and 120% ET_c. Plant growth regulator treatments were applied using different Mepiquat chloride management strategies. PGR treatments consisted of a control (1), in which no PGR was applied; a moderate treatment (2), where 0.88 L.ha⁻¹ of a 4% solution of Mepiquat chloride (w/v) was applied at the first flower stage, and 1.17 L.ha⁻¹ at two weeks after the first application; and an aggressive treatment (3), where 0.73 L.ha⁻¹ was applied at the squaring growth stage, a second application of 0.88 L.ha⁻¹ was made at first flower, and a third application of 1.17 L.ha⁻¹ was made two weeks after first flower. Lastly, the cotton cultivar used was DP 1646 B2XF (Bayer Crop Science, Leverkusen, Germany), which is widely grown in Georgia and is the most widely grown cultivar in the state at the beginning of the experiment. This cultivar has a somewhat shorter stature during the early season, early maturity, and less vegetative growth than other cultivars [27].

2.1.1. Irrigation Scheduling

The irrigation systems used in both years were equipped with VRI technology, which enabled the application of the three different irrigation rates. The irrigation rates and frequency were defined using the SmartIrrigation Cotton App version 1.1.6 [28]. The SmartIrrigation App estimates the crop evapotranspiration (ET_c) on a daily basis from reference ET (ET₀) using an on-site meteorological weather station (University of Georgia Weather Station Network) installed at SIRP near the fields. The SmartIrrigation Cotton App calculates daily (K_c), which is derived from a K_c curve validated at SIRP for more than five years. The daily K_c is then multiplied by daily ET₀ to estimate ET_c. The deficit between daily ET_c and rainfall was used to drive recommendations for the well-watered (100% ET_c) plots. The irrigation for over-irrigated (125% ET_c) plots was calculated by multiplying the daily ET_c and precipitation deficit by 1.25. Standard irrigation rates per irrigation event were 19.05 mm in 2018 and 20.3 mm in 2020. The standard rate per irrigation event was defined based on the irrigation system capacity and soil properties to avoid runoff. In both seasons, a uniform irrigation rate was applied at the entire field, including the dryland plots, until the squaring stage to guarantee initial plant growth. After the squaring stage, dryland plots did not receive any supplemental water.

2.1.2. Weather Conditions

Daily minimum temperatures during the cotton growing season from May to September ranged from 12 to 24 °C and 8 to 25 °C in 2018 and 2020, respectively. Maximum temperatures in 2018 ranged from 26 to 36 °C and in 2020 from 23 to 37 °C during the same period. The relative humidity in both years remained above 55% during the growing season, reaching humidity up to 90% on multiple days. Precipitation was the most differing condition. In 2018, total precipitation from planting to harvest was 828 mm, while in 2020, it was 347 mm. Precipitation patterns during the two seasons were also different. In 2020, larger precipitation amounts were concentrated at the end of the season in August and September, while in 2018, precipitation was higher in the early and mid-season than later in the season. The overall precipitation difference between the two years resulted in higher total water applied via irrigation in 2020 when compared to 2018 [29].

2.2. Ground Data Collection

2.2.1. In-Season Measurements

Cotton growth parameters were collected multiple times during both seasons from the squaring stage until the last week of irrigation. In 2018, measurements were tentatively taken weekly (36, 44, 50, 65, 71, 86, and 113 days after planting; DAP), while in 2020, field data collection followed a two-week interval (35, 56, 71, 92, and 108 DAP) due to reduced personnel during COVID. Weeks with bad weather were skipped in both years. The average DAP for a normal cotton crop to reach specific growth stages is shown in

Table 1. Average number of days after planting for a normal cotton crop to reach different growth stages (modified from [30]).

Growth Stage	Days After Planting (DAP)
Emergence	5
First Square	38
First Flower	59
Open Boll	116
Harvest	140

. Plant height was measured for five representative plants in each plot using a ruler and averaged for analysis. The total number of mainstem nodes was also measured for five representative plants per plot, and average values were used as the final node count. This measurement was only taken in the 2018 growing season. Leaf area index (LAI) and the fraction of incident photosynthetically active radiation (IPAR_f) intercepted by the canopy were calculated from data collected using a AccuPAR LP-80 (Decagon Devices, Inc., Pullman, WA, USA) ceptometer. The ceptometer consists of two light probes (a small quantum sensor placed on a tripod 1.5 m above the surface and a light-elongated probe) connected to a datalogger that measures photosynthetically active radiation (PAR) above and below the crop canopy. In this study, two measurements were taken in each plot to calculate average LAI and IPAR_f values. The first measurement was taken with the light probe positioned parallel to the crop row, and the second measurement was taken with the light probe perpendicular to the rows (Figure 2). IPAR_f was calculated using Equation (1).

$${\mathrm{I}\mathrm{P}\mathrm{A}\mathrm{R}}_{\mathrm{f}}={\displaystyle \frac{(\mathrm{a}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{g}\mathrm{e}\mathrm{P}\mathrm{A}\mathrm{R}\mathrm{a}\mathrm{b}\mathrm{o}\mathrm{v}\mathrm{e}-\mathrm{a}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{g}\mathrm{e}\mathrm{P}\mathrm{A}\mathrm{R}\mathrm{b}\mathrm{e}\mathrm{l}\mathrm{o}\mathrm{w})}{\mathrm{a}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{g}\mathrm{e}\mathrm{P}\mathrm{A}\mathrm{R}\mathrm{b}\mathrm{e}\mathrm{l}\mathrm{o}\mathrm{w}}}$$

(1)

2.2.2. End-of-the-Season Measurements

In both seasons, cotton was defoliated when approximately 60% of the cotton bolls opened in the latest maturing plots. In both years, defoliant was applied 2 weeks before the cotton was harvested. Defoliant application helped promote leaf drop and opening mature bolls that were unopened. Cotton lint yield and lint and seed yield were estimated in 2018 and 2020, respectively, by harvesting the two center rows of each plot using a John Deere 9930 two-row spindle picker (Deere & Company, Moline, IL, USA). A hanging scale was used to weigh the seedcotton in the field right after harvest. Samples were subsequently sent to the University of Georgia MicroGin located in Tifton, Georgia. Lint weight was taken after the ginning process, and lint yield (in kg ha⁻¹) was estimated by multiplying the original seedcotton weight by the gin turnout in the harvested area [27]. Cotton lint quality parameters were obtained by sending a portion of the samples to the USDA classing office in Memphis, TN to be analyzed for fiber uniformity, strength, length, and micronaire.

2.3. Aerial Data Collection

Multispectral images were collected on the same days of ground measurements in both seasons using a 3DR Solo quadcopter (3D Robotics, Berkeley, CA, USA) equipped with a Parrot Sequoia multispectral sensor (Parrot SAS, Paris, France) (Figure 3a). This camera has four different narrow-band sensors, including green (530–570 nm), red (640–680 nm), red edge (730–740 nm), and near-infrared (NIR; 770–810 nm) with a pixel resolution of 1280 × 960. The Parrot Sequoia sensor was adapted to the quadcopter using a fixed mount and a power and data board that enabled images to be geotagged during the flight using the quadcopter GPS receiver.

Two different flight plans were created for the 2018 and 2020 growing seasons due to the difference in field sizes. In 2018, flights were performed at a height of 50 m from the surface, with a 4 m/s speed, and 80% front and side overlap resulting in a spatial resolution of approximately 5.6 cm. In 2020, flights were conducted at a higher height of 90 m from the surface and a speed of 9 m/s due to the larger field size. Front and side overlap were maintained at 80%. The spatial resolution of images acquired in 2020 was 8.5 cm.

Images of the camera’s radiometric calibration panel were taken immediately before and after each flight to ensure that calibration panel images had the same light conditions as the flight and that flight images could be properly calibrated during the stitching process (Figure 3b). The Pix4Dmapper software version 4.8.4 (Pix4D SA, Lausanne, Switzerland) was used to stitch images and perform geographic and radiometric calibrations. To calibrate the flight images, the pictures taken of the panel in each band were uploaded to Pix4D during the stitching process, and the known radiometric factor of the panel in each one of the four bands provided by the manufacturer were input. The software uses the radiometric factor to correct the panel reflectance value captured by the camera. This correction was then applied to all flight images. For geometric calibration, four ground control points (GCPs) with known geographic coordinates were placed at the corners of the fields during the flight for later calibration.

UAV Image Processing and Vegetation Indices (VIs) Calculation

Reflectance maps of all four bands obtained from the Pix4Dmapper stitching software were stacked using the QGIS 3.36.0 open-source software on each date. The stacked images were created using the “build virtual raster” tool, where the individual bands were used as inputs and the stacked bands as the output.

After using the virtual raster construction tool within QGIS, each bandset was imported into the open-access software R for further analysis. The initial step in R after the insertion and visualization of the band sets was to utilize the fieldIndex function from the FIELDimageR package [31] to calculate the normalized difference vegetation index (NDVI). The NDVI is calculated using the NIR and red bands and allows for better differentiation between vegetation and soil than using individual band reflectance. This is because the reflectance of NIR is higher for plants than for soil, and the absorbance of red is lower for plants compared to soil, resulting in higher NDVI values in the index for vegetation and lower values for soil [32]. Following this distinction, the fieldView and fieldsegment functions from the FIELDimageR.EXTRA package [33] were employed to create random samples of soil and plants, respectively, and to perform a supervised image classification using the random forest algorithm. Upon completion of the classification, the fieldMask function from the same package was utilized to create a mask of the soil class and remove it from the stacked image, thereby retaining only the crop pixels in the image.

Using the new image with the extracted soil, the fieldIndex function was applied once again to calculate sixteen different vegetation indices (VIs) (Table 2) for the prediction of the in-season growth parameters and end-of-season cotton fiber yield and quality. Many VIs were used to explore different combinations of mathematical equations using the crop reflectance in the four bands in an attempt to identify a certain band combination that will better capture variability of the specific parameters targeted in this study.

The selected indices have shown a correlation with plant biomass, chlorophyll content, and other pigments present in plant cells [34] and have demonstrated efficiency in predicting or correlating well with crop parameters in other crops. The red and NIR bands included in the VIs are well-associated with plant biomass, while the green and red-edge bands are more related to chlorophyll content [35,36].

After calculating the VIs, the plots mentioned earlier in Figure 1 were overlaid on the VI maps, and the mean VI values for each plot were extracted using the fieldInfo_extra function. The table generated by this function was exported as a .CSV file for further analysis.

Table 2. Selected VIs calculated from the multispectral images after the soil extraction process.

Vegetation Indices	Formula	Reference
CIRE (Chlorophyll Index Red Edge)	$\left({\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}}{\mathrm{R}\mathrm{E}}}\right)-1$	[37]
DVI (Difference Vegetation Index)	$\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{E}$	[38]
NDRE (Normalized Difference Red Edge Index)	$\left(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{E}\right)\times \left(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}\mathrm{E}\right)$	[39]
GNDVI (Green Normalized Difference Vegetation Index)	${\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{G}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{G}}}$	[40]
RDVI (Renormalized Difference Vegetation Index)	${\left({\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}}}\right)}^{0.5}$	[41]
NGRDI (Normalized Green/Red Difference Index)	${\displaystyle \frac{\mathrm{G}-\mathrm{R}}{\mathrm{G}+\mathrm{R}}}$	[42]
TCARI (Transformed Chlorophyll Absorption Reflectance Index)	$3\times \left(\mathrm{R}\mathrm{E}-\mathrm{R}\right)-0.2\times \left(\mathrm{R}\mathrm{E}-\mathrm{G}\right)-({\displaystyle \frac{\mathrm{R}\mathrm{E}}{\mathrm{R}}})$	[43]
NDVI (Normalized Difference Vegetation Index)	${\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}}}$	[44]
PSRI (Plant Senescence Reflectance Index)	${\displaystyle \frac{\mathrm{R}-\mathrm{G}}{\mathrm{R}\mathrm{E}}}$	[45]
TVI (Triangular Vegetation Index)	$0.5\times 120\times \left(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{G}\right)-200\times (\mathrm{R}-\mathrm{G})$	[46]
MSR (Modified Simple Ratio)	${\left({\displaystyle \frac{\left(\frac{\mathrm{N}\mathrm{I}\mathrm{R}}{\mathrm{R}-1}\right)}{\left(\frac{\mathrm{N}\mathrm{I}\mathrm{R}}{\mathrm{R}+1}\right)}}\right)}^{0.5}$	[47]
MTCI (MERIS * Terrestrial Chlorophyll Index)	${\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{E}}{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}}}$	[48]
MSAVI (Modified Soil Adjusted Vegetation Index)	${\displaystyle \frac{\left(2\mathrm{N}\mathrm{I}\mathrm{R}+1\right)-\sqrt{\left(2\mathrm{N}\mathrm{I}\mathrm{R}+1\right)2-8\left(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\right)}}{2}}$	[49]
SAVI (Soil Adjusted Vegetation Index)	$\left(1+0.5\right)\left({\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}+0.5}}\right)$	[50]
RVI (Ratio Vegetation Index)	${\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}}{\mathrm{R}}}$	[51]
OSAVI (Optimized Soil Adjusted Vegetation Index)	$\left(1+0.16\right)\times \left({\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}+0.16}}\right)$	[52]

* MERIS: Medium Resolution Imaging Spectrometer, NIR: Near Infrared; R: Red; RE: Red-Edge; G: Green, bands.

Table 2. Selected VIs calculated from the multispectral images after the soil extraction process.

Vegetation Indices	Formula	Reference
CIRE (Chlorophyll Index Red Edge)	$\left({\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}}{\mathrm{R}\mathrm{E}}}\right)-1$	[37]
DVI (Difference Vegetation Index)	$\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{E}$	[38]
NDRE (Normalized Difference Red Edge Index)	$\left(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{E}\right)\times \left(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}\mathrm{E}\right)$	[39]
GNDVI (Green Normalized Difference Vegetation Index)	${\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{G}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{G}}}$	[40]
RDVI (Renormalized Difference Vegetation Index)	${\left({\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}}}\right)}^{0.5}$	[41]
NGRDI (Normalized Green/Red Difference Index)	${\displaystyle \frac{\mathrm{G}-\mathrm{R}}{\mathrm{G}+\mathrm{R}}}$	[42]
TCARI (Transformed Chlorophyll Absorption Reflectance Index)	$3\times \left(\mathrm{R}\mathrm{E}-\mathrm{R}\right)-0.2\times \left(\mathrm{R}\mathrm{E}-\mathrm{G}\right)-({\displaystyle \frac{\mathrm{R}\mathrm{E}}{\mathrm{R}}})$	[43]
NDVI (Normalized Difference Vegetation Index)	${\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}}}$	[44]
PSRI (Plant Senescence Reflectance Index)	${\displaystyle \frac{\mathrm{R}-\mathrm{G}}{\mathrm{R}\mathrm{E}}}$	[45]
TVI (Triangular Vegetation Index)	$0.5\times 120\times \left(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{G}\right)-200\times (\mathrm{R}-\mathrm{G})$	[46]
MSR (Modified Simple Ratio)	${\left({\displaystyle \frac{\left(\frac{\mathrm{N}\mathrm{I}\mathrm{R}}{\mathrm{R}-1}\right)}{\left(\frac{\mathrm{N}\mathrm{I}\mathrm{R}}{\mathrm{R}+1}\right)}}\right)}^{0.5}$	[47]
MTCI (MERIS * Terrestrial Chlorophyll Index)	${\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{E}}{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}}}$	[48]
MSAVI (Modified Soil Adjusted Vegetation Index)	${\displaystyle \frac{\left(2\mathrm{N}\mathrm{I}\mathrm{R}+1\right)-\sqrt{\left(2\mathrm{N}\mathrm{I}\mathrm{R}+1\right)2-8\left(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\right)}}{2}}$	[49]
SAVI (Soil Adjusted Vegetation Index)	$\left(1+0.5\right)\left({\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}+0.5}}\right)$	[50]
RVI (Ratio Vegetation Index)	${\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}}{\mathrm{R}}}$	[51]
OSAVI (Optimized Soil Adjusted Vegetation Index)	$\left(1+0.16\right)\times \left({\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}+0.16}}\right)$	[52]

* MERIS: Medium Resolution Imaging Spectrometer, NIR: Near Infrared; R: Red; RE: Red-Edge; G: Green, bands.

2.4. Model Development

2.4.1. Spearman Correlation Rank

After VI value extraction, the data were merged into a spreadsheet and were subjected to descriptive analyses to identify the distribution and dispersion of the data. For these initial analyses, resources such as histograms and box plots were used aiding in the identification of outliers. Outliers were identified using the interquartile range (IQR) method and were corrected using the median of the values distributed in each quartile.

Following outlier correction, a Spearman correlation rank analysis (r_s) [53] using p < 0.05 was used as the first step to help determine the significance of the VIs in potential prediction models for the cotton in-season and end-of-the-season parameters. The Spearman correlation method was selected following the non-normal data distribution. Compared to Pearson correlation (r_p), Spearman correlation exhibits a lower standard deviation, which is perceived as a more accurate method for this kind of data distribution [54]. Furthermore, Spearman correlation tends to perform better than the other correlation models with a low number of samples and data.

2.4.2. Feature Selection

The feature importance selection was developed using the random forest algorithm, which is based on classification and regression trees [55,56]. This algorithm works by selecting a layer of features at a time, analyzing the possible combinations between all the features inserted as inputs and outputs, removing a specific portion of less important variables relative to the output variables, and generating a new decision tree [57]. In other words, random forest does not inherently remove any features during training. Instead, each tree is built using random subsets of features in each split. After training, random forest provides a feature importance ranking based on metrics mean decrease in impurity (Gini importance). This ranking enables a post-training feature selection step, where features can be ranked according to the degree of importance.

All the VIs were submitted to the feature importance selection process to further narrow down the list of predictive variables. The random forest model was implemented using scikit-learn’s RandomForestRegressor version 1.6.1 (https://scikit-learn.org). By employing the function’s default hyperparameter values, the model was trained and tested on each dataset. After the feature importance selection, only the VIs with the highest importance values were selected to be used and combined in linear plots analysis, varying between 3 and 5 VIs.

The linear correlation value plots were generated for each index over the course of degree days, analyzing the behavior of their relationship with each variable throughout the season. This analysis was utilized to select the best days after planting (DAP) and the best indices resulting in a combination to perform the regression models to predict the variable for an exact period of time during the season.

2.4.3. Regression Models

After selecting the best indices, the dataset was separated according to DAP within each year, and regression analyses using the chosen models commenced. The selected regression models were linear regression, in which a unique input parameter for each date in each variable was used, and multiple linear regression and second-degree polynomial regression, in which the three best VIs selected by the random forest feature importance algorithm were combined as the input parameters. The comparison among the three models aimed to evaluate whether a combination of multiple VIs would result in a better model fit and prediction outcome than using a single index in a simple linear regression when predicting the studied parameters. The dataset collected in each date in 2018 (n = 54) was divided into training (70%) and validation (30%). In 2020, the data were not split into training and validations sets due to the lower number of datapoints (n = 27). Regressions in 2020 were used to show the relationship between VIs and cotton parameters.

The regression model development and statistical models for data analyses were created on the Google Colaboratory platform using the Python 3.10 programming language, with regressions being developed using the scikit-learn library [58] and the results and data visualization graphs developed using the seaborn and matplotlib libraries [59,60].

2.5. Statistical Analysis

After developing the regression models, analyses of precision and model compatibility were conducted to compare the performance of the different models. These metrics are as follows: determination coefficient (R²) (Equation (2)), root mean square deviation (RMSE) (Equation (3)), and the Willmott’s agreement (WA) index (Equation (4)), an index that ranges from 0.0 to 1.0, indicating a complete disagreement of the model (0.0) and a perfect agreement between the model and the variables to be predicted (1.0) [61].

$${\mathrm{R}}^2={\displaystyle \frac{{\sum }_{\mathrm{t}=1}^{\mathrm{N}}(\mathrm{Y}\mathrm{e}\mathrm{s}\mathrm{t}-\overline{\mathrm{Y}})}{(\mathrm{Y}\mathrm{o}\mathrm{b}\mathrm{s}-\overline{\mathrm{Y}})}},$$

(2)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{{\sum }_{\mathrm{t}=1}^{\mathrm{N}}{\left(\mathrm{Y}\mathrm{o}\mathrm{b}\mathrm{s}-\mathrm{Y}\mathrm{e}\mathrm{s}\mathrm{t}\right)}^2}{\mathrm{N}}}},$$

(3)

$$\mathrm{d}=1-{\displaystyle \frac{{\sum }_{\mathrm{t}=1}^{\mathrm{N}}{(\mathrm{Y}\mathrm{o}\mathrm{b}\mathrm{s}-\mathrm{Y}\mathrm{e}\mathrm{s}\mathrm{t})}^2}{{\sum }_{\mathrm{t}=1}^{\mathrm{N}}\left(\left|\mathrm{Y}\mathrm{e}\mathrm{s}\mathrm{t}-\overline{\mathrm{Y}}\right|+\left|\mathrm{Y}\mathrm{o}\mathrm{b}\mathrm{s}-\overline{\mathrm{Y}}\right|\right)}},$$

(4)

where R² is the determination coefficient, RMSE is the root mean square of the average error, and d is the Willmott index coefficient. Yobs indicates the observed values, Yest indicates the estimated values by the models, N is the total number of data points analyzed, and $\overline{\mathrm{Y}}$ is the average value of the estimated variable.

3. Results

3.1. In-Season Growth Parameters

The Spearman correlations between the whole season average VI values and plant height, mainstem nodes, LAI, and IPAR_f averages are shown in

Table A1. Spearman correlation coefficients between whole season average VI values and the in-season growth parameters for the 2018 and 2020 cotton growing seasons.

	2018				2020
Vegetation Indices	Height	Mainstem Nodes	LAI	IPARf	Height	LAI	IPARf
NDVI	0.88 *	0.82 *	0.90 *	0.95 *	0.93 *	0.88 *	0.93 *
TVI	0.87 *	0.86 *	0.85 *	0.94 *	0.84 *	0.80 *	0.81 *
NGRDI	0.77 *	0.69 *	0.85 *	0.92 *	0.87 *	0.87 *	0.84 *
PSRI	−0.43 *	−0.38 *	−0.55 *	−0.62 *	−0.62 *	−0.60 *	−0.53 *
GNDVI	0.93 *	0.90 *	0.88 *	0.96 *	0.88 *	0.79 *	0.89 *
NDRE	0.67 *	0.69 *	0.73 *	0.77 *	0.73 *	0.63 *	0.75 *
CIRE	0.67 *	0.69 *	0.73 *	0.76 *	0.74 *	0.64 *	0.76 *
DVI	0.78 *	0.80 *	0.79 *	0.84 *	0.77 *	0.71 *	0.80 *
RDVI	0.89 *	0.87 *	0.87 *	0.96 *	0.87 *	0.84 *	0.87 *
MSR	0.88 *	0.82 *	0.91 *	0.97 *	0.92 *	0.89 *	0.93 *
MTCI	0.50 *	0.55 *	0.58 *	0.47 *	0.67 *	0.58 *	0.70 *
SAVI	0.87 *	0.88 *	0.83 *	0.92 *	−0.25 *	−0.12 NS	−0.09 NS
RVI	0.88 *	0.84 *	0.89 *	0.96 *	0.11 NS	0.28 *	0.34 *
MSAVI	0.88 *	0.89 *	0.84 *	0.93 *	−0.21 *	−0.09 NS	−0.06 NS
TCARI	−0.77 *	−0.77 *	−0.77 *	−0.94 *	−0.29 *	−0.43 *	−0.43 *
OSAVI	0.89 *	0.89 *	0.85 *	0.93 *	−0.18 NS	−0.07 NS	−0.03 NS

* Significant correlations; ^NS non-significant correlations.

, Appendix A. In 2018, significant correlations were observed between all VIs and cotton growth parameters. The lowest correlation coefficient (r) was observed between PSRI and the mainstem nodes with a correlation value of −0.38, while the highest correlation was observed between GNDVI and plant height with an r-value of 0.93. Other VIs, such as RDVI, MSR, OSAVI, NDVI, MSAVI, and RVI, all showed a strong positive correlation with plant height with r values above 0.87. The highest correlated VIs with the mainstem nodes were GNDVI with an r-value of 0.9, MSAVI and OSAVI with values of 0.89, and SAVI with an r-value of 0.88. NDVI had the highest correlation with LAI, followed by RVI, GNDVI, and RDVI with r-values of 0.89, 0.88, and 0.87, respectively. The highest correlations with IPAR_f were observed for MSR (r = 0.97) and GNDVI, RVI, and RDVI with r-values of 0.96, which were among the highest correlated VIs for LAI.

In 2020, correlations were not all significant. VIs, such as SAVI, RVI, MSAVI, and OSAVI, all had non-significant correlations with one or more of the growth parameters. NDVI showed the highest correlation with plant height and IPAR_f with an r-value of 0.93 for both parameters, while MSR exhibited the greatest correlation (r = 0.89) with LAI. The lowest significant correlations for plant height, LAI, and IPAR were observed for MSAVI (r = −0.21), RVI (r = 0.28), and RVI (r = 0.34), respectively.

3.1.1. Random Forest Feature Selection Algorithm

The results of the random forest feature selection algorithm used to rank the importance of the VIs in predicting each in-season growth parameter are shown in Figure 4 for 2018 and Figure 5 for 2020. The most important VIs for plant height in 2018 were GNDVI with an importance value above 0.5, followed by RDVI, TVI, NDVI, and MSR (Figure 4a), while in 2020, NDVI was the most important VI followed by MSR, NGRDI, GNDVI, and TCARI (Figure 5a). MSR and NDVI were also among the most important VIs for LAI predictions in both seasons. In 2018, MSR was ranked the most important followed by NDVI, GNDVI, and RDVI (Figure 4c). In 2020, NDVI and MSR were ranked the highest with importance values above 0.3. TCARI and GNDVI were the third and fourth VIs, but the importance value was significantly lower with values of 0.1 and below (Figure 5c).

The feature importance results for IPAR_f in 2018 showed that OSAVI was the most important VI with an importance value above 0.2, followed by RDVI and GNDVI with importance values above 0.1 (Figure 4b). In 2020, although the overall importance values were lower, the differences among them were smaller. CIRE was the most important VI, followed by DVI, MSR, GNDVI, and NDRE (Figure 5b). Mainstem nodes results were shown only for 2018, since the data were not collected in 2020. Multiple VIs were deemed important. Highest importance scores were observed for GNDVI, RDVI, NGRDI, MSR, OSAVI, and TVI (Figure 4d).

The random forest feature selection results were used to narrow down the list of VIs to only the most important for each in-season cotton growth parameter. The selected VIs are shown in

Table A2. Most significant VIs for each in-season cotton growth parameter for both growing seasons selected based on the random forest feature selection algorithm.

Variables	Year	Selected VIs
Height	2018	GNDVI, RDVI, TVI, NDVI
IPARf		OSAVI, RDVI, GNDVI, TVI
LAI		MSR, NDVI, GNDVI, RDVI
Nodes		GNDVI, RDVI, NGRDI, MSR, OSAVI
Height	2020	NDVI, MSR, NGRDI, GNDVI, TCARI
IPARf		CIRE, DVI, MSR, GNDVI, NDRE
LAI		NDVI, MSR, TCARI, GNDVI, NGRDI

. Subsequently, the Spearman correlation was used to indicate the correlation between the selected VIs and the growth parameters by DAP. Correlation coefficients for each VI at various periods during the season are shown in Figure 6 and Figure 7. The correlations by DAP were utilized to identify at which growth stages multispectral data can potentially be used to estimate different cotton growth parameters with higher accuracy.

Figure 6a shows the correlations between VIs and cotton height in 2018. The VIs showed the lowest correlation at the beginning of the season when plants are smaller. Correlations significantly increased at 44 DAP and reached its peak at approximately 50 DAP. All VIs showed a positive strong correlation with plant height on this date, with the highest correlation VI being observed for NDVI (r = 0.88). The correlations decreased after 50 DAP and slightly increased towards the end of the season at 113 DAP when cotton bolls were opening. A similar response was seen for LAI correlations (Figure 6c). However, although correlations were high at 50 DAP, the highest correlations were seen earlier at 44 DAP. The highest correlation coefficient was observed for MSR (r = 0.85). Correlations for IPAR_f followed a similar pattern during the season, with low correlations being observed at 36 DAP when plants were on the first square stage and peaked at 50 DAP (Figure 6b). On this date, SAVI showed the lowest significant correlation (r = 0.56), and RDVI and MSR were the most highly correlated with an r value of 0.82. The highest correlations with mainstem nodes were observed at two different stages during the season (Figure 6d). RDVI and MSR were the highest correlated VIs (r = 0.69) with mainstem nodes during early to mid-season at 50 DAP, while NGRDI and RDVI were the highest correlated (r = 0.76 and 0.74, respectively) at the end of the season at 113 DAP.

In 2020, three VIs consistently showed high correlations with cotton height, excluding during early season at 35 DAP, wherein all correlations were low (Figure 7a). At 56 DAP, when the plant was close to flowering, TCARI was the highest correlated with an r-value of −0.85 followed by NGRDI (r = 0.82). At 71 DAP, the highest correlations in the season were observed for NGRDI (r = 0.92) and PSRI (r = −0.92). NGRDI was again the highest correlated (r = 0.87) at 92 DAP, followed by TCARI and PSRI, both showing correlation coefficients of −0.78. Similar to 2018, correlations significantly increased at around 50 DAP. However, at the end of the 2020 season, correlations only showed a more significant decreased after 71 DAP. While NGRDI consistently showed high correlations for height, particularly during the flowering stage (71 and 92 DAP), for IPAR_f, MSR consistently showed the highest correlation throughout the season, with the highest r-value of 0.86 at 92 DAP (Figure 7b). MSR also showed high positive correlations with LAI (Figure 7c) at the early to late season (56, 71, and 92 DAP) with r-values of 0.76, 0.88, and 0.88. TCARI had the highest correlation coefficients among the VIs at 56 and 92 DAP, with the highest value of the season (r = −0.91) observed at 92 DAP.

3.1.2. Summary Statistics for Height, LAI, IPAR_f, and Mainstem Nodes

The summary statistics for the in-season cotton growth parameters are listed on

Table A3. Summary of range, mean, and standard deviation (SD) values for cotton in-season growth parameters by DAP in 2018 and 2020.

	2018				2020
Parameter	DAP	Range	Mean	SD	DAP	Range	Mean	SD
Height (cm)	36	26.8–43.4	36.9	3.55	35	22.8–36.0	28.0	3.32
	44	35.2–64.4	51.2	6.74	56	44.0–95.4	69.5	14.80
	50	25.1–62.7	47.3	9.51	71	56.2–128.2	90.6	20.13
	65	49.3–99.6	80.8	12.2	92	62.0–182.2	107.8	32.15
	71	67.1–125	101.2	12.3	108	-	-	-
	86	70.6–149.9	118.1	17.2
	113	78.2–175.3	129.4	24.5
LAI	36	0.22–0.90	0.51	0.15	35	0.08–0.67	0.36	0.19
	44	0.21–1.07	0.57	0.22	56	1.1–4.32	2.93	0.85
	50	0.43–2.6	1.03	0.40	71	1.77–7.03	4.49	1.33
	65	1.16–6.22	3.46	1.48	92	2.67–8.17	5.05	1.55
	71	2.00–7.62	5.38	1.44	108	2.17–6.74	4.24	1.42
	86	1.73–8.7	5.04	1.57
	113	1.87–5.67	4.09	0.78
IPARf	36	0.08–0.25	0.18	0.03	35	0.01–0.23	0.12	0.06
	44	0.13–0.50	0.27	0.10	56	0.37–0.94	0.79	0.13
	50	0.14–0.58	0.28	0.08	71	0.77–0.99	0.93	0.06
	65	0.37–0.91	0.76	0.13	92	0.87–1.00	0.96	0.04
	71	0.51–0.98	0.87	0.11	108	0.87–1.00	0.96	0.04
	86	0.61–0.99	0.93	0.08
	113	0.64–0.97	0.90	0.06
Nodes	36	6.00–8.40	7.14	0.45
	44	8.40–11.00	9.79	0.60
	50	8.40–12.00	10.6	0.77
	65	11.6–16.20	14.1	0.94
	71	12.4–15.80	14.6	0.63
	86	14.60–19.80	17.4	1.14
	113	17.00–21.40	18.6	1.18

in the Appendix A. Overall, variability in the data for all parameters in both years tended to increase at mid and late season compared to early season. The average plant height in 2018 was higher than in 2020. At first square, plants had an average height of 36.9 cm in 2018, while in 2020, the average was 28 cm. At peak flowering, before open bolls, plants in 2018 had an average height of 118.1 cm and 107.8 cm in 2020. However, higher variability in plant height was seen in 2020. LAI was also higher in 2018, with the highest LAI (5.38) being observed at 71 DAP, while the highest LAI in 2020 was observed only later in the season at 92 DAP. For a number of mainstem nodes, variability also increased as the season progressed, with a difference of up to five nodes between plots being observed.

3.1.3. Regression Models for Height, LAI, IPAR_f, and Mainstem Nodes

The stages in the season in which VIs showed the highest correlation with each in-season growth parameter were selected for further investigation. Different regression models were developed using varied combinations of VIs to estimate plant height, LAI, IPAR_f, and mainstem nodes (

Table 3. Performance metrics for the 2018 and 2020 in-season cotton growth parameters (height, leaf area index (LAI), and mainstem nodes) prediction models using different combinations of selected vegetation indices (VIs). Four models for the multiple and polynomial regression were developed starting with the five most important VIs combined, followed by the removal of the least important VI from each model until only two VIs were used. In this table, only the best performing multiple and linear regression models are shown.

Growth Parameter	Model Input Variables	Model	R2	RMSE	Willmott Agreement	R2	RMSE	Willmott Agreement
2018
			Training			Validation
Height (50 DAP)	NDVI	Linear	0.80	3.90	0.94	0.62	6.17	0.88
	NDVI, GNDVI, RDVI *	Multiple *	0.81	3.79	0.94	0.62	6.15	0.88
		Polynomial	0.83	3.90	0.95	0.60	5.99	0.81
IPARf (50 DAP)	RDVI *	Linear *	0.65	0.04	0.89	0.30	0.08	0.60
	RDVI, GNDVI, TVI, OSAVI	Multiple	0.78	0.03	0.93	0.21	0.08	0.63
		Polynomial	0.62	0.05	0.86	0.28	0.04	0.71
LAI (44 DAP)	MSR	Linear	0.65	0.10	0.89	0.52	0.17	0.83
	MSR, NDVI *	Multiple *	0.65	0.11	0.88	0.60	0.14	0.86
		Polynomial	0.67	0.11	0.89	0.61	0.18	0.77
Mainstem Nodes (50 DAP)	RDVI	Linear	0.38	0.49	0.74	0.19	0.45	0.69
	RDVI, MSR, NGRDI, GNDVI, OSAVI *	Multiple *	0.54	0.47	0.82	0.44	0.71	0.67
		Polynomial	0.67	0.4	0.89	0.19	0.84	0.41
2020
Height (71 DAP)	NGRDI	Linear	0.83	8.12	0.77	-	-	-
	NGRDI, NDVI, MSR, GNDVI, TCARI *	Multiple	0.88	6.7	0.96	-	-	-
		Polynomial *	0.96	3.9	0.98	-	-	-
LAI (71 DAP)	MSR	Linear	0.78	0.61	0.77	-	-	-
	MSR, TCARI, NDVI, NGRDI, GNDVI *	Multiple	0.81	0.57	0.77	-	-	-
		Polynomial *	0.87	0.47	0.83	-	-	-
IPARf (71 DAP)	MSR	Linear	0.63	0.04	0.59	-	-	-
	MSR, GNDVI, DVI, NDRE, CIRE *	Multiple	0.77	0.83	0.77	-	-	-
		Polynomial *	0.91	0.02	0.86	-	-	-

* Selected models based on the highest performance using the least number of input variables.

). Only the highest performing multiple and polynomial regression models are shown in the table.

The linear, multiple linear, and polynomial regressions developed for plant height at 50 DAP in 2018 showed very similar performance, with R² values of 0.8, 0.81, and 0.83, respectively, for training, and a lower performance for validation with linear, multiple, and polynomial models, explaining 62, 62, and 60% of the variability in height. Validation models also had higher errors of 6.17, 6.15, and 5.99 cm compared to training (3.9, 3.79, and 3.9 cm) for linear, multiple, and polynomial regressions, respectively.

Despite having the lowest agreement values for the validation set, the multiple linear regression using NDVI, GNDVI, and RDVI showed the lowest difference between model performance and error between training and validation sets with R² and WA values of 0.81 and 0.94 for training, 0.62 and 0.88 for validation (Figure 8a), and RMSE 3.79 and 6.15. Regression models for IPAR_f were developed for the same day as plant height. Despite the high R² and WA values seen for model training, all models’ performances significantly decreased for validation. The multiple linear regression using RDVI, GNDVI, TVI, and OSAVI was the best performing model for training, explaining 78% of the IPAR_f variation, with an error of 0.03 and an agreement of 0.93. However, the best performing model for validation was the simple linear regression using only the RDVI (R² = 0.30, RMSE = 0.08, and WA = 0.60).

Models for LAI were developed at 44 DAP. The best-performing model was the multiple linear regression using MSR and NDVI with R² values of 0.65 and 0.60, errors of 0.11 and 0.14, and WA of 0.88 and 0.86 for training and validation, respectively (Figure 8b). This model was selected due to its consistent performance metrics between the training and validation datasets. Models for mainstem nodes were developed for 50 DAP. Although correlations were higher at 113 DAP, estimating the number of nodes would be a more useful earlier in the season, since it is used in the quantification of plant growth and growth stages [62]. The multiple linear regression model using the five most important VIs showed the highest performance between training and validation, with a R² of 0.54, an error of 0.47, and a WA value of 0.82 for training and 0.44, 0.71, and 0.67 for validation, respectively.

The 2020 regression models were not subdivided into training and validation due to a lower number of data points. Nonetheless, models were still developed to show relationships for 2020. Models were developed for height, LAI, and IPAR_f at 71 DAP.

All regression models showed a strong relationship with height. The highest agreement was observed for the polynomial regression using the five highest correlated VIs. The combination of VIs was able to explain 96% of the height data variation and 98% based on the WA. The prediction error was 3.9 cm. The remaining models, using fewer input variables, still resulted in strong relationships with height. However, the accuracy of the models decreased, with the highest RMSE of 8.12 cm observed for the linear model. Similar results were seen for LAI and IPAR_f. The polynomial models using five VIs had the highest agreement (R² = 0.87 and WA = 0.83 for LAI and R² = 0.91 and WA = 0.86 for IPAR_f, respectively) and lowest errors (RMSE = 0.47 and 0.02, for LAI and IPAR_f, respectively).

3.2. End-of-the-Season Parameters

Correlations between VIs and cotton yield and quality parameters taken at the end of the season were developed and shown in Table A3, Appendix A. Overall, the whole season correlations for lint yield and micronaire in 2018 and 2020 and seed yield in 2020 were significant for the majority of VIs, but correlation values were low. In 2018, the highest significant correlations observed were −0.30 between MSR and lint yield and −0.27 between MSR and micronaire. In 2020, RVI was the highest correlated index with lint yield (r = 0.21) and seed yield (r = 0.34). TCARI was the highest correlated with micronaire (r = 0.55), fiber length (r = −0.30), and fiber strength (r = 0.23).

3.2.1. Random Forest Feature Selection Algorithm

The importance of each VI in predicting the end-of-the-season parameters for the 2018 (Figure 9) and 2020 (Figure 10) seasons was ranked using a random forest feature selection algorithm. The most important VIs for each end-of-the-season variable were analyzed at individual dates during the season. NDVI and TCARI were among the six most important indices for lint yield and all fiber quality parameters in 2018 (Figure 9). While NDRE and the soil-adjusted-related indices consistently appeared among the lowest important VIs.

NDVI was the most important index for lint yield (Figure 9c), NGRDI was the most important for fiber length (Figure 9a), while TCARI was the most important for fiber strength and micronaire (Figure 9b,d). In 2020, PSRI was again among the most important VI, together with TCARI, NGRDI, and RVI (Figure 10). TCARI was the most important index for micronaire (Figure 10c), lint (Figure 10d), and seed yield (Figure 10e), while NGRDI was the most important for fiber length (Figure 10a), and RVI for the fiber strength (Figure 10b). Indices calculated using reflectance on the visible range seemed to be more relevant for cotton fiber yield and quality. Indices using reflectance on the green and red bands showed higher importance than popular NIR-based VIs.

The most important VIs (

Table A5. Most significant VIs for end-of-the-season parameter for both growing seasons selected based on the random forest feature selection algorithm.

Variables	Year	Selected VIs
Fiber Length	2018	NGRDI, PSRI, MTCI, NDVI
Fiber Strength		TCARI, PSRI, NGRDI, NDVI
Lint Yield		NDVI, TCARI, RVI, NGRDI, MSR
Micronaire		NDVI, PSRI, DVI, NGRDI, MTCI
Fiber Length	2020	NGRDI, TCARI, RVI, PSRI
Fiber Strength		RVI, TCARI, NGRDI, PSRI
Lint Yield		TCARI, NGRDI, PSRI, DVI, RVI
Seed Yield		TCARI, RVI, NGRDI, PSRI, DVI
Micronaire		TCARI, NGRDI, RVI, PSRI

) in both years were selected for an in-season correlation analysis by DAP (Figure 11 and Figure 12). Results showed that, in both years, the VIs collected throughout the season did not have high correlations with fiber length and strength. Correlation coefficients were all lower than 0.4 in 2018 for negative and positive correlations (Figure 11a,b), while in 2020, only one date showed a slightly higher correlation (r = 0.53) for fiber strength (Figure 12b). The highest correlation for micronaire was observed at the end of the season at 113 DAP for TVI with an r-value of −0.64 (Figure 11c). However, earlier in the season, TCARI and MSR showed cosisntently higher correlations. The correlation between micronaire and TCARI was highest at 71 DAP with an r-value of 0.63, while the highest correlation with MSR was seen at 36 DAP with an r-value of 0.62. However, in 2020, multiple VIs showed high positive correlations with micronaire in varies dates. Starting at 71 DAP, all four selected VIs showed to be highly correlated with micronaire (Figure 12c). Correlations increased as the season progressed, reaching the highest r-value of 0.82 observed at 92 DAP for TCARI.

In 2018, the VIs showed the highest correlations with lint yield, among all parameters. Various VIs started showing high correlations with lint yield as early as 44 DAP and remained relatively high for the majority of the season, until it reached peak values at 113 DAP (Figure 11d). MSR had the highest correlation with lint yield at 44 DAP with an r-value of 0.71, while an r value of 0.73 was observed between lint yield and RVI at 113 DAP.

3.2.2. Summary Statistics for Lint Yield, Seed Yield, and Micronaire

The summary statistics for the cotton yield and fiber quality parameters are shown in

Table A6. Summary of range, mean, and standard deviation (SD) values for cotton end-of-the-season yield and fiber quality parameters in 2018 and 2020.

	2018			2020
Parameter	Range	Mean	SD	Range	Mean	SD
Fiber Length (cm)	1.05–1.16	1.12	0.02	1.17–1.33	1.25	0.04
Fiber Strength (g tex−1)	25.1–34.5	30.8	2.03	28.7–32.5	30.3	0.90
Micronaire	3.2–5.1	4.09	0.45	3.5–4.9	4.31	0.35
Lint Yield (kg.ha−1)	314.29–2091.21	997.30	349.82	987.71–1742.95	1382.38	233.01
Seed Yield (kg.ha−1)	-	-	-	2562.22–4209.37	3492.37	502.48

in Appendix A. For end-of-the-season parameters, higher variability was observed in 2018 for fiber strength, micronaire, and lint yield. Average fiber strength in 2018 (30.8) was higher than 2020 (30.3), and the data showed higher variability for 2018 (SD = 2.3) than in 2020 (SD = 0.90). For lint yield average values for 2018 and 2020 were 997.3 kg.ha⁻¹ and 1382.38 kg.ha⁻¹, respectively, while the SD was 349.82 for 2018 and 233.01 for 2020. For micronaire, higher average was observed in 2020 (4.31) than in 2018 (4.09); however, the SD for 2018 was 0.45, while it was 0.35 in 2020. Fiber length had a slightly higher variability with a standard deviation (SD) of 0.04 compared to 0.02 for 2018. The mean fiber length was 1.12 and 1.25 for 2018 and 2020, respectively. Seed yield was only collected in 2020 and had an average value of 3492.37 kg.ha⁻¹.

3.2.3. Regression Models for Lint Yield, Seed Yield, and Micronaire

In 2018, regression models were developed for only lint yield and micronaire due to weak correlation values between VIs and remaining end-of-the-season cotton quality parameters. The highest correlations for lint yield were observed at 44 DAP, 65 DAP, and 113 DAP. The earliest date in the season, at 44 DAP, was selected to develop regression models (

Table 4. Performance metrics for the 2018 and 2020 end-of-the-season cotton lint yield prediction models using different combinations of selected vegetation indices (VIs). Four models for the multiple and polynomial regression were developed starting with the five most important VIs combined, followed by the removal of the least important VI from each model until only two VIs were used. In this table, only the best performing multiple and linear regression models are shown.

Parameter	Model Input Variables	Model	R2	RMSE	Willmott Agreement	R2	RMSE	Willmott Agreement
2018
			Training			Validation
Lint Yield (44 DAP)	MSR	Linear	0.41	271.23	0.76	0.22	244.29	0.73
	MSR, NDVI, TCARI, NGRDI, RVI	Multiple	0.50	250.16	0.81	0.19	248.18	0.76
		Polynomial	0.58	236.48	0.84	0.67	174.68	0.90
	MSR, NDVI, TCARI, NGRDI *	Multiple	0.49	250.84	0.81	0.17	251.70	0.76
		Polynomial *	0.58	235.77	0.84	0.66	179.10	0.89
Micronaire (86 DAP)	TCARI	Linear	0.38	0.35	0.7	0.2	0.37	0.68
	TCARI, MSR	Multiple	0.45	0.32	0.7	0.33	0.41	0.73
		Polynomial *	0.50	0.31	0.81	0.30	0.41	0.66
2020
Lint Yield (108 DAP)	DVI	Linear	0.43	171.24	0.54	-	-	-
	DVI, RVI, NGRDI, PSRI, TCARI *	Multiple	0.51	159.91	0.57	-	-	-
		Polynomial *	0.76	111.32	0.92	-	-	-
Seed Yield (108 DAP)	DVI	Linear	0.55	327.33	0.60	-	-	-
	DVI, RVI, NGRDI, TCARI, PSRI *	Multiple	0.63	295.87	0.67	-	-	-
		Polynomial *	0.73	253.6	0.91	-	-	-
Micronaire (71 DAP)	RVI	Linear	0.56	0.22	0.65	-	-	-
	RVI, NGRDI, TCARI, PSRI *	Multiple	0.60	0.21	0.66	-	-	-
		Polynomial *	0.68	0.19	0.89	-	-	-

* Selected models based on the highest performance using the least number of input variables.

). Polynomial regressions outperformed simple and multiple linear regressions, irrespective of the number of input variables. Linear and multiple regressions showed a significant decrease in performance in the validation with lower R², higher RMSE, and lower WA for almost all models tested. The highest performance model was the polynomial regression using MSR, NDVI, TCARI, and NGRDI, which explained 58% of the lint yield variation, had an error of 235.77 kg/ha, and had a WA value of 0.84 for the training model (Figure 13). While the higher R² of 0.66 during validation may indicate a possible overfitting, the prediction error decreased to 179.1 kg/ha. For micronaire, the highest correlations were observed at 113 DAP followed by correlations at 86 DAP. However, 86 DAP was selected for the development of regression models, as it is earlier in the season, and differences in the correlation coefficients were of only 0.01 and 0.03 between the highest and second highest correlated VIs for both dates. The results showed that including the second highest correlated VI to multiple and polynomial regressions helped increase performance when comparing to the simple linear regression. The highest performing model was the polynomial regression using TCARI and MSR, which explained 50% of the variability for the training model and 30% of the validation. The RMSE and WA values observed were 0.31 and 0.41 and 0.81 and 0.66 for training and validation, respectively.

Cotton lint and seed yield and micronaire regression models were developed for 2020 (Table 4). Regression models for lint and seed yield were developed at 108 DAP due to lower correlations during the early and mid-season, while for micronaire, VIs showed a high correlation at 71 DAP. The polynomial regressions with the DVI, RVI, NGRDI, PSRI, and TCARI as input variables were the best performing models for all three parameters, explaining 76, 73, and 68% of the variation in lint yield, seed yield, and micronaire, respectively. The prediction error indicated by the RMSE value for lint yield was 111.32 kg/ha, 253.6 kg/ha for seed yield, and 0.19 for micronaire. The agreement values based on the WA were 0.92, 0.91, and 0.89, for lint and seed yield and micronaire, respectively.

4. Discussion

The primary goal of this research was to evaluate the use of UAV-based multispectral images in estimating both in-season and end-of-the-season cotton parameters and identify the best stages during the season to estimate each parameter. Multispectral data have been shown to accurately depict green vegetation responses, as well as detecting differences in plant reflectance due to variability in growth stages in the NIR band [63]. Several VIs combining bands in the visible and NIR ranges that have been reported in the literature to correlate well with general crop growth and biomass were calculated in this study, and their correlations with cotton growth, yield, and quality parameters were evaluated.

4.1. In-Season Growth Parameters Estimation

The approach used in this study for VI selection involved an initial whole-season average values correlation analysis using Spearman correlations and an importance rank analysis using a random forest selection algorithm. Only selected VIs were then used to develop stage-specific correlations to evaluate the best time in the season to estimate each parameter. Results from this 2-year study showed that most VIs had high importance and high season-long correlations with the in-season cotton growth and development parameters, revealing great potential in enabling the investigation of their spatial variability. The majority of VIs calculated in this study used a combination of NIR and visible bands. Some of these VIs were repeatably among the most important indices for cotton growth indicators in both years, such as NDVI, GNDVI, CIRE, and MSR. The NIR and red bands used in NDVI, CIRE, and MSR have been widely studied for cotton growth parameters, such as LAI, biomass, nitrogen assessment, yield, and lint yield [64,65,66,67]. The red and NIR bands have been shown to be directly related with foliar pigments and photosynthetic activity [68,69]. The foliar pigments, especially chlorophyll, at high concentrations in the leaves presented a high peak of absorbance in the red band in the wavelengths 660 and 680 nanometers [36]. Differently from the red band, plants present a low peak of absorbance in the NIR band [36,68]. This difference occurs due to the interaction of the water content with the light in the mesophyll and parenchyma cells [36]. The stomatal opening regulates the gas exchange, water absorption, and metabolic production by the photosynthetic routes [36,69]. This photosynthetic production increases the chlorophyll pigments and biomass production, which are strongly correlated with the growth parameters, leading to the high correlation values observed in this study.

The correlations among selected VIs and plant height, LAI, IPAR_f, and mainstem nodes at different DAP showed that multispectral data only start to detect differences in these parameters after the crop reaches at least 44 DAP, which is when cotton plants are at the squaring stage [30]. These results show the possibility of estimating cotton growth at even earlier dates when compared with previous studies in the southeast region that reported high correlations with cotton height starting at 60 DAP [14,70]. This indicates that the correlation between these indices and cotton growth are dependent on the weather conditions of each season. In 2018, correlations between the indices and cotton growth parameters fluctuated between 44 and 113 DAP, which were the dates with the two major peak correlations during the season. On the other hand, in 2020, correlations had a more consistent increase, until they reached a plateau during the mid to late season at 92 DAP. In 2018, there was excessive precipitation at the beginning of the season and during other important stages, while in 2020, more sunny days with less precipitation led to a more rapid accumulation of heat units and gradual response, despite the crop being smaller at first square. The difference in weather conditions impacted canopy growth and reflectance correlations between the years.

Both in the whole season and growth stage, correlations between VIs and IPAR_f were high. However, the 2018 regression models developed with the highest correlated VIs at 50 DAP did not have a consistent performance for training and validation. The coefficient of determination for training explained 62 to 78% of the IPAR_f variation, while this percentage was much lower for validation, varying from 0.14 to 0.30%. The highest value for training was observed for a multiple linear regression model using RDVI, GNDVI, TVI, and OSAVI as the input variable, while the linear regression model using only RDVI had the strongest relationship for validation. In 2020, training models for IPAR_f were developed later in the season at 71 DAP. All three regression models had a higher precision than in 2018, with R² values of 0.63, 0.77, and 0.91 for linear, multiple, and polynomial regressions, respectively. Correlations between IPAR_f and VIs were already increasing at 56 DAP, but they had not reached peak correlation yet. IPAR_f prediction models using similar VIs for the same region showed higher performance than the models developed in this study. Models based on RVI, red edge chlorophyll index (RECI), NDRE, and simplified canopy chlorophyll content index (SCCCI) explained 89 to 93% of the variation in IPAR_f [26]. The superior performance can be attributed to a few factors. Authors in this study incorporated GDD in the regression models, had a higher number of data points available, and a more significant difference in VI and IPAR values, since the models were generalized for the entire season and not by DAP, as in the current study. These factors potentially helped the development of more robust models, once predicting crop parameters within specific dates is more challenging than predicting seasonal differences.

When comparing the 2018 linear, multiple, and polynomial regression models using the highest-correlated VIs at 50 DAP for height and 44 DAP for LAI, polynomial and multiple regression models tended to explain more of the variation in these two parameters and have a lower error, even if these differences were small. The multiple linear regression using NDVI, GNDVI, and RDVI was the selected model for plant height due to the consistent performance showing high R² and WA and low RMSE between the training and validation sets. However, based on the results, any of the other models could result in a good prediction model. Similarly, models for LAI prediction had similar performance when comparing all the metrics, but the multiple linear regression using MSR and NDVI was the selected model due to a combination of high accuracy for training and validation using the least number of input variables. In 2020, the regression models were developed later in the season at 71 DAP for all three parameters. For training models, the linear, multiple, and polynomial models explained 83 to 0.96, 0.78 to 0.87, and 0.63 to 0.91 of the variability in height, LAI, and IPAR_f, respectively. The validation of these models was not possible due to insufficient data. However, these results show the relationship between these VIs and the growth parameters and the effects of including additional VIs in the models’ performances.

The results highlight that correlations between VIs and cotton growth and the best timing to predict growth parameters are highly dependent on the conditions of each season. In 2018, the crop showed higher vigor at the beginning of squaring stage (around 36 DAP) than in 2020; however, in 2020, the crop grew faster and achieved higher height, LAI, and IPAR_f earlier in the season. Irrespective of other external factors that may affect the performance of prediction and monitoring of cotton growth, the prediction models’ performance may be increased with the use of more complex models, such as extreme learning machine (ELM), which has shown an increased ability to explain variation in LAI of up to 90% during validation [64].

Models for the mainstem nodes showed an overall lower performance when compared to plant height and LAI. The precision and accuracy of models significantly decreased during validation, which can be attributed to the limited number of data points. However, it has been reported that the mainstem node is mostly related to the simple ratio (SR) index, which uses a simple equation dividing the NIR reflectance by the red reflectance [71].

4.2. End-of-the-Season Yield and Quality Parameters Estimation

The estimation of end-of-the-season fiber quality parameters using in-season crop reflectance is slightly more challenging, as reflected by this study’s results. Low correlations for the end-of-the-season parameters and VIs were more frequently observed when compared to in-season growth parameters. One potential explanation is the variety of factors that are involved in determining fiber quality parameters during the cotton developmental stages. Although the canopy architecture and leaf shape regulate the photosynthetic rate and play a big role in affecting micronaire and fiber length and strength [15,72], final cotton fiber quality and yield are affected by the combination of genetics [73], environment [74], and management practices, such as irrigation [75] and fertilization [76], which may not always be picked up by VIs. This makes it more difficult to draw direct correlations between VIs and fiber quality than VIs and plant biomass and growth. Despite the complex combination of factors driving cotton fiber quality parameters, it has been observed that poor growing conditions and high fruit number are associated with low micronaire [77]. It has been reported in the literature that, in some circumstances, reflectance in the 350–920 nm and 1400–2500 nm ranges shows sensitivity to predict fiber length, strength, and micronaire during the boll opening period [21].

In the current study, bands in the longer wavelength range were not available, but among the VIs that used bands in the 350–920 nm range, the NGRDI, TCARI, NDVI, and RVI were ranked the most important for fiber length, strength, lint and seed yield, and micronaire. Similar to the in-season parameters, the most important VIs were further analyzed, and correlations were developed for different growth stages. When comparing both years, VIs were more sensitive to the differences in fiber quality and yield parameters throughout the 2020 season than in 2018. In 2018, the highest correlations were seen for lint yield followed by micronaire. For lint yield, TCARI had a high correlation at 36 DAP, while the remaining VIs only showed higher correlations at 44 DAP. Correlations stayed high throughout the squaring and flowering stages until 71 DAP, after which correlations decreased and increased again at the end of the season at 113 DAP. For micronaire, correlations with TCARI showed a steady increase, reaching their highest at 71 and 86 DAP (r = 0.63 and 0.62, respectively). MSR showed moderate-to-high correlations throughout the season; however, the highest r-value was observed at 36 DAP. In 2020, the lower precipitation allowed for more rapid plant growth to occur due to the different irrigation treatments compared to 2018, which received excessive rainfall at early growth stages. This higher variability captured by the VIs was translated into the correlations with cotton quality and yield parameters during the season. Correlations with fiber length and strength were slightly higher during most stages. The same was observed for micronaire, in which TCARI and RVI showed increased sensitivity to micronaire starting close to 56 DAP, which coincides with previous results [21].

Prediction models developed for lint yield at 44 DAP in 2018 showed that a polynomial regression using MSR, NDVI, TCARI, and NGRDI resulted in the best prediction model. Models using a lower number of VIs or a single VI resulted in poor accuracy and precision. For micronaire, the models were developed at 86 DAP. A polynomial regression with TCARI and MSR had the best performance. In 2020, models for lint and seed yield were developed later in the season at 108 DAP and at mid-season (71 DAP) for micronaire. When comparing performances between the two growing seasons, it is important to consider that both years had very distinct weather conditions. Temperature, water availability, and sunlight during the initial growth stages highly influence fiber composition, length, and strength [15,17]. These factors also drive crop growth rates and affect correlations with VIs, the predictive ability of these models, and the timing for more accurate predictions.

Overall, the performance for models developed in this work is lower than other studies, such as the one conducted by Haghverdi et al. [78], in which machine learning lint yield prediction models were developed from Landsat 8-based VIs, achieving a higher precision with R² values of 0.69 and 0.74. Another recent study used a random forest model and had an average R² of 0.77 and an average accuracy of 7.5% [79]. However, it is important to note that these yield prediction studies used data from later in the season (close to harvest), while the goal of the current study is to estimate yield in earlier stages using growth parameters as an indicative of yield potential.

4.3. Limitations

This study was conducted in two small experimental fields inside one of the University of Georgia’s research centers during two growing seasons. Despite the use of different cultivars and irrigation levels, the low soil variability coupled with a very rainy season in 2018 limited the range of growth and quality parameter values. The multispectral-based VIs were not sensitive enough to the low variability achieved in some of these parameters. The multispectral camera used in the study also had a limited spectral resolution of only four bands. Expanding the data collection area to include more soil variability may result in increased variation in cotton growth and quality parameters. Moreover, the use of newer camera systems with a higher spectral resolution may increase the precision and accuracy of the prediction models.

5. Conclusions

This study aimed to evaluate the potential of using time series UAV-based multispectral data to monitor and estimate in-season cotton growth parameters and end-of-the-season lint yield and quality. The VIs calculated in this study were mostly sensitive to differences in cotton growth and final yield, showed moderate sensitivity to micronaire, but were less sensitive in detecting variation in cotton fiber quality indicators, such as length, and strength. The highest correlated VIs selected using a random forest feature selection were selected for each of the observed parameters, and the correlation at specific growth stages evaluated. The results from this study showed that multispectral-based VIs can be applied as early as 44 DAP to estimate most cotton growth indicators and final lint yield. Overall, the regression models tested for these parameters showed a fair performance, but including a combination of two or more VIs on multiple or polynomial regressions seemed beneficial to improve the precision and accuracy of the prediction models, such as for height, LAI, and lint yield. Based on these results, it is fair to conclude that VIs that use a combination of visible and NIR bands can be used with certain accuracy to monitor in-season cotton growth and yield as early as the crop reaching the squaring stage. The development of predicted maps can enable the evaluation of spatial variability in crop growth and potential yield. Nonetheless, the correlations between remotely sensed VIs and cotton parameters are weather-dependent, as shown by the strikingly different results between the two seasons. There was a significant difference in precipitation between the first and second years, which most likely affected how the models performed. Further studies, with additional years of data collection can help increase the robustness of models with the incorporation of additional data.