The study was conducted at the Lang-Rigdon Farm (Figure 1) of the University of Georgia Tifton Campus in Tifton, Georgia, USA (31° 52’ N, 83° 55’ W) where the predominant soil type is classified as Tifton loamy sand (fine-loamy, kaolinite, plinthic kandiudults) (Soil Survey Staff, Natural Resources Conservation Service and United States Department of Agriculture, no date). During the 2021 and 2022 growing seasons, a field experiment was arranged in a randomized complete block design with five replications using a cotton cultivar DP 1646 B2XF and five nitrogen application rates of 0, 44, 89, 135, and 179 kg N ha^-1. DP 1646 B2XF was the most widely grown cotton cultivar in the southeastern US during the time of this experiment (USDA Agricultural Marketing Service, 2020), and the five nitrogen application rates were implemented to create variability in crop growth and yield. N was applied as granular urea (46-0-0): 25% at planting and 75% at the initiation of floral bud development (squaring stage). All other agronomic management practices, except N application, were conducted based on the recommendations outlined in the University of Georgia Cotton Production Guide (Hand et al., 2022). The cotton crop was sown on June 1, 2021 and April 26, 2022, and machine harvested on October 25, 2021 and September 22, 2022. The weather data which includes daily maximum and minimum temperatures, and daily solar radiation ( Supplementary Figure 1 ) for the study site from planting until harvest was obtained from the University of Georgia Weather Network (http://www.georgiaweather.net/). The 2022 growing season had higher average daily minimum and maximum temperatures by 0.15°C and 1.16°C, respectively than the 2021 growing season. The 2021 growing season had cloudier days with daily solar radiation below 10 MJ m^-2 than the 2022 growing season. Daily maximum and minimum temperatures were used to calculate GDDs throughout the season for cotton during both years.

The accumulated GDDs of cotton over time is a predictor of growth and phenological stage during the season. The GDDs is calculated based on the average daily temperatures and base temperature of cotton using equation 1 (Ritchie et al., 2004).

where i =1 signifies first day of planting and n is each sampling date. For cotton, the base temperature is 15.6°C (Snider et al., 2021b; Hand et al., 2022) and the upper threshold of maximum temperature is 33.9°C (Hand et al., 2022). The growth and development of cotton plants is assumed to be zero below the base temperature and above the upper threshold temperature.

Physiological measurements and UAV-based imagery were collected on the same dates, beginning 5 weeks after planting, throughout the 2021 and 2022 growing seasons in approximately two-week intervals, depending on weather conditions on target sampling dates. The actual dates of data collection along with the GDD for each sampling date are provided in Table 1 .

The IPAR_f models generated using training data had generalized R² values ranging from 0.66 to 0.90 ( Table 4 ). The model equations are provided in the Supplementary Table 1. Upon comparing the model performance for both training and cross-validation, RVI, RECI, NDRE, and SCCCI were the top four VIs with the highest R² and R²cv values, and lowest AICc, BIC, and RMSEcv values. The predicted IPAR_f from these four models was able to explain 89% to 93% of the variation in measured IPAR_f with RMSEcv ranging from 0.080 to 0.097 during cross-validation ( Figure 4 ). Compared to the VIs, models from near-infrared and red reflectance bands explained 85% and 87% of measured IPAR_f; however, the RMSEcv was higher for these raw bands (0.139, and 0.115, respectively).

For above-ground biomass, models developed from the training dataset had generalized R² values ranging from 0.69 to 0.87 indicating a moderate to strong relationship with the VIs ( Table 5 ). These model equations are provided in Supplementary Table 2. Based on the model performance, MSAVI, OSAVI, RVI, and SAVI were the four VI-based models with the highest R² and R²cv values, and the lowest AICc, BIC, and RMSEcv values. The estimated above-ground biomass from these models was able to explain 83-84% of variation with RMSEcv ranging from 240.64 to 251.34 g m^-2 during cross-validation ( Figure 5 ). GNDVI and NIR/G had the highest R² values of 0.86 and 0.87, respectively, during training, but the R² values reduced to 0.75 and 0.76, respectively during cross-validation. Out of the five raw bands, the model with the NIR band was able to explain 79% of the variation in the measured above-ground biomass during cross-validation.

The models created using training data to predict RUE explained 2% to 43% of the variation in RUE indicating a poor to low relationship of VIs to the RUE ( Table 6 ). The model equations are provided in Supplementary Table 3. While some of the VIs such as WDRVI, OSAVI, MSAVI, and TCARI explained as much as 43% variation in the RUE for the training data, they did not perform well during validation (Rcv = 25-33%). RECI, NIR/G, NDRE, and SCCCI performed best during training and cross-validation with lower AICc, BIC, and RMSEcv values, and the highest R²cv value. A plot of predicted versus measured RUE using validation data ( Figure 6 ) showed that the above-mentioned models explained 37% to 40% of the variation in cotton RUE with around 0.4 g MJ^-1 RMSEcv. Among the five raw reflectance bands, the RUE prediction model using the red-edge and green bands explained 37% and 34%, respectively, of variation in RUE during cross-validation.

Based on the model identified for IPAR_f and biomass in sections 3.1 and 3.2, the models that performed the best during validation were used to derive RUE using equation 3. The estimated RUE from MSAVI-based biomass model and RVI-based IPAR_f model only explained 18% of the variation in measured RUE with an RMSE of 0.462 ( Figure 7A ). However, replacing the MSAVI-based biomass model with the ground-measured biomass increased the prediction accuracy to 84% with the RMSE of 0.207 ( Figure 7B ).

Cotton fiber Index (CFI) from UAV-based RGB imagery showed a linear relationship with machine-harvested lint yield and explained 69% of the variation (RMSE = 244.07 kg ha^-1) in the lint yield ( Figure 8 ). However, the estimated harvest index (HI) utilizing the CFI-based lint yield model and the four models that performed the best in predicting above-ground biomass ( Figure 9 ) wasn’t significant (p-value > 0.05) in explaining measured HI. A plot of predicted versus measured HI for the two growing seasons (2021 and 2022) ( Figure 9 ) shows that the estimated HI was able to explain 40% to 49% of the variation in cotton HI, for the 2022 growing season, while only 1% to 4% for the 2021 growing season.

The first objective of this study was to develop and validate models to estimate the fraction of IPAR (IPAR_f), RUE, and biomass of cotton throughout the season using reflectance data (which includes multispectral vegetation indices (VIs) and raw reflectance bands) and growing degree days (GDD). Past research has demonstrated that the relationship between crop biophysical characteristics and VIs can be altered by growth stage (Gutierrez et al., 2012; Li Z. et al., 2022). For example, during cotton growth and development, biomass distribution, specific leaf nitrogen and chlorophyll contents in the canopy change, resulting in varied spectral responses throughout the season (Snider et al., 2021b). Gutierrez et al. (2012) showed the need of separate models to explain above-ground biomass and LAI of cotton plants at different days after planting. In this study, the inclusion of the GDD term as a predictor helped in addressing the above limitation and allowed us to directly account for the changes in the spectral reflectance of cotton plants at different growth stages during the growing season. The proposed model of VIs along with GDD can be used to estimate the IPAR_f and aboveground biomass throughout the season.

During cross-validation, the developed models explained up to 93% of the variation in measured IPAR_f of the cotton canopy ( Table 4 ), where RVI, RECI, NDRE, and SCCCI had the highest coefficient of determination ( Figure 4 ). Past studies have relied on either a physical radiative transfer model (Clevers et al., 1989) or hyperspectral VIs (Tan et al., 2013; Tan et al., 2018) to estimate the fraction of light absorption by crop canopies. However, there is limited information on the relationship between multispectral VIs and IPAR_f. The amount of light intercepted by a cotton canopy is highly influenced by canopy structure and architectural traits such as leaf area index (Brodrick et al., 2013; Bhattacharya, 2019). Multiple studies have reported the association between cotton leaf area index (LAI) and multispectral VIs (Hatfield et al., 1984; Zhao et al., 2007b; Gutierrez et al., 2012), and insights on the relationship between IPAR_f and VIs can be drawn from these studies. RVI, the ratio of near-infrared to red reflectance, has been shown to be effective (R² = 0.69 to 0.93) in explaining cotton LAI (Zhao et al., 2007a; Gutierrez et al., 2012; Chen, 2019). In the visible and near-infrared wavelength region, the near-infrared is reflected by the canopy and is sensitive to green photosynthetically active vegetation, whereas the red band is strongly absorbed by chlorophyll, making RVI a good predictor of LAI (Tucker, 1979). In this study, models based on raw bands of near-infrared and red reflectance were also able to moderately estimate IPAR_f throughout the growing season, further highlighting the importance of these two reflectance bands in explaining the variability in IPAR_f. NDVI, which is also a function of near-infrared and red reflectance bands, has been used to predict cotton LAI and canopy cover (Ritchie et al., 2010; Adams et al., 2021). Guillen-Climent et al. (2012) also reported that out of several VIs, NDVI was best correlated with IPAR_f for crops with homogenous canopies such as wheat, maize, and soybean. However, in the current study, NDVI saturated after the third sampling date (800 GDD) with strong absorption of red wavelength light by the canopy (red reflectance moved closer to zero) and couldn’t explain the variation in IPAR_f during the latter part of the season. A similar trend of NDVI saturating at high LAI was reported by Chen (2019) and Gutierrez et al. (2012). The other best performing VIs, RECI, NDRE, and SCCCI, are also reported to explain LAI and canopy chlorophyll content. RECI has been found to be closely related to maize LAI (R² = 0.95), even at higher LAI values (0 to > 6), without reaching saturation. Similarly, RECI performed the best without saturation even at higher GDD in this study. This is because RECI is based on near-infrared and red-edge reflectance, and the red-edge wavelength of light is not strongly absorbed by chlorophyll compared to the red wavelength (Gitelson et al., 2003b). In a similar manner, NDRE, and SCCCI are also based on near-infrared and red-edge reflectance and have been found to be associated with leaf and canopy chlorophyll content (Gitelson et al., 2003a; Raper and Varco, 2014; Ballester et al., 2017), which explains their effectiveness in explaining IPAR_f in this study.

The developed models using reflectance data and GDD were able to explain 55% to 84% of the variation in above-ground biomass for the validation dataset ( Table 5 ). MSAVI, OSAVI, RVI, and SAVI were the best performing VIs with the highest coefficient of determination ( Figure 5 ). The near-infrared band explained the most variation (79%) in above-ground biomass out of the five raw bands. For crops and other vegetation, the red band is highly absorbed by chlorophyll, whereas the near-infrared is reflected from the outer leaf surfaces and spongy mesophyll cells. (Jackson, 1986). The VIs computed based on these two bands, such as NDVI and RVI, have been linked to above-ground biomass previously (Gitelson et al., 2003a; Zhao et al., 2005; Zhao et al., 2007b; Chao et al., 2019). RVI is one of the best VIs to explain above-ground biomass in this study, which is similar to previous findings. NDVI is also widely used VI that has been extensively related to above-ground biomass. However, it tends to saturate early with increasing biomass, and is influenced by soil reflectance, particularly under low vegetation cover (Huete, 1988; Qi et al., 1994; Gutierrez et al., 2012). VIs such as SAVI, MSAVI, and OSAVI are all modifications of NDVI which are more sensitive to changes in vegetation cover and are less influenced by soil reflectance during the lag phase of vegetative growth (Rouse, 1974; Huete, 1988; Qi et al., 1994; Rondeaux et al., 1996), which explain their strong relationship with biomass throughout the season in the current study. Previous studies have also reported RVI (Gutierrez et al., 2012), MSAVI (Brandão et al., 2015), and OSAVI (Zhao et al., 2007b) to be a good predictor of cotton above-ground biomass at early and peak growth stages. However, recent studies have shifted their attention towards use of hyperspectral (Junhua et al., no date) and LIDAR (Sun et al., 2018; Furbank et al., 2019) sensors but these are currently expensive and require extensive data analysis for crop aboveground biomass estimation. In our study, we noticed that with increasing GDD, the estimated above-ground biomass values moved further away from measured values. This source of error and uncertainty in the biomass model could be attributed to the indeterminate growth habit of cotton because the above-ground biomass during the later season is a combination of reproductive and vegetative biomass, and the reflectance data cannot fully capture the reproductive structures in the lower canopy.

The VIs developed from multispectral imagery only explained up to 40% of the variation in cotton RUE ( Table 6 ). This could be due to the fact that RUE is affected by various environmental factors and the interaction of different physiological processes, making it difficult to accurately assess using multispectral remote sensing (Furbank et al., 2019). However, the relationship between multispectral VIs and RUE can help in the identification of valuable predictors for future modeling. RECI, NIR/G, NDRE, and SCCCI were the top four VIs in predicting RUE ( Figure 6 ). Similarly, the red-edge and green raw bands explained 37% and 34% of the variation in RUE for the validation data set. All the aforementioned VIs and raw reflectance bands have shown to be sensitive to canopy chlorophyll content for crops including corn and cotton (Tumbo et al., 2002; Gitelson et al., 2003a; Bausch et al., 2008; Ritchie et al., 2010; Raper and Varco, 2014; Ballester et al., 2017). The specific absorption coefficient of chlorophyll for the green and red-edge bands is lower than the red and blue bands, making the VIs based on green and red-edge bands sensitive to changes in chlorophyll content in plant tissues (Gitelson and Merzlyak, 1997; Gitelson et al., 2003b). Because the quantum efficiency of primary photochemistry is influenced by canopy chlorophyll content per unit leaf area (Porcar-Castell et al., 2021), the aforementioned VIs may relate to canopy-level RUE. There are limited studies estimating canopy RUE from multispectral VIs. However, Robles-zazueta et al. (2021) showed potential of hyperspectral VIs and partial least square regression (PSLR) modelling to predict RUE in wheat, where EVI and PRI predicted RUE with 70% accuracy.

Additionally, this study showed the potential of using a mechanistic model (RUE = Biomass/IPAR) from estimated IPAR and biomass to derive RUE in cotton. The estimated RUE from the measured biomass and RVI-based IPAR_f model was able to explain 84% of measured RUE ( Figure 7 ). A similar approach to obtain RUE from measured biomass and remotely sensed light interception has been used in corn (Tewes and Schellberg, 2018) and soybean (Phillips et al., 2020) as well. Further, in this study, actual biomass measurements were replaced with estimates from the MSAVI-based biomass model, which resulted in a reduced prediction accuracy of 18%. The reason for the reduced performance could be due to the error and inaccuracies associated with biomass prediction (Robles-zazueta et al., 2021). As previously discussed, cotton is an indeterminate crop, and its above-ground biomass during the season consists of both vegetative and reproductive structures. Therefore, the spectral VIs may not represent the biomass contributed by the bolls in lower canopy. Moreover, due to cotton’s indeterminate growth pattern, it is challenging to uniformly sample above-ground biomass throughout the season.

The second objective of this study was to estimate cotton lint yield using cotton fiber index (CFI). CFI explained 69% of the variation in cotton lint yield ( Figure 8 ), which is somewhat lower than the result reported by Huang et al. (2016) (R² = 0.83) and Feng et al. (2020) (R² = 0.90). However, the performance of CFI in explaining lint yield variation was better than previous studies that used in-season vegetation indices to predict yield (R² = 0.47 to 0.60) (Gutierrez et al., 2012; Ballester et al., 2017). There are a few limitations of using CFI from UAV-based RGB imagery which could introduce some error and possibly the reason for low accuracy observed in our study. CFI does not consider the cotton bolls present in the lower and middle of the canopy (Feng et al., 2020; Siegfried et al., 2023), and there is a potential for misclassification of soil or background pixels as white cotton bolls. Also, the UAV flight height can have an influence on the image resolution for cotton pixel detection. The lower flight height can increase the detection accuracy but can limit the amount of area covered as well as increase time for image processing. To increase the yield-predicting performance of CFI, future research could focus on identifying low cost remote sensing predictors, such as UAV-based plant height or white flower detection during the season, which can account for the cotton bolls in the middle and below the canopy. Utilizing machine learning and convolutional neural networks to accurately delineate cotton pixels from the surrounding soil pixels could also improve yield prediction (Li F. et al., 2022; Shi et al., 2022).

Finally, the last objective of this study was to investigate the utility of the developed models for lint yield and above-ground biomass in predicting cotton harvest index (HI). No significant relationship was observed between the measured and predicted HI when the data from both years was combined. It was also found that the HI models performed differently for each growing season ( Figure 9 ), with the estimated HI explaining 40-49% of the variation in measured HI for the 2022 growing season. The lack of a significant relationship over the two years of the study could be attributed to the combined errors in the lint yield and biomass prediction. As discussed earlier, for an indeterminate crop such as cotton, nadir UAV imagery cannot fully account for the reproductive structure or cotton bolls in the lower and middle canopy that contribute to both above-ground biomass and lint yield (Shi et al., 2022; Siegfried et al., 2023). Furthermore, previous studies have reported varying responses of cotton HI (indifferent to inverse relationship) to different nitrogen levels (Kimball and Mauney, 1993; Gerik et al., 1994; Li et al., 2017), which could be another reason for the varied responses observed in the different seasons.

The models developed to estimate the yield-contributing physiological parameters in cotton showed low to strong performance. The models developed in this study can estimate the yield determining physiological parameters throughout the growing season and are relatively simple with only two required input predictors- VI from multispectral imagery and GDD (growth stage parameter). This makes use of these parameters a viable method to measure the light interception and above-ground biomass rapidly and accurately during the cotton growing season and also expands the potential of these models to be used by researchers and industry in cotton management and breeding. Previous work showed that low nitrogen stress in cotton primarily reduces the light interception of the cotton canopy, resulting in significant yield loss (Pokhrel et al., 2023). The models, therefore, can be applied to measure light interception and identify yield-limiting low nitrogen circumstances in-season. The models have ability to aid in agronomic decision making in order to prevent significant yield loss. Similarly, excessive irrigation can cause excessive above-ground biomass resulting in a lower number of cotton bolls and boll mass, ultimately causing significant yield loss (Ermanis et al., 2020). The models from this study can be potentially applied to regulate the irrigation requirement of a cotton plant during the growing season based on their light interception and biomass gain to avoid yield loss due to excess growth. For cotton breeding, the models can be utilized for high-throughput phenotyping of a large number of cotton genotypes for yield determining physiological parameters without the need for labor-intensive and time-consuming manual measurements. Finally, results attained in this study can be used to expand and improve future research on predicting functional yield drivers of cotton from aerial imagery.

One of the limitations of this study is that the modeling and validation were limited to data collected within 372 GDD to 1253 GDD, with no extrapolation, and it would be beneficial to test these models outside this range in future studies. Furthermore, the models were developed based on a two-year study of a single cotton cultivar’s response to five different nitrogen application rates at an experimental site. Inclusion of training data from different cotton cultivars across multiple production environments can enhance model performance. It is suggested that future research should focus on increasing the model’s robustness for its transferability across a wide range of genotypes and environmental conditions. Independent validation from different locations would also help in determining the model’s prediction ability. In the future, more robust models can be created by exposing the crop to a broader range of growth and yield-altering factors. Additionally, due to the limited number of data points, validation for the lint yield modeling was not performed. Validation for the use of CFI in lint yield prediction can be done in the future. There is also the possibility of misclassification while separating cotton white pixels and background soils for CFI calculation. Future research can work on assessing quality of those filtering techniques to avoid the misclassifications. Moreover, top-view UAV imagery from 45 m height may not fully capture the cotton bolls present in the middle section of cotton plants. Future research can investigate the influence of different flight parameters such as flight height and sensor resolution on accurate estimation of cotton lint yield and above-ground biomass prediction. This would potentially aid in increasing the estimation accuracy of the cotton harvest index. Overall, this study is the first attempt of its kind to use low-cost UAV RGB and multispectral imagery to predict yield-determining physiological parameters of cotton throughout the season and the results obtained here shows a strong potential to utilize and expand the use of remotely-sensed imagery for estimating yield-driving functional traits of cotton.