[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Next Article in Journal
Assessing Comparative Yield and Yield Contributing Traits of Hybrid Rice Varieties Transplanted as Over-Aged Seedlings in the South-Central Coastal Ecosystem of Bangladesh
Previous Article in Journal
Impact of Mutations in Soybean Oleate and Linoleate Desaturase Genes on Seed Germinability of Heat-Stressed Plants
You seem to have javascript disabled. Please note that many of the page functionalities won't work as expected without javascript enabled.
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Crops
Volume 5
Issue 1
10.3390/crops5010003
crops-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Estimating Switchgrass Biomass Yield and Lignocellulose Composition from UAV-Based Indices

by
Daniel Wasonga
1,2,
Chunhwa Jang
1,2,*,
Jung Woo Lee
1,
Kayla Vittore
1,
Muhammad Umer Arshad
1,2,
Nictor Namoi
1,2,
Colleen Zumpf
3 and
DoKyoung Lee
1,2,3
1
Department of Crop Sciences, University of Illinois at Urbana-Champaign, Urbana, IL 61801, USA
2
Agroecosystem Sustainability Center, Institute for Sustainability, Energy, and Environment, University of Illinois at Urbana-Champaign, Urbana, IL 61801, USA
3
Energy Systems Division, Argonne National Laboratory, 9700 South Cass Ave., Argonne, IL 60439, USA
*
Author to whom correspondence should be addressed.
Crops 2025, 5(1), 3; https://doi.org/10.3390/crops5010003
Submission received: 30 November 2024 / Revised: 6 January 2025 / Accepted: 8 January 2025 / Published: 16 January 2025
Browse Figures
Figure 1
<p>Seasonal trajectories of the normalized difference vegetation index (NDVI), green normalized difference vegetation index (GNDVI), normalized difference red-edge index (NDRE), green ratio vegetation index (GRVI), and simple ratio (SR) index as influenced by cultivar and N treatments. The indices were calculated from multispectral images taken over large-scale switchgrass plots at the Urbana Energy Farm, IL, during the 2021, 2022, and 2023 growing seasons. Error bars represent the mean standard errors of three replicates.</p> "> Figure 2
<p>Pearson correlation coefficients between vegetation indices measured at different times (June, July, August, and September) and biomass yield and lignocellulose components. NDVI, normalized difference vegetation index; GNDVI, green normalized difference vegetation index; NDRE, normalized difference red-edge index; GRVI, green ratio vegetation index; SR, simple ratio index. Significance levels *, **, *** indicate <span class="html-italic">p</span> &lt; 0.05, <span class="html-italic">p</span> &lt; 0.01, <span class="html-italic">p</span> &lt; 0.001, respectively.</p> "> Figure 3
<p>Scatter plots of predicted and actual switchgrass biomass yields based on linear, exponential, partial least square, and random forest regression models for the training and validation sets using the univariate GNDVI from mid-August imagery. The shaded areas indicate the 95% confidence intervals associated with model fit. The dots (blue points) represent the actual data points from the dataset, while the red lines indicate how well the model fits or predicts the data.</p> "> Figure 4
<p>Scatter plots of predicted and actual switchgrass biomass yields based on linear, exponential, partial least square, and random forest regression models for the training and validation sets using a multivariate combination of the GNDVI and NDRE from mid-August imagery. The shaded areas indicate the 95% confidence intervals associated with model fit. The dots (blue points) represent the actual data points from the dataset, while the red lines indicate how well the model fits or predicts the data.</p> "> Figure 5
<p>Scatter plots of predicted and actual cellulose concentration based on linear, exponential, partial least square (PLSR), and random forest (RF) regression models for the training and validation sets using the normalized difference red-edge index (NDRE) from mid-August imagery. The shaded areas indicate the 95% confidence intervals associated with model fit. The dots (blue points) represent the actual data points from the dataset, while the red lines indicate how well the model fits or predicts the data.</p> "> Figure 6
<p>Scatter plots of predicted and actual hemicellulose concentration based on linear, exponential, partial least square (PLSR), and random forest (RF) regression models for the training and validation sets using a multivariate combination of the GNDVI, NDRE, NDVI, and GRVI from mid-July imagery. The shaded areas indicate the 95% confidence intervals associated with model fit. The dots (blue points) represent the actual data points from the dataset, while the red lines indicate how well the model fits or predicts the data.</p> "> Figure 7
<p>Scatter plots of predicted and actual acid detergent lignin (ADL) based on linear, exponential, partial least square (PLSR), and random forest (RF) regression models for the training and validation sets using a combination of the GNDVI, NDVI, and GRVI from mid-July imagery. The shaded areas indicate the 95% confidence intervals associated with model fit. The dots (blue points) represent the actual data points from the dataset, while the red lines indicate how well the model fits or predicts the data.</p> ">
Versions Notes

Abstract

:
Innovative methods for estimating commercial-scale switchgrass yields and feedstock quality are essential to optimize harvest logistics and biorefinery efficiency for sustainable aviation fuel production. This study utilized vegetation indices (VIs) derived from multispectral images to predict biomass yield and lignocellulose concentrations of advanced bioenergy-type switchgrass cultivars (“Liberty” and “Independence”) under two N rates (28 and 56 kg N ha−1). Field-scale plots were arranged in a randomized complete block design (RCBD) and replicated three times at Urbana, IL. Multispectral images captured during the 2021–2023 growing seasons were used to extract VIs. The results show that linear and exponential models outperformed partial least square and random forest models, with mid-August imagery providing the best predictions for biomass, cellulose, and hemicellulose. The green normalized difference vegetation index (GNDVI) was the best univariate predictor for biomass yield (R2 = 0.86), while a multivariate combination of the GNDVI and normalized difference red-edge index (NDRE) enhanced prediction accuracy (R2 = 0.88). Cellulose was best predicted using the NDRE (R2 = 0.53), whereas hemicellulose prediction was most effective with a multivariate model combining the GNDVI, NDRE, NDVI, and green ratio vegetation index (GRVI) (R2 = 0.44). These findings demonstrate the potential of UAV-based VIs for the in-season estimation of biomass yield and cellulose concentration.

1. Introduction

Switchgrass (Panicum virgatum L.) is a warm-season perennial grass with a high biomass yield and favorable lignocellulose composition, making it an ideal feedstock for biofuel production [1,2]. It requires minimal agricultural inputs and can thrive on marginal lands while generating a wide range of ecosystem services [3,4]. Switchgrass is expected to play a critical role in achieving medium- and long-term sustainable aviation fuel (SAF) goals, contributing 15% of the 1.5 billion dry tons of biomass required in a medium-mature market scenario [5,6]. With the push toward commercial-scale switchgrass, efficient preharvest methods to estimate biomass yield and feedstock quality are essential for optimizing harvest and biorefinery planning [7,8].
Cultivar selection and management are crucial for maximizing biomass yield and quality in switchgrass [2]. Recently developed bioenergy switchgrass varieties offer high yield potential, increased fiber concentration, and lower lignin and ash concentrations [9,10]. Effective management practices, including N optimization and strategic harvest timing, particularly post-frost harvesting, are recommended to enhance quality and support sustainable long-term biomass production [11]. Higher proportions of cellulose and hemicellulose in biomass are generally preferred due to their positive impact on ethanol production potential during biochemical conversion, while high lignin concentration is undesirable [12,13,14,15]. Traditionally, the estimation of feedstock yield and quality traits is conducted post-harvest, involving plant sampling, drying, grinding, and wet chemical assays or infrared (NIR) spectroscopy [16]. While generally considered accurate, these methods are laborious and time-consuming and do not provide ample time for biorefineries to adjust conversion processes based on biomass quality and yields [17].
Recent advancements in phenotyping technologies have significantly improved the monitoring and evaluation of bioenergy grasses in the United States [18,19]. Remote sensing tools, including hyperspectral and multispectral imaging, have enabled the precise estimation of traits such as biomass, chlorophyll content, and water-use efficiency in bioenergy grasses, facilitating the identification of high-performing genotypes across diverse environmental conditions [20,21,22]. Field studies have further highlighted the potential of phenotyping platforms in assessing phenological traits, such as flowering time and senescence, which are critical for optimizing harvest schedules and maximizing biomass quality [23,24,25]. Moreover, machine learning algorithms and data fusion techniques have been integrated with phenotyping data to improve the predictive modeling of yield and stress responses [24,26]. Therefore, by leveraging UAV-based technologies, switchgrass phenotyping can overcome the challenges of traditional methods, offering scalable, accurate, and time-efficient solutions for optimizing biomass yield and quality to meet growing bioenergy demands.
Remote sensing with Unmanned Aerial vehicles (UAVs) equipped with multispectral sensors has emerged as a promising approach for collecting high temporal and spatial resolution imagery datasets that can be used to predict various plant traits [27,28]. The strong correlation between spectral bands and vegetation parameters enables the indirect estimation of plant traits, such as biomass, canopy cover, leaf area index, and plant chemical composition [29,30]. Reflectance values measured from multispectral images in the visible and near-infrared (NIR) bands are combined, either linearly or non-linearly, to calculate vegetation indices (VIs) [29]. The VIs derived from UAV-based multispectral sensors have recently been used to estimate biomass yield in various perennial energy crops, including energy cane [31,32], Miscanthus [33], and corn [34]. However, the estimation of switchgrass biomass yield has mainly been performed using VIs derived from satellite imagery [35,36], which has limitations, including low spatial resolution and cloud cover obstruction during imaging, which lowers image quality and requires complex modeling and statistics [37,38].
Cellulose, hemicellulose, and lignin are carbon-based molecules influencing biomass structural and chemical properties [39]. The chemical bonds in these compounds absorb NIR wavelengths, causing excitation to higher energy levels. By analyzing their unique spectral signatures, mathematical models can correlate specific chemical information with NIR spectral data, allowing for lignocellulose estimation from images [40,41]. Multispectral imaging facilitates the monitoring of lignocellulose composition by capturing spectral information linked to these molecular structures [42]. Limited research has utilized UAV imagery to estimate switchgrass lignocellulose composition. To our knowledge, only Xu et al. [43] successfully predicted lignin concentration using the normalized difference vegetation index (NDVI) and the normalized difference vegetation index (NDRE).
Multispectral measurements often yield highly correlated spectral data, making multivariate statistical techniques essential for improving prediction accuracy [41]. Among these techniques, partial least square regression (PLSR) is particularly effective for multivariate prediction in collinear data, as it combines principal component analysis with multiple linear regression to capture variability in both predictor (X) and response (Y) variables [44]. Random forest (RF), an ensemble machine learning method, enhances prediction accuracy by bootstrapping multiple decision trees from randomly selected variables and aggregating their outputs [45]. Linear regression, commonly used to model relationships between variables, can take the form of simple linear regression (SLR), with one independent variable, or multiple linear regression (MLR) with two or more [46]. For relationships involving exponential change rates, exponential regression is used with either single (SER) or multiple independent variables (MER). Linear, exponential, RF, and PLSR models perform variably in predicting the biomass yield of herbaceous energy crops, highlighting the need for tailored model selection based on specific crop and data characteristics [33,34,35,36,43]. Few studies have, however, compared the predictive capabilities of linear and non-linear models specifically for switchgrass. Linear models are simpler and require fewer assumptions about the underlying data structure, making them interpretable and computationally efficient, whereas non-linear models can capture complex relationships and interactions in the data that linear models might miss [44,45,46]. Therefore, understanding the strengths and limitations of both approaches enables the selection of the most suitable model to improve prediction accuracy and extract meaningful insights from multispectral imaging data. The objective of this study was to evaluate the use of UAV-derived indices with various regression models to estimate biomass yield and feedstock composition in advanced bioenergy switchgrass cultivars “Liberty” and “Independence” [9,47].

2. Materials and Methods

2.1. Site Description and Experimental Design

This study was conducted at the University of Illinois Energy Farm in Urbana, Illinois (40°4′7.68″ N, –88°11′26.78″ W), during the 2021–2023 growing seasons. The site had a history of corn–soybean rotation and is classified as marginal [35]. The soil is a Dana silt loam (Fine-silty, mixed, superactive, mesic Oxyaquic Argiudolls). The experimental layout followed a split-plot arrangement within a randomized complete block design with three replicates. The main plots were two bioenergy switchgrass cultivars (Independence and Liberty), while the subplots were two N application rates (28 and 56 kg N ha−1). The individual plot size was approximately 0.2 ha (N = 12) with a 3 m alley between the blocks (Supplementary Figure S1). Planting was accomplished using a no-till drill (Great Plains Plot Planter, Salina, KS, USA) in the spring of 2020 at a seeding rate of 6.8 kg of pure live seed (PLS) per acre for Independence and 3.6 kg PLS per acre for Liberty, with row spacing of 19 cm and planting depth of 1.3 cm. Weed control was accomplished using pre-emergence application of atrazine (2-chloro-4ethylamine-6-isopropylamino-s-triazine) at 2 kg of active ingredient per ha. Nitrogen (N) fertilizer was broadcast as urea (40-0-0) beginning in the spring of the second year (2021) and continued each spring in 2022 and 2023.

2.2. UAV Imaging and Processing

High-resolution multispectral data were acquired using a DJI Inspire 2 quadcopter (SZ DJI Technology Co., Ltd., Shenzhen, China) in mid-June, mid-July, mid-August, and mid-September during the 2021–2023 growing seasons (Table 1). The quadcopter was flown in a single north–south serpentine flight pattern and was equipped with nadir-fixed 3.2 MP Micasense Altum sensors (AgEagle Sensor Systems Inc., Seattle, WA, USA), which included blue (459–491 nm), green (546.5–573.5 nm), red (661–675 nm), red-edge (711–723 nm), near-infrared (813.5–870.5 nm), and long-wave infrared thermal (8–14 μm) sensors. The images were collected at 45 m above ground level with 80% front and 75% side overlap, obtaining a ground sample surface (GSD) of 0.02 m. Flights were conducted within two hours of solar noon with a wind speed of less than 8 km h−1 to avoid image distortion. A Micasense calibration panel was recorded before and after each flight with the Altum to aid post-processing for georeferencing purposes. The immediate GPS coordinates of the UAV were recorded in the metadata of each image and utilized during image reconstruction by the Agisoft Metashape Professional software version 2.0.1 (Agisoft LLC., St. Petersburg, Russia).
During image processing, the multispectral images were uploaded to Metashape, converted into a local coordinate system (WWGS 84/World Mercator; EPSG 3395), and processed using structure from motion (SfM) techniques. Calibration panels were masked out, and reflectance was calibrated from Altum sensor images. Images were aligned with high accuracy, and ground control points (GCPs) were identified and incorporated. Dense point clouds were generated with high-quality and mild depth filtering. The orthomosaic was exported in a common coordinate reference system (WGS 84/World Mercator; EPSG 3395) in GeoTIFF format without file compression.

2.3. Vegetation Indices

We calculated five vegetation indices based on the generated orthomosaics, as shown in Table 2. These indices are associated with five multispectral bands commonly used for estimating plant biomass, chlorophyll, and plant N content [29,48]. The indices were computed using the band arithmetic function within the ArcGIS Pro 2.9.1 (Esri, Redlands, CA, USA). To improve classification accuracy, we applied the majority filter module, which replaced the pixel values with the most frequent neighboring values. Subsequently, the zonal statistic module in ArcGIS Pro 2.9.1 was used to extract the average index values from each large-scale plot. The calculated indices are included in Table 2.

2.4. Model Training and Validation

In developing the model, we used 70% of the combined 2021 and 2022 datasets for training and 30% for testing and validated the model with the 2023 dataset. In each model, biomass yield, cellulose, hemicellulose, and lignin were the dependent variables, with VIs as the independent variables. Linear models were built in R using the “lm” function, PLSR with the “plsr” function, and RF with the “rf” function. To prevent overfitting and ensure reliable performance, we employed 10-fold repeated cross-validation on the training dataset. Cross-validation evaluates model performance by averaging results across k different validation sets. In the 10-fold cross-validation, the sample is divided into 10 subsamples, each used once as a hold-out group while the remaining 9 are used for training. The performance of the 10 models on these hold-out samples is recorded and averaged. For machine learning RF regression, hyperparameter tuning was conducted using a surrogate Bayesian optimization approach implemented through the DeepHyper framework [36]. This framework not only facilitates hyperparameter tuning within specified ranges but also supports the evaluation of contingent parameters. Its highly configurable search space enables automated machine learning (AutoML), a technique that evaluates a variety of machine learning models with minimal manual intervention. For the ensemble optimization task, the search space included the choice of algorithm (RF) as a hyperparameter. Models were evaluated using the mean absolute error (MAE), root mean squared error (RMSE), and coefficient of determination (R2). The R2 value provides a measure of how well the model can fit the data. Typically, R2 values range from 0 to 1, with higher values indicating a better fit. The MAE evaluates the average of the absolute differences between the predicted and actual values. Lower MAE and RMSE values are preferred because they indicate that the model’s predictions are closer to the observed values [53].

2.5. Biomass Yield Collection and Compositional Analysis

Switchgrass biomass yield was determined by harvesting the whole plot after a killing frost when the plants had senesced in early to mid-December of 2021–2023. The plots were harvested using a self-propelled mower-conditioner (New Holland model H8080, 750 HD Specialty Head, New Holland, PA, USA) at 10–15 cm cutting height. The biomass was baled using a square baler (New Holland model BB9080, New Holland, PA, USA) and weighed. Biomass subsamples (~1.0 kg) were collected from windrows in each plot prior to baling, dried in a forced-air drier at 60 °C for 72 h, and moisture content determined and used to adjust yields to a dry-matter basis (Mg ha−1).
The oven-dried subsamples were then ground to pass a 2 mm sieve using a Wiley mill (Model 4, Thomas Scientific, Swedesboro, NJ, USA), and compositional analysis was conducted. The concentrations of neutral detergent fiber (NDF), acid detergent fiber (ADF), and acid detergent lignin (ADL) were determined using a sequential extraction and filtration process (Ankom Technology, 2002 and 2003) and an Ankom 200 Fiber Analyzer (ANKOM Technology, Fairport, NY, USA). Cellulose concentration, ADL, and ADF were determined using procedures outlined by Van Soest et al. [54].

2.6. Data Analysis

The treatment effects on biomass yield, cellulose, hemicellulose, lignin concentrations, and VIs were analyzed using analysis of variance (ANOVA) using the generalized mixed model procedure. The cultivar, N rate, and year were considered fixed factors, while the blocks were considered random. The data normality and homogeneity were assessed using a Shapiro–Wilk test and an equal variance test to meet the ANOVA assumption. For data that deviated from normality, Tukey’s ladder of powers was applied to transform the data using the “TransformTukey” function from the “rcompanion” package in R. Where treatments showed significant differences, the means were compared using Tukey’s HSD test at p ≤ 0.05. Pearson correlations were also calculated between the VIs and the observed biomass, cellulose, hemicellulose, and lignin for the different measurement periods. All statistical analyses were performed in R software version 4.3.0 [55].

3. Results

3.1. Precipitation and Temperature Changes over the Growing Years

The monthly temperature and precipitation for the years 2021–2023 and the 30-year averages (1993–2023) are shown in Table 3. The annual precipitation in the years 2021–2023 was lower than the long-term average of 1009 mm. However, 2021 recorded higher annual precipitations than the years 2022 and 2023. The growing season (April–September) total precipitation was also higher in 2021 (540 mm) and lower in 2022 (405 mm) and 2023 (352 mm) compared to the long-term average (603 mm). Average annual temperatures were 11.9 °C in 2021, 11.2 °C in 2022, 12.7 °C in 2023, and 10.9 °C for the long term.

3.2. Trends in Biomass Yield and Lignocellulose Concentrations

Significant differences in biomass yield due to year, cultivar, and N application were detected (Table 4). Biomass yields during 2021–2023 ranged from 6.9 to 9.2 Mg ha⁻¹, with lower yields in 2022 due to drought conditions (Table 2). Averaged across the years, Liberty produced 22% more biomass than Independence (p < 0.001), and the 56 kg N ha−1 treatment resulted in a 33% increase in biomass yield compared to 28 kg N ha−1 (p < 0.001).
The concentrations of cellulose, hemicellulose, and lignin varied across the three years, with slightly lower concentrations in 2022 and 2023 due to drought (Table 3). The hemicellulose concentrations were significantly (p < 0.05) impacted by the cultivar treatment but not the cellulose and lignin concentrations (Table 3). Averaged across the years, Liberty showed a higher hemicellulose concentration than Independence. The lignocellulose concentrations showed differences due to N treatment, with 56 kg N ha−1 application resulting in improved cellulose, hemicellulose, and lignin concentrations compared to 28 kg N ha−1 (Table 4).

3.3. Seasonal Trajectory of the Vegetation Indices

The NDVI, GNDVI, NDRE, GRVI, and SR increased across the years from mid-June and peaked around mid-July or mid-August and then decreased to their lowest values in September (Figure 1). This trend likely reflects the growth pattern of switchgrass, where the canopy reaches maximum greenness around mid-summer, followed by a decline toward the end of the growing season as the crop matures and sets seed [56]. Peak VI values were slightly lower in 2022 compared to 2021 and 2023. For instance, the NDRE was 10% and 7% lower in 2022 relative to 2021 and 2023, which exhibited similar VI levels. Liberty showed higher index values than Independence with a clear distinction in both 2022 and 2023. The N application rates also impacted the NDVI, GNDVI, NDRE, GRVI, and SR with 56 kg N ha−1, resulting in higher index values than 28 kg N ha−1 (Figure 1).

3.4. Correlation Between Biomass Yield, Lignocellulose Concentrations, and VIs

Biomass showed positive correlations with all vegetation indices (VIs), with stronger correlations observed from mid to late season (July to August). Notably, biomass had consistently strong correlations with the NDVI (r = 0.85 and p < 0.001), GNDVI (r = 0.83 and p < 0.001), and SR (r = 0.84 and p < 0.001) in mid-August (Figure 2). Cellulose exhibited moderate to strong positive correlations, particularly with the GNDVI (r = 0.53 and p < 0.001) and NDVI (r = 0.50 and p < 0.01) in mid-July. In mid-July, lignin also showed moderate positive correlations (r > 0.5 and p < 0.001) with the GNDVI, GRV, and NDVI. In contrast, hemicellulose had weaker correlations across the indices and measurement periods (Figure 2).

3.5. Biomass Yield Estimation

We found strong predictive accuracy for biomass yield using mid-August imagery, with the GNDVI emerging as the most effective univariate predictor for biomass yield across the models (Figure 3). Using the GNDVI from mid-August imagery, the linear model showed the best model fit on the training set (R2 = 0.86, RMSE = 0.54 Mg ha−1, and MAE = 0.43 Mg ha−1) and maintained a high model fit on the validation set (R2 = 0.86, RMSE = 0.59 Mg ha−1, and MAE = 0.43 Mg ha−1). The exponential model showed slightly better validation performance (R2 = 0.86, RMSE = 0.44 Mg ha−1, and MAE = 0.32 Mg ha−1) than the training set. The PLSR model demonstrated moderate non-linear fit comparable to the exponential model, especially on the validation set (R2 = 0.86, RMSE = 0.44 Mg ha−1, and MAE = 0.32 Mg ha−1). The RF model achieved the highest fit on the training data (R2 = 0.93, RMSE = 0.39 Mg ha−1, and MAE = 0.33 Mg ha−1) but exhibited a notable decrease in predictive accuracy on the validation set (R2 = 0.72, RMSE = 0.82 Mg ha−1, and MAE = 0.55 Mg ha−1) (Supplemental Table S1).
The multivariate combination of the GNDVI and NDRE from mid-August imagery resulted in improved biomass yield predictions compared to univariate predictors alone (Figure 4). Model performance with the multivariate GNDVI and NDRE on both training and validation sets followed similar trends with the univariate GNDVI predictor. In the training sets, the RF model provided a stronger fit (R2 = 0.94, RMSE = 0.36 Mg ha−1, and MAE = 0.30 Mg ha−1) with predicted values closely aligned to actual values. The linear model also performed well on the training set (R2 = 0.88, RMSE = 0.51 Mg ha−1, and MAE = 0.44 Mg ha−1), while the exponential and PLSR models demonstrated moderate fit with wider confidence intervals. In the validation set, the linear and exponential models maintained similar predictive accuracy to their training performance, showing consistent model fit across datasets. The PLSR model displayed moderate consistency between training and validation sets, though with slightly lower predictive accuracy than the linear and exponential models. The RF model’s performance declined on the validation set (R2 = 0.69, RMSE = 0.98 Mg ha−1, MAE = 0.68 Mg ha−1) compared to the training set (Figure 4). Generally, biomass yield predictions were weak when using imagery collected in mid-June and mid-September across models, regardless of univariate or multivariate indices.

3.6. Lignocellulose Estimation

The NDRE was the best univariate predictor of cellulose, with imagery from mid-August showing more accurate predictions than from other time points (Figure 5). Cellulose was moderately predicted across the models using the NDRE from mid-August imagery, with slight variations in the training and validation sets (Figure 5). In the training set, we observed comparable performance with both linear (R2 = 0.47, RMSE = 13.67 g kg−1, MAE = 11.51 g kg−1) and exponential (R2 = 0.46, RMSE = 13.65 g kg−1, MAE = 11.45 g kg−1) models, which exhibited moderate fit with narrow confidence intervals. The PLSR model also exhibited a moderate fit on the training set (R2 = 0.44), but it was weaker compared to the linear and exponential models. In contrast, the RF model demonstrated the strongest fit on the training set with an R2 of 0.57 and the lowest RMSE (8.62 g kg−1) and MAE (7.01 g kg−1) values. On the validation set, the linear and exponential models showed improved fits (R2 = 0.53 and 0.51, respectively). The PLSR model also maintained model fit on the validation set with an R2 of 0.46, but a higher RMSE of 18.21 g kg−1, and an MAE of 15.07 g kg−1. The RF model showed decreased performance on the validation set with an R2 of 0.55 and increased RMSE (17.75 g kg−1) and MAE (15.89 g kg−1) values. The multivariate combination of VIs did not improve cellulose prediction.
Hemicellulose prediction was generally moderate and most effective with mid-July imagery compared to other time points (Figure 6). Univariate VIs were ineffective for predicting hemicellulose, but the multivariate combination of the GNDVI, NDVI, NDRE, and GRVI from mid-July produced the best predictions across models. In the training set, both the linear and the exponential models exhibited moderate predictive accuracy (R2 = 0.48) for hemicellulose. The PLSR model showed a weak correlation between the predicted and actual values with high RMSE and MAE values (R2 = 0.23, RMSE = 18.24 g kg−1, MAE = 15.99 g kg−1). Although the RF demonstrated the best fit for the training data (R2 = 0.59, RMSE = 9.57 g kg−1, MAE = 8.7 g kg−1), it showed a notable decline in predictive accuracy on the validation set (R2 = 0.35, RMSE = 15.22 g kg−1, MAE = 11.27 g kg−1) for hemicellulose prediction. In the validation set, the linear and exponential models maintained relatively consistent performance in the validation set (R2 = 0.44, RMSE = 10.73 g kg−1, MAE = 8.71 g kg−1). The PLSR model’s performance declined significantly in the validation set, showing further reduced hemicellulose prediction (R2= 0.21, RMSE = 17.16 g kg−1, MAE =14.85 g kg−1) (Figure 6).
Lignin prediction was most effective using mid-July imagery with a multivariate approach. While univariate predictors proved ineffective for lignin prediction and are therefore not shown, the combination of the GNDVI, NDVI, and GRVI provided better predictive accuracy though weak (Figure 7). The linear and exponential models demonstrated weaker performance lignin predictions but achieved improved predictive accuracy with the validation set (R2 = 0.35) and higher RMSE and MAE values compared to the training set. The PLSR model exhibited consistently weak performance for lignin prediction in both training and validation sets. The RF model performed best on the training set but showed a substantial decline in predictive accuracy on the validation set (R2 = 0.35).

4. Discussion

4.1. Performance of VIs in Predicting Biomass Yields and Lignocellulose Concentrations

Our study demonstrates switchgrass biomass yield increases with 56 kg N ha−1 N, with the yields reaching levels comparable to those reported by Lee et al. [2] under the same N rate. Previous research also indicates that switchgrass yields often increase within the first 1–3 years of production with added N when the initial N fertility is low [2,57], a trend also observed in this study. This study reveals year-to-year variability in switchgrass production, with drought hampering yields in some years, such as in 2022. While the yields of the two cultivars were generally comparable, Liberty tended to outperform Independence, likely due to its greater tolerance to drought and cold [9].
Differences in VIs due to cultivar and N treatments were evident, with the NDVI, GNDVI, NDRE, GRVI, and SR showing distinct trajectory patterns. The phenological variations in the index corresponded to switchgrass growth patterns, with peak values in mid-season reflecting increased leaf greenness after canopy closure and lower values at the end of the season indicating stand senescence [58]. The higher index values observed with Liberty compared to Independence are consistent with previously documented field observations during the summer, where Liberty exhibited a greener canopy than Independence due to the inherent genetic traits of the cultivars [9,47]. The clear effect of N application on the VIs, especially the NDVI, GNDVI, and NDRE, suggests that these indices can be valuable in distinguishing nitrogen application effects in bioenergy switchgrass production systems. Previous studies have also explored spatio-temporal variations in chlorophyll and nitrogen content during the growing season, further emphasizing the strong relationship between VIs and key plant physiological parameters [59]. In separate studies, Amaral et al. [60] observed differences in sugarcane biomass at different N rates using the NDVI. Differences in VIs measured from sorghum fertilized with different N rates have also been observed [61].
The effective prediction using mid-summer-growing season (mid-July to mid-August) indices can be associated with the positive correlations observed between the Vis, biomass, and lignocellulose components. There were strong correlations (r > 0.7) between the VIs and biomass yields and moderate correlations (r > 0.5) between the VIs and cellulose with the mid-summer-growing season (mid-July to mid-August) indices, suggesting their usefulness for prediction. Moreover, the VIs tended to reach peak values around mid-July to mid-August, driven by maximum canopy cover and high leaf chlorophyll content, which further informs why the mid-season imageries were effective for predictions. Previous studies have shown that VIs representing peak greenness, measured from late June to early August, are the most reliable predictors of crop yield [62,63]. Previous studies on Miscanthus [33] and switchgrass [35] have also reported reliable biomass estimations using indices derived from mid-season imagery.
Our results indicate that the GNDVI (R2 = 0.86) was the strongest univariate predictor of biomass yield, followed by the NDRE (R2 = 0.75), as these indices demonstrated the lowest RMSE and MAE values with the highest R2 across the models. Hamada et al. [35] also found high switchgrass biomass prediction (R2 = 0.879) using the GNDVI derived from satellite imagery. In a separate study, Namoi et al. [33] found that the NDRE from mid-summer had the highest relationship with Miscanthus biomass yields (R2 = 0.82–0.97). In our study, combining the GNDVI and NDRE improved biomass yield estimation, explaining 88% of the variance in biomass yield in the validation set. These findings contrast those of Li et al. [61], who reported that the NDRE alone provided less accurate switchgrass biomass yield forecasts than other indices. However, their study relied on UAV imagery captured later in the growing season (August to November).
Accurate information on biomass cellulose, hemicellulose, and lignin concentrations is essential for effectively assessing biofuel yield potential and evaluating the economic feasibility of various conversion processes [64]. In our study, the NDRE was the best predictor for cellulose and explained 53% of the variance in cellulose. The favorable prediction of cellulose using the NDRE can be indirectly explained by the fact that the red-edge portion of the spectrum, where the NDRE is measured, captures changes in chlorophyll and nitrogen, both of which influence the plant’s photosynthesis, with higher photosynthesis often correlating with increased biomass, and consequently, greater cellulose accumulation in the cell walls [65]. The moderate prediction of hemicellulose using a multivariate combination of the GNDVI, NDVI, NDRE, and GRVI aligns with the existing literature, which shows that combining appropriate variable predictors can result in better prediction than single predictors [55]. However, the results showed weak predictive accuracy for lignin with the multivariate combination of the GNDVI, NDVI, and GRVI from mid-July imagery, offering limited predictive accuracy. Compared to lignin, the higher predictive accuracy for cellulose and hemicellulose can likely be attributed to their higher abundance as polysaccharides in plant biomass [61]. Moreover, cellulose and hemicellulose in biomass can be readily hydrolyzed into sugars for biofuel production, while lignin is more resistant to breakdown [66].

4.2. Model Performance for Biomass Yield and Lignocellulose Prediction

Linear and exponential regression appeared to be the most suitable models for biomass yield prediction, outperforming PLSR and RF models. The linear model displayed the most consistent predictive performance across both training and validation datasets. The PLSR model showed lower accuracy on the training data but performed well on the validation set. While the RF model achieved the best performance on the training data, its decreased accuracy on the validation set suggests potential overfitting. Overfitting in the RF model may be influenced by the small dataset size and model complexity. To mitigate overfitting, techniques such as hyperparameter tuning (e.g., reducing the maximum depth of trees, limiting the number of features considered at each split, or increasing the minimum sample size per leaf) can be effective [36,45]. Additionally, increasing the dataset size through further experiments or augmentation could enhance model robustness [45]. Linear and exponential models have also been reported as the best predictors for Miscanthus biomass yield [33], sorghum biomass yield [61], and corn yield [34]. Moreover, the model performance slightly improved after 10-fold cross-validation, with multivariate predictors resulting in higher prediction accuracy than univariate predictors.
The linear and exponential models maintained a relatively consistent performance across the training and validation sets for cellulose and hemicellulose prediction, though with moderate predictive power. This consistency suggests that the linear model generalizes well across both training and validation data. The PLSR showed low predictive accuracy for cellulose estimation in both training and validation sets, which suggests a weak correlation between the predicted and actual values. The RF model performed exceptionally well on the training set but showed reduced accuracy on the validation set. This suggests that RF may have overfitted the training data, capturing complex patterns that do not generalize well to new data. Linear and exponential models exhibited consistent yet limited performance in predicting lignin. As indicated by the consistently low R2 values across different models, the weak prediction of lignin suggests the need to examine additional indices using other sensors, such as hyperspectral imagery, for reliable lignin prediction. Our findings align with previous research showing that cellulose prediction models typically perform better than those for hemicellulose and lignin [67]. The performance drop in the RF model highlights the importance of selecting models that balance complexity with generalizability. Overall, methods that enable the reliable determination of switchgrass biomass yield and composition in the field can provide valuable estimates for biofuel yield estimates and enhance biomass utilization.

5. Conclusions

The accurate estimation of biomass and lignocellulose components is critical for optimizing the use of switchgrass in bioenergy production. Our study developed statistical models for predicting biomass yield, cellulose, hemicellulose, and lignin concentrations in bioenergy switchgrass using VIs calculated from UAV-based multispectral sensors. The findings show that the VIs collected in mid-July and mid-August exhibited stronger correlations with both biomass yield and cellulose. The biomass yield was strongly predicted using a multivariate combination of the GNDVI and NDRE in linear and exponential regression models. Cellulose was satisfactorily estimated using the univariate NDRE. The linear and exponential models outperformed the other models for predicting biomass, cellulose, and hemicellulose based on R2, RMSE, and MAE values. Hemicellulose was moderately predicted with a multivariate combination of the GNDVI, NDRE, NDVI, and GRVI, explaining 44% of the variability using either linear or exponential models. In contrast, lignin was weakly predicted across all models, regardless of univariate or multivariate predictors. These results demonstrate the potential of UAV-based VIs for the in-season estimation of biomass yield, cellulose, and hemicellulose concentrations. The models developed offer acceptable accuracy as a baseline for field-scale predictions. Future studies should focus on exploring additional VIs, particularly those derived from hyperspectral sensors, to improve predictions of lignocellulose components like lignin. Integrating data from multiple UAV-mounted sensors, such as hyperspectral, thermal, and LiDAR, could also provide complementary insights and enhance overall model accuracy. Additionally, validating models under diverse environmental conditions, soil types, and switchgrass ecotypes is critical for ensuring broader applicability. Field-scale implementation with automated workflows could streamline practical adoption and integrate UAV-based phenotyping into precision agriculture systems. Furthermore, economic and environmental assessments are crucial to evaluate the cost-effectiveness and sustainability of UAV-based phenotyping, supporting its large-scale adoption in bioenergy production.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/crops5010003/s1, Figure S1: Plot layout in Urbana, IL, showing switchgrass (Panicum virgatum L) cultivars Independence and Liberty and their respective annual N application (28 or 56 kg N ha−1) rates. Table S1: Switchgrass biomass yield estimation based on linear, exponential, partial least square, and random forest regression models using single and multiple predictor indices under sampling periods in July and August of 2021–2023.

Author Contributions

Conceptualization and design, D.L. and C.J.; methodology, D.W., C.J., C.Z., M.U.A., J.W.L., N.N. and D.L.; software, J.W.L., C.J. and D.W.; validation, D.W. and M.U.A.; formal analysis, D.W., N.N. and C.J.; field investigation, D.W., J.W.L., N.N. and D.L.; resources, D.L.; data curation, D.W, M.U.A., K.V. and C.J.; writing—original draft preparation, D.W.; writing—review and editing, all authors; visualization, D.W. and C.J.; supervision, D.L.; funding acquisition, D.L. All authors have read and agreed to the published version of the manuscript.

Funding

This work was funded by the U.S. Department of Energy, Bioenergy Technologies Office (DOE-BETO) under Award Number (DE-EE0008521). Any opinions, findings, conclusions, or recommendations expressed in this publication are those of the author(s) and do not necessarily reflect the views of the U.S. Department of Energy.

Data Availability Statement

The raw data supporting the conclusions of this article will be made available by the authors upon request.

Acknowledgments

The authors are grateful to the following University of Illinois Urbana-Champaign colleagues for technical field assistance: Tim Mies, Sunbong Jung, and Dana Landry.

Conflicts of Interest

The authors declare no conflicts of interest. The funders had no role in the design of this study, in the collection, analyses, or interpretation of the data, in the writing of this manuscript, or in the decision to publish the results.

References

  1. Vogel, K.P. Energy Production from Forages. J. Soil Water Conserv. 1996, 51, 137–139. [Google Scholar]
  2. Lee, D.K.; Aberle, E.; Anderson, E.K.; Anderson, W.; Baldwin, B.S.; Baltensperger, D.; Barrett, M.; Blumenthal, J.; Bonos, S.; Bouton, J.; et al. Biomass Production of Herbaceous Energy Crops in the United States: Field Trial Results and Yield Potential Maps from the Multiyear Regional Feedstock Partnership. GCB Bioenergy 2018, 10, 698–716. [Google Scholar] [CrossRef]
  3. Varvel, G.E.; Vogel, K.P.; Mitchell, R.B.; Follett, R.F.; Kimble, J.M. Comparison of Corn and Switchgrass on Marginal Soils for Bioenergy. Biomass Bioenergy 2008, 32, 18–21. [Google Scholar] [CrossRef]
  4. LaGory, K.E.; Cacho, J.F.; Zumpf, C.R.; Lee, D.; Feinstein, J.; Dematties, D.; Walston, L.J.; Namoi, N.; Negri, M.C. Bird Species Use of Bioenergy Croplands in Illinois, USA—Can Advanced Switchgrass Cultivars Provide Suitable Habitats for Breeding Grassland Birds? Sustainability 2024, 16, 4807. [Google Scholar] [CrossRef]
  5. U.S. Department of Energy; U.S. Department of Transportation; U.S. Department of Agriculture; U.S. Environmental Protection Agency. SAF Grand Challenge Roadmap: Flight Plan for Sustainable Aviation Fuel; U.S. Department of Energy: Washington, DC, USA, 2022; p. 128. [Google Scholar]
  6. U.S. Department of Energy. 2023 Billion-Ton Report: An Assessment of U.S. In Renewable Carbon Resources; Langholtz, M.H., Lead; Oak Ridge National Laboratory: Oak Ridge, TN, USA, 2024; ORNL/SPR-2024/3103. [Google Scholar] [CrossRef]
  7. McLaughlin, S.B.; De La Torre Ugarte, D.G.; Garten, C.T.; Lynd, L.R.; Sanderson, M.A.; Tolbert, V.R.; Wolf, D.D. High-Value Renewable Energy from Prairie Grasses. Environ. Sci. Technol. 2002, 36, 2122–2129. [Google Scholar] [CrossRef]
  8. Schmer, M.R.; Mitchell, R.B.; Vogel, K.P.; Schacht, W.H.; Marx, D.B. Spatial and Temporal Effects on Switchgrass Stands and Yield in the Great Plains. Bioenergy Res. 2010, 3, 159–171. [Google Scholar] [CrossRef]
  9. Vogel, K.P.; Mitchell, R.B.; Casler, M.D.; Sarath, G. Registration of ‘Liberty’ Switchgrass. J. Plant. Regist. 2014, 8, 242–247. [Google Scholar] [CrossRef]
  10. Casler, M.D. Biomass Yield Evaluation for Switchgrass Breeding: Seeded Swards vs. Transplanted Plots Yield Different Results. Bioenergy Res. 2021, 14, 1093–1105. [Google Scholar] [CrossRef]
  11. Anderson, E.K.; Parrish, A.S.; Voigt, T.B.; Owens, V.N.; Hong, C.-H.; Lee, D.K. Nitrogen Fertility and Harvest Management of Switchgrass for Sustainable Bioenergy Feedstock Production in Illinois. Energies 2020, 13, 1234. [Google Scholar] [CrossRef]
  12. Lemus, R.; Brummer, E.C.; Moore, K.J.; Molstad, N.E.; Burras, C.L.; Barker, M.F. Biomass Yield and Quality of 20 Switchgrass Populations in Southern Iowa, USA. Biomass Bioenergy 2002, 23, 433–442. [Google Scholar] [CrossRef]
  13. Lee, D.; Owens, V.N.; Boe, A.; Jeranyama, P. Composition of Herbaceous Biomass Feedstocks; South Dakota State University Publication, SGINC1-07: Brookings, SD, USA, 2007. [Google Scholar]
  14. Labbé, N.; Ye, X.P.; Franklin, J.A.; Womac, A.R.; Tyler, D.D.; Rials, T.G. Analysis of Switchgrass Characteristics Using Near Infrared Spectroscopy. Bioenergy Res. 2008, 3, 1329–1348. [Google Scholar] [CrossRef]
  15. Monono, E.M.; Nyren, P.E.; Berti, M.T.; Pryor, S.W. Variability in Biomass Yield, Chemical Composition, and Ethanol Potential of Individual and Mixed Herbaceous Biomass Species Grown in North Dakota. Ind. Crops Prod. 2013, 41, 331–339. [Google Scholar] [CrossRef]
  16. Zeng, L.; Chen, C. Using Remote Sensing to Estimate Forage Biomass and Nutrient Contents at Different Growth Stages. Biomass Bioenergy 2018, 115, 74–81. [Google Scholar] [CrossRef]
  17. Makepa, D.C.; Chihobo, C.H. Barriers to Commercial Deployment of Biorefineries: A Multi-Faceted Review of Obstacles Across the Innovation Chain. Heliyon 2024, 10, e32649. [Google Scholar] [CrossRef]
  18. Ayankojo, I.T.; Thorp, K.R.; Thompson, A.L. Advances in the Application of Small Unoccupied Aircraft Systems (sUAS) for High-Throughput Plant Phenotyping. Remote Sens. 2023, 15, 2623. [Google Scholar] [CrossRef]
  19. Li, F.; Piasecki, C.; Millwood, R.J.; Wolfe, B.; Mazarei, M.; Stewart, C.N., Jr. High-Throughput Switchgrass Phenotyping and Biomass Modeling by UAV. Front. Plant Sci. 2020, 11, 574073. [Google Scholar] [CrossRef]
  20. Impollonia, G.; Croci, M.; Ferrarini, A.; Brook, J.; Martani, E.; Blandinières, H.; Marcone, A.; Awty-Carroll, D.; Ashman, C.; Kam, J.; et al. UAV Remote Sensing for High-Throughput Phenotyping and for Yield Prediction of Miscanthus by Machine Learning Techniques. Remote Sens. 2022, 14, 2927. [Google Scholar] [CrossRef]
  21. Lu, B.; Dao, P.D.; Liu, J.; He, Y.; Shang, J. Recent Advances of Hyperspectral Imaging Technology and Applications in Agriculture. Remote Sens. 2020, 12, 2659. [Google Scholar] [CrossRef]
  22. Wang, T.; Liu, Y.; Wang, M.; Fan, Q.; Tian, H.; Qiao, X.; Li, Y. Applications of UAS in Crop Biomass Monitoring: A Review. Front. Plant Sci. 2021, 12, 616689. [Google Scholar] [CrossRef] [PubMed]
  23. Li, J.; Schachtman, D.P.; Creech, C.F.; Wang, L.; Ge, Y.; Shi, Y. Evaluation of UAV-Derived Multimodal Remote Sensing Data for Biomass Prediction and Drought Tolerance Assessment in Bioenergy Sorghum. Crop J. 2022, 10, 1016–1023. [Google Scholar] [CrossRef]
  24. Impollonia, G.; Croci, M.; Martani, E.; Ferrarini, A.; Kam, J.; Trindade, L.M.; Clifton-Brown, J.; Amaducci, S. Moisture Content Estimation and Senescence Phenotyping of Novel Miscanthus Hybrids Combining UAV-Based Remote Sensing and Machine Learning. GCB Bioenergy 2022, 14, 639–656. [Google Scholar] [CrossRef]
  25. Galli, G.; Horne, D.W.; Collins, S.D.; Jung, J.; Chang, A.; Fritsche-Neto, R.; Rooney, W.L. Optimization of UAS-Based High-Throughput Phenotyping to Estimate Plant Health and Grain Yield in Sorghum. Plant Phenome J. 2020, 3, e20010. [Google Scholar] [CrossRef]
  26. Kumar, C.; Mubvumba, P.; Huang, Y.; Dhillon, J.; Reddy, K. Multi-Stage Corn Yield Prediction Using High-Resolution UAV Multispectral Data and Machine Learning Models. Agronomy 2023, 13, 1277. [Google Scholar] [CrossRef]
  27. Ahamed, T.; Tian, L.; Zhang, Y.; Ting, K. A Review of Remote Sensing Methods for Biomass Feedstock Production. Biomass Bioenergy 2011, 35, 2455–2469. [Google Scholar] [CrossRef]
  28. Dutta, G.; Goswami, P. Application of Drone in Agriculture: A Review. Int. J. Chem. Stud. 2020, 8, 181–187. [Google Scholar] [CrossRef]
  29. Bannari, A.; Morin, D.; Bonn, F.; Huete, A.R. A Review of Vegetation Indices. Remote Sens. Rev. 1995, 13, 95–120. [Google Scholar] [CrossRef]
  30. Tsouros, D.C.; Bibi, S.; Sarigiannidis, P.G. A Review on UAV-Based Applications for Precision Agriculture. Information 2019, 10, 349. [Google Scholar] [CrossRef]
  31. Sanches, G.M.; de Carvalho, C.J.R.; Pereira, G.T.; Ribeiro, M.; de Souza, A.F.; dos Santos, J.C.; Silva, A.A. The Potential for RGB Images Obtained Using Unmanned Aerial Vehicle to Assess and Predict Yield in Sugarcane Fields. Int. J. Remote Sens. 2018, 39, 5402–5414. [Google Scholar] [CrossRef]
  32. Cholula, U.; da Silva, J.A.; Marconi, T.; Thomasson, J.A.; Solorzano, J.; Enciso, J. Forecasting Yield and Lignocellulosic Composition of Energy Cane Using Unmanned Aerial Systems. Agronomy 2020, 10, 718. [Google Scholar] [CrossRef]
  33. Namoi, N.; Jang, C.; Robins, Z.; Lin, C.-H.; Lim, S.-H.; Voigt, T.; Lee, D. Aerial Imagery Can Detect Nitrogen Fertilizer Effects on Biomass and Stand Health of Miscanthus × giganteus. Remote Sens. 2022, 14, 1435. [Google Scholar] [CrossRef]
  34. Jang, C.; Namoi, N.; Wolske, E.; Wasonga, D.; Behnke, G.; Bowman, N.D.; Lee, D.K. Integrating Plant Morphological Traits with Remote-Sensed Multispectral Imageries for Accurate Corn Grain Yield Prediction. PLoS ONE 2024, 19, e0297027. [Google Scholar] [CrossRef] [PubMed]
  35. Hamada, Y.; Zumpf, C.R.; Cacho, J.F.; Lee, D.; Lin, C.-H.; Boe, A.; Heaton, E.; Mitchell, R.; Negri, M.C. Remote Sensing-Based Estimation of Advanced Perennial Grass Biomass Yields for Bioenergy. Land 2021, 10, 1221. [Google Scholar] [CrossRef]
  36. Cacho, J.F.; Feinstein, J.; Zumpf, C.R.; Hamada, Y.; Lee, D.J.; Namoi, N.L.; Lee, D.; Boersma, N.N.; Heaton, E.A.; Quinn, J.J.; et al. Predicting Biomass Yields of Advanced Switchgrass Cultivars for Bioenergy and Ecosystem Services Using Machine Learning. Energies 2023, 16, 4168. [Google Scholar] [CrossRef]
  37. Zhang, C.; Kovacs, J.M. The Application of Small Unmanned Aerial Systems for Precision Agriculture: A Review. Precis. Agric. 2012, 13, 693–712. [Google Scholar] [CrossRef]
  38. Jansen, M.; Pinto, F.; Nagel, K.A.; van Dusschoten, D.; Fiorani, F.; Rascher, U.; Schneider, H.U.; Walter, A.; Schurr, U. Non-Invasive Phenotyping Methodologies Enable the Accurate Characterization of Growth and Performance of Shoots and Roots. In Genomics of Plant Genetic Resources; Tuberosa, R., Graner, A., Frison, E., Eds.; Springer: Dordrecht, The Netherlands, 2014; pp. 173–203. [Google Scholar]
  39. Lee, D.K.; Boe, A. Biomass Production of Switchgrass in Central South Dakota. Crop Sci. 2005, 45, 2583–2590. [Google Scholar] [CrossRef]
  40. Sanderson, M.A.; Agblevor, F.; Collins, M.; Johnson, D.K. Compositional Analysis of Biomass Feedstocks by Near Infrared Reflectance Spectroscopy. Biomass Bioenergy 1996, 11, 365–370. [Google Scholar] [CrossRef]
  41. Park, J.I.; Liu, L.; Ye, X.P.; Jeong, M.K.; Jeong, Y.-S. Improved Prediction of Biomass Composition for Switchgrass Using Reproducing Kernel Methods with Wavelet Compressed FT-NIR Spectra. Expert Syst. Appl. 2011, 38, 8649–8655. [Google Scholar] [CrossRef]
  42. Feng, X.; Yu, C.; Liu, X.; Chen, Y.; Zhen, H.; Sheng, K.; He, Y. Nondestructive and Rapid Determination of Lignocellulose Components of Biofuel Pellet Using Online Hyperspectral Imaging System. Biotechnol. Biofuels 2018, 11, 88. [Google Scholar] [CrossRef] [PubMed]
  43. Xu, Y.; Shrestha, V.; Piasecki, C.; Wolfe, B.; Hamilton, L.; Millwood, R.J.; Mazarei, M.; Stewart, C.N. Sustainability Trait Modeling of Field-Grown Switchgrass (Panicum virgatum) Using UAV-Based Imagery. Plants 2021, 10, 2726. [Google Scholar] [CrossRef] [PubMed]
  44. Wold, S.; Ruhe, H.; Wold, H.; Dunn, W.J. The Collinearity Problem in Linear Regression: The Partial Least Squares (PLS) Approach to Generalized Inverses. SIAM J. Sci. Stat. Comput. 1984, 5, 735–743. [Google Scholar] [CrossRef]
  45. Breiman, L. Random Forests. Mach. Learn. 2001, 45, 5–32. [Google Scholar] [CrossRef]
  46. Su, X.; Yan, X.; Tsai, C. Linear Regression. Wiley Interdiscip. Rev. Comput. Stat. 2012, 4, 275–294. [Google Scholar] [CrossRef]
  47. Lee, M.-S.; Casler, M.D.; Lee, D.K. Registration of ‘Independence’ Switchgrass. Plant Regist. 2024, 14, e20384. [Google Scholar] [CrossRef]
  48. Peñuelas, J.; Filella, I. Visible and Near-Infrared Reflectance Techniques for Diagnosing Plant Physiological Status. Trends Plant Sci. 1998, 3, 151–156. [Google Scholar] [CrossRef]
  49. Rouse, J.W.; Haas, R.H.; Schell, J.A.; Deering, D.W. Monitoring vegetation systems in the Great Plains with ERTS. NASA Spec. Publ. 1974, 351, 309. [Google Scholar]
  50. Gitelson, A.; Merzlyak, M.N. Spectral reflectance changes associated with autumn senescence of Aesculus hippocastanum L. and Acer platanoides L. leaves. Spectral features and relation to chlorophyll estimation. J. Plant Physiol. 1994, 143, 286–292. [Google Scholar] [CrossRef]
  51. Sripada, R.P.; Heiniger, R.W.; White, J.G.; Meijer, A.D. Aerial color infrared photography for determining early in-season nitrogen requirements in corn. Agron. J. 2006, 98, 968–977. [Google Scholar] [CrossRef]
  52. Birth, G.S.; McVey, G.R. Measuring color of growing turf with a reflectance spectrophotometer. Agron. J. 1968, 60, 640–649. [Google Scholar] [CrossRef]
  53. Afolabi, I.C.; Epelle, E.I.; Gunes, B.; Güleç, F.; Okolie, J.A. Data-Driven Machine Learning Approach for Predicting the Higher Heating Value of Different Biomass Classes. Clean Technol. 2022, 4, 1227–1241. [Google Scholar] [CrossRef]
  54. Van Soest, P.J.; Robertson, J.B.; Lewis, B.A. The Chemical Composition of Feeds in Relation to Digestion. J. Dairy Sci. 1991, 74, 3563–3578. [Google Scholar]
  55. R Core Team. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing. 2023. Available online: https://www.R-project.org/ (accessed on 2 February 2024).
  56. Tejera-Nieves, M.; Abraha, M.; Chen, J.; Hamilton, S.K.; Robertson, G.P.; Walker, J.B. Seasonal Decline in Leaf Photosynthesis in Perennial Switchgrass Explained by Sink Limitations and Water Deficit. Front. Plant Sci. 2023, 13, 1023571, Erratum in Front. Plant Sci. 2023, 14, 1204150. [Google Scholar] [CrossRef]
  57. Fike, J.H.; Pease, J.A.; Owens, V.N.; Farris, R.L.; Hansen, J.L.; Heaton, E.A.; Hong, C.O.; Mayton, H.S.; Mitchell, R.B.; Viands, D.R. Switchgrass Nitrogen Response and Estimated Production Costs on Diverse Sites. GCB Bioenergy 2017, 9, 1526–1542. [Google Scholar] [CrossRef]
  58. Mitchell, R.; Schmer, M. Switchgrass Harvest and Storage. In Switchgrass. Green Energy and Technology; Monti, A., Ed.; Springer: London, UK, 2012; pp. 73–100. [Google Scholar] [CrossRef]
  59. Catania, P.; Ferro, M.V.; Orlando, S.; Vallone, M. Grapevine and Cover Crop Spectral Response to Evaluate Vineyard Spatio-Temporal Variability. Sci. Hortic. 2024, 202, 113844. [Google Scholar] [CrossRef]
  60. Amaral, L.R.; Molin, J.P.; Portz, G.; Finazzi, F.B.; Cortinove, L. Comparison of Crop Canopy Reflectance Sensors Used to Identify Sugarcane Biomass and Nitrogen Status. Precis. Agric. 2015, 16, 15–28. [Google Scholar] [CrossRef]
  61. Li, J.; Shi, Y.; Veeranampalayam-Sivakumar, A.N.; Schachtman, D.P. Elucidating Sorghum Biomass, Nitrogen and Chlorophyll Contents with Spectral and Morphological Traits Derived from Unmanned Aircraft System. Front. Plant Sci. 2018, 9, 1406. [Google Scholar] [CrossRef] [PubMed]
  62. Mkhabela, M.; Bullock, P.; Raj, S.; Wang, S.; Yang, Y. Crop yield forecasting on the Canadian Prairies using MODIS NDVI data. Agric. For. Meteorol. 2011, 151, 385–393. [Google Scholar] [CrossRef]
  63. Bolton, D.K.; Friedl, M.A. Forecasting crop yield using remotely sensed vegetation indices and crop phenology metrics. Agric. For. Meteorol. 2013, 173, 74–84. [Google Scholar] [CrossRef]
  64. Okolie, J.A. Can Biomass Structural Composition Be Predicted from a Small Dataset Using a Hybrid Deep Learning Approach? Ind. Crops Prod. 2023, 203, 117191. [Google Scholar] [CrossRef]
  65. Loqué, D.; Scheller, H.V.; Pauly, M. Engineering of Plant Cell Walls for Enhanced Biofuel Production. Front. Plant Sci. 2015, 6, 288. [Google Scholar] [CrossRef]
  66. Azadi, P.; Khan, S.; Strobel, F.; Azadi, F.; Farnood, R. Hydrogen Production from Cellulose, Lignin, Bark, and Model Carbohydrates in Supercritical Water Using Nickel and Ruthenium Catalysts. Appl. Catal. B Environ. Energy 2012, 117–118, 330–338. [Google Scholar] [CrossRef]
  67. Jin, X.; Chen, X.; Shi, C.; Li, M.; Guan, Y.; Yu, C.Y.; Yamada, T.; Sacks, E.J.; Peng, J. Determination of Hemicellulose, Cellulose and Lignin Content Using Visible and Near Infrared Spectroscopy in Miscanthus sinensis. Bioresour. Technol. 2017, 241, 603–609. [Google Scholar] [CrossRef] [PubMed]
Crops 05 00003 g001
Figure 1. Seasonal trajectories of the normalized difference vegetation index (NDVI), green normalized difference vegetation index (GNDVI), normalized difference red-edge index (NDRE), green ratio vegetation index (GRVI), and simple ratio (SR) index as influenced by cultivar and N treatments. The indices were calculated from multispectral images taken over large-scale switchgrass plots at the Urbana Energy Farm, IL, during the 2021, 2022, and 2023 growing seasons. Error bars represent the mean standard errors of three replicates.
Figure 1. Seasonal trajectories of the normalized difference vegetation index (NDVI), green normalized difference vegetation index (GNDVI), normalized difference red-edge index (NDRE), green ratio vegetation index (GRVI), and simple ratio (SR) index as influenced by cultivar and N treatments. The indices were calculated from multispectral images taken over large-scale switchgrass plots at the Urbana Energy Farm, IL, during the 2021, 2022, and 2023 growing seasons. Error bars represent the mean standard errors of three replicates.
Crops 05 00003 g001
Crops 05 00003 g002
Figure 2. Pearson correlation coefficients between vegetation indices measured at different times (June, July, August, and September) and biomass yield and lignocellulose components. NDVI, normalized difference vegetation index; GNDVI, green normalized difference vegetation index; NDRE, normalized difference red-edge index; GRVI, green ratio vegetation index; SR, simple ratio index. Significance levels *, **, *** indicate p < 0.05, p < 0.01, p < 0.001, respectively.
Figure 2. Pearson correlation coefficients between vegetation indices measured at different times (June, July, August, and September) and biomass yield and lignocellulose components. NDVI, normalized difference vegetation index; GNDVI, green normalized difference vegetation index; NDRE, normalized difference red-edge index; GRVI, green ratio vegetation index; SR, simple ratio index. Significance levels *, **, *** indicate p < 0.05, p < 0.01, p < 0.001, respectively.
Crops 05 00003 g002
Crops 05 00003 g003
Figure 3. Scatter plots of predicted and actual switchgrass biomass yields based on linear, exponential, partial least square, and random forest regression models for the training and validation sets using the univariate GNDVI from mid-August imagery. The shaded areas indicate the 95% confidence intervals associated with model fit. The dots (blue points) represent the actual data points from the dataset, while the red lines indicate how well the model fits or predicts the data.
Figure 3. Scatter plots of predicted and actual switchgrass biomass yields based on linear, exponential, partial least square, and random forest regression models for the training and validation sets using the univariate GNDVI from mid-August imagery. The shaded areas indicate the 95% confidence intervals associated with model fit. The dots (blue points) represent the actual data points from the dataset, while the red lines indicate how well the model fits or predicts the data.
Crops 05 00003 g003
Crops 05 00003 g004
Figure 4. Scatter plots of predicted and actual switchgrass biomass yields based on linear, exponential, partial least square, and random forest regression models for the training and validation sets using a multivariate combination of the GNDVI and NDRE from mid-August imagery. The shaded areas indicate the 95% confidence intervals associated with model fit. The dots (blue points) represent the actual data points from the dataset, while the red lines indicate how well the model fits or predicts the data.
Figure 4. Scatter plots of predicted and actual switchgrass biomass yields based on linear, exponential, partial least square, and random forest regression models for the training and validation sets using a multivariate combination of the GNDVI and NDRE from mid-August imagery. The shaded areas indicate the 95% confidence intervals associated with model fit. The dots (blue points) represent the actual data points from the dataset, while the red lines indicate how well the model fits or predicts the data.
Crops 05 00003 g004
Crops 05 00003 g005
Figure 5. Scatter plots of predicted and actual cellulose concentration based on linear, exponential, partial least square (PLSR), and random forest (RF) regression models for the training and validation sets using the normalized difference red-edge index (NDRE) from mid-August imagery. The shaded areas indicate the 95% confidence intervals associated with model fit. The dots (blue points) represent the actual data points from the dataset, while the red lines indicate how well the model fits or predicts the data.
Figure 5. Scatter plots of predicted and actual cellulose concentration based on linear, exponential, partial least square (PLSR), and random forest (RF) regression models for the training and validation sets using the normalized difference red-edge index (NDRE) from mid-August imagery. The shaded areas indicate the 95% confidence intervals associated with model fit. The dots (blue points) represent the actual data points from the dataset, while the red lines indicate how well the model fits or predicts the data.
Crops 05 00003 g005
Crops 05 00003 g006
Figure 6. Scatter plots of predicted and actual hemicellulose concentration based on linear, exponential, partial least square (PLSR), and random forest (RF) regression models for the training and validation sets using a multivariate combination of the GNDVI, NDRE, NDVI, and GRVI from mid-July imagery. The shaded areas indicate the 95% confidence intervals associated with model fit. The dots (blue points) represent the actual data points from the dataset, while the red lines indicate how well the model fits or predicts the data.
Figure 6. Scatter plots of predicted and actual hemicellulose concentration based on linear, exponential, partial least square (PLSR), and random forest (RF) regression models for the training and validation sets using a multivariate combination of the GNDVI, NDRE, NDVI, and GRVI from mid-July imagery. The shaded areas indicate the 95% confidence intervals associated with model fit. The dots (blue points) represent the actual data points from the dataset, while the red lines indicate how well the model fits or predicts the data.
Crops 05 00003 g006
Crops 05 00003 g007
Figure 7. Scatter plots of predicted and actual acid detergent lignin (ADL) based on linear, exponential, partial least square (PLSR), and random forest (RF) regression models for the training and validation sets using a combination of the GNDVI, NDVI, and GRVI from mid-July imagery. The shaded areas indicate the 95% confidence intervals associated with model fit. The dots (blue points) represent the actual data points from the dataset, while the red lines indicate how well the model fits or predicts the data.
Figure 7. Scatter plots of predicted and actual acid detergent lignin (ADL) based on linear, exponential, partial least square (PLSR), and random forest (RF) regression models for the training and validation sets using a combination of the GNDVI, NDVI, and GRVI from mid-July imagery. The shaded areas indicate the 95% confidence intervals associated with model fit. The dots (blue points) represent the actual data points from the dataset, while the red lines indicate how well the model fits or predicts the data.
Crops 05 00003 g007
Table 1. Dates for multispectral imagery and biomass yield harvest, and the corresponding phenological stages for Liberty and Independence switchgrass cultivars during the 2021–2023 growing periods.
Table 1. Dates for multispectral imagery and biomass yield harvest, and the corresponding phenological stages for Liberty and Independence switchgrass cultivars during the 2021–2023 growing periods.
Activity202120222023Phenological Stages
Date, Month
Multispectral imagery16-Jun17-Jun14-JunStem elongation
19-Jul14-Jul19-JulBoot/heading stage
17-Aug18-Aug14-AugAnthesis
17-Sep19-Sep13-SepSeed development
Biomass yield harvest3-Dec14-Dec14-DecSenescence
Table 2. Summary of vegetation indices examined in this study.
Table 2. Summary of vegetation indices examined in this study.
IndexIndex NameFormulaSource
NDVINormalized difference vegetation index[(NIR − R)/(NIR + R)]Rouse et al. [49]
GNDVIGreen normalized difference vegetation index[(NIR − G)/(NIR + G)]Gitelson et al. [50]
NDRENormalized difference red-edge index[(NIR − RE)/(NIR + RE)]Gitelson et al. [50]
GRVIGreen ratio vegetation index(NIR/G)Sripada et al. [51]
SRSimple ratio index(NIR/R)Birth and McVey [52]
G = green band; R = red band; NIR = near-infrared band; RE = red-edge band.
Table 3. Monthly, total annual, and 30-year average precipitation (mm), as well as monthly and 30-year average temperatures (°C) for Urbana, Illinois (IL), from 2020 to 2023.
Table 3. Monthly, total annual, and 30-year average precipitation (mm), as well as monthly and 30-year average temperatures (°C) for Urbana, Illinois (IL), from 2020 to 2023.
Total Precipitation, mm Average Temperature (°C)
Year20212022202330-Year Avg20212022202330-Year Avg
Jan4511.24552.1−1.1−51.6−4
Feb35.130.254.451.3−5.6−2.12.9−1.7
Mar109113.586.469.97.55.75.24.4
Apr48.869.127.296.311.610.411.811.1
May89.795.555.1115.615.519.218.816.9
Jun166.63229.5105.923.52422.222.3
Jul105.96372.9110.222.924.123.723.9
Aug50.588.498.391.723.522.322.623
Sep78.257.468.883.320.818.819.919
Oct136.957.9121.279.815.51213.412.2
Nov30.545.718.588.94.45.66.25.2
Dec55.965.585.9644.7−0.24.1−1.7
Total952.1729.4763100911.911.212.710.9
GS Total539.7405.4351.860319.619.819.819.4
GS Total indicates the growing season (April–September) total; a 30-year average of monthly precipitation, and temperature for 30 years (1993–2023) at Urbana, IL. The weather data were obtained from the National Oceanic and Atmospheric Administration (NOAA) at Willard Airport Station, IL, USA.
Table 4. Effect of cultivar and N treatments on biomass yield, cellulose, hemicellulose, and lignin concentrations of bioenergy switchgrass across three years (2021–2023) at Urbana, IL. Data shown are means of three replicates (n = 12).
Table 4. Effect of cultivar and N treatments on biomass yield, cellulose, hemicellulose, and lignin concentrations of bioenergy switchgrass across three years (2021–2023) at Urbana, IL. Data shown are means of three replicates (n = 12).
YearCultivarN Rate (kg ha−1)Biomass Yield (Mg ha−1)Cellulose (g kg−1)Hemicellulose (g kg−1)Lignin (g kg−1)
2021Independence285.6a407a335ab71.6a
Independence568.1b419ab322a78.6a
Liberty286.4a424ab339b75.2a
Liberty567.5b430ab331ab76.7a
2022Independence283.5a397a365a61.7a
Independence565.5b413ab351a68.4a
Liberty286.0b404a373a62.8a
Liberty567.1c404a357a69.4a
2023Independence286.6c381a366b63.4a
Independence569.8ab401a347a72.3a
Liberty288.9b391a370b64.5a
Liberty5611.7a410ab360ab74.8ab
p-values (<0.05)Year<0.001<0.001<0.001<0.001
Cultivar<0.0010.0690.0030.458
N rate<0.0010.004<0.001<0.001
Year × cultivar0.0020.2820.9070.968
Year × N rate0.0220.3640.6160.409
Cultivar × N rate0.0660.2790.3450.666
Year × cultivar × N rate0.8810.7320.6570.657
Means followed by different letters in the same year are significantly different (Tukey’s HSD test p < 0.05).
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Wasonga, D.; Jang, C.; Lee, J.W.; Vittore, K.; Arshad, M.U.; Namoi, N.; Zumpf, C.; Lee, D. Estimating Switchgrass Biomass Yield and Lignocellulose Composition from UAV-Based Indices. Crops 2025, 5, 3. https://doi.org/10.3390/crops5010003

AMA Style

Wasonga D, Jang C, Lee JW, Vittore K, Arshad MU, Namoi N, Zumpf C, Lee D. Estimating Switchgrass Biomass Yield and Lignocellulose Composition from UAV-Based Indices. Crops. 2025; 5(1):3. https://doi.org/10.3390/crops5010003

Chicago/Turabian Style

Wasonga, Daniel, Chunhwa Jang, Jung Woo Lee, Kayla Vittore, Muhammad Umer Arshad, Nictor Namoi, Colleen Zumpf, and DoKyoung Lee. 2025. "Estimating Switchgrass Biomass Yield and Lignocellulose Composition from UAV-Based Indices" Crops 5, no. 1: 3. https://doi.org/10.3390/crops5010003

APA Style

Wasonga, D., Jang, C., Lee, J. W., Vittore, K., Arshad, M. U., Namoi, N., Zumpf, C., & Lee, D. (2025). Estimating Switchgrass Biomass Yield and Lignocellulose Composition from UAV-Based Indices. Crops, 5(1), 3. https://doi.org/10.3390/crops5010003

Article Metrics

Back to TopTop