Estimating Leaf Area Index of Wheat Using UAV-Hyperspectral Remote Sensing and Machine Learning ^†

Abstract

Hyperspectral remote sensing using Unmanned Aerial Vehicles (UAVs) provides accurate, near real-time, and large-scale spatial estimation of the leaf area index (LAI), a significant crop variable for monitoring crop growth. In the present study, the LAI of wheat crops was estimated using high-resolution UAV-borne hyperspectral data. The PLS (Partial Least Squares) regression combined with the VIP (Variable Importance in the Projection) was used for selecting the optimum indices as feature vectors to the Extreme Gradient Boosting (Xgboost) model for predicting LAI. Twelve of twenty-seven vegetation indices were selected to develop the prediction model. On validation against the in situ measured LAI values, the prediction model shows good accuracy with an R² of 0.71. The model was used to generate a spatial map showing the variability of the LAI. Accurate mapping of LAI from high-resolution hyperspectral UAV data using machine learning models facilitates near-real-time monitoring of crop health.

1. Introduction

The leaf area index (LAI) is a dimensionless variable measured as the total one-sided foliar area per unit area on the ground [1]. Since LAI is closely related to many plant canopies’ physical and biological processes such as photosynthesis, evapotranspiration, carbon uptake, energy balance, etc., it is essential to monitor timely, accurate, and temporal manner. However, the conventional approach to LAI estimation for larger fields is time-consuming, inaccurate, and labour-intensive [2]. Recent advances in remote sensing have made a significant breakthrough in estimating LAI more accurately, rapidly and reliably for larger areas. The capability of unmanned aerial vehicles (UAVs) to fly at low altitudes generates high-resolution images, which can be used to produce LAI maps showing high spatial variability. The potential of hyperspectral sensors mounted on UAVs has been explored successfully for estimating LAI utilizing various physical, empirical, machine learning or deep learning regression models [3,4]. Hybrid models developed using process-based radiative transfer (RT) models and machine learning models retrieve accurate measurements of crop traits [5]. Both spectral data [6] and spectral indices [7] are strongly related to various crop traits. After applying suitable feature extraction or transformation techniques, they serve as potential feature vectors for predicting these important crop variables [8,9]. Recently, frameworks made using a combination of preprocessing algorithms, feature selection models and machine learning/deep learning regression models were found to be able to successfully predict crop variables with better fit and accuracy [10]. On applying these optimized models to UAV-acquired data, operational mapping of crop traits can be achieved for smart nutrient input applications.

The LAI of wheat crops in the research farm of ICAR—Indian Agricultural Research Institute (IARI), New Delhi, was successfully estimated using different machine learning models from UAV image spectra and RT simulated spectra as input feature vectors [11,12]. In the present study, we attempted to estimate LAI using vegetation indices calculated from the UAV hyperspectral data in the spectral range of 400–1000 nm using the Extreme Gradient Boosting (Xgboost) regression model. A combined approach of PLSR with VIP scores was used for selecting optimal indices.

2. Materials and Methods

The flowchart showing LAI estimation using UAV data is shown in Figure 1. The major steps involved are the following: (i) Field Experimental Design, Data Acquisition and Pre-processing and (ii) Selection of Vegetation Indices and Model Development.

2.1. Field Experimental Design, Data Acquisition and Pre-Processing

The wheat experiments were conducted in 45 plots in the research farm of ICAR—Indian Agricultural Research Institute (IARI), New Delhi, during the rabi season (winter cropping season from October to April) of 2021–2022 (as shown in Figure 2a,b). The field was maintained with three replications of fifteen plots consisting of five graded nitrogen levels (0, 50, 100, 150, and 200 kgNha⁻¹) and three irrigation treatments (soil moisture sensor-based treatment-I1; crop water stress index (CWSI)-based treatment-I2; and conventional treatment-I3 (as shown in Figure 2c).

The five graded nitrogen levels were represented as N0, N1, N2, N3, and N4. The hyperspectral data, having 269 bands in the spectral range of 400–1000 nm, was captured on 17 March 2022 using a Nano-Hyperspec sensor (Headwall Photonics Inc., Bolton, MA, USA) mounted on a DJI Matrice 600 Pro hexacopter platform (DJI Sky City, Shenzhen, China) with a Ronin gimbal. Image acquisition was carried out between 11:00 AM and 12:00 PM on a bright sunny day. The mission planning was carried out using the flight planning UgCS mission planning software v3.4.609 (Universal Ground Control Software, SPH Engineering, Riga, Latvia). The drone was flown with a speed of 3 m/s at 21 m altitude to achieve a spatial resolution of 4 cm. The acquired data underwent a series of processing steps, including (i) radiometric calibration, (ii) orthorectification, (iii) mosaicking, (iv) spectral smoothing, and (v) crop area segmentation [6]. The LAI was also measured on the same day for each plot using an LAI-2000 plant canopy analyser (Li-Cor, Inc., Lincoln, NE, USA).

2.2. Selection of Vegetation Indices and Model Development

After a detailed review of the literature, a total of 27 vegetation indices, as listed in

Table 1. List of vegetation indices used in the study.

Sl No.	Vegetation Index	Formulation	Ref.
1	Angular insensitivity vegetation index (AIVI)	${\displaystyle \frac{{\mathrm{R}}_{445}\left({\mathrm{R}}_{720}+{\mathrm{R}}_{735}\right)-{\mathrm{R}}_{573}\left({\mathrm{R}}_{720}-{\mathrm{R}}_{735}\right)}{{\mathrm{R}}_{720}\left({\mathrm{R}}_{573}+{\mathrm{R}}_{445}\right)}}$	[13]
2	Chlorophyll index (ClI)	${\mathrm{R}}_{750}/\left({\mathrm{R}}_{700}+{\mathrm{R}}_{710}\right)-1$	[14]
3	Two-band Enhanced vegetation index (EVI2)	2.5[(R800 − R660)/(1 + R800 + 2.4 R660)]	[15]
4	Green chlorophyll index (CIgreen)	R780/R550 − 1	[16]
5	Modified simple ratio (mSR)	(R750 − R445)/(R705 − R445)	[17]
6	MERIS terrestrial chlorophyll index (MTCI2)	(R754 − R709)/(R709 − R681)	[18]
7	Modified triangular vegetation index (MTVI2)	${\displaystyle \frac{1.5\left[1.2\left({\mathrm{R}}_{800}-{\mathrm{R}}_{550}\right)-2.5\left({\mathrm{R}}_{670}-{\mathrm{R}}_{550}\right)\right]}{{\left[{\left(2{\mathrm{R}}_{800}+1\right)}^2-6{\mathrm{R}}_{800}+{5\left({\mathrm{R}}_{670}\right)}^{0.5}-0.5\right]}^{0.5}}}$	[19]
8	Modified chlorophyll absorption ratio indices (MCARI3)	$\left[\left({\mathrm{R}}_{750}-{\mathrm{R}}_{705}\right)-0.2\left({\mathrm{R}}_{750}-{\mathrm{R}}_{550}\right)\right]\times {\mathrm{R}}_{750}/{\mathrm{R}}_{705}$	[20]
9	Modified chlorophyll absorption ratio index 1 (MCARI1)	$1.2\left[2.5\left({\mathrm{R}}_{800}-{\mathrm{R}}_{670}\right)-1.3\left({\mathrm{R}}_{800}-{\mathrm{R}}_{550}\right)\right]$	[19]
10	Modified chlorophyll absorption ratio index 2 (MCARI2)	[1.5 × (2.5 × (R800 − R670)  −  1.3 × (R800 − R550))]/[Sqrt((2 × R800 + 1)2  −  6 × R800 +  5 × sqrtR670) − 0.5]	[19]
11	Modified Red Edge Normalized Difference Vegetation Index (MRENDVI)	(R750 − R705)/(R750 + R705 − 2 R445)	[17]
12	Modified soil adjusted vegetation index (MSAVI2)	(2 R810 + 1 − sqrt((2 R810 + 1)2 − 8 (R810 − R660)))/2	[21]
13	MCARI3/OSAVI	${\displaystyle \frac{\left[\left({\mathrm{R}}_{750}-{\mathrm{R}}_{705}\right)-0.2\left({\mathrm{R}}_{750}-{\mathrm{R}}_{550}\right)\right]{\mathrm{R}}_{750}/{\mathrm{R}}_{705}}{(1+0.16)({\mathrm{R}}_{800}-{\mathrm{R}}_{670})/({\mathrm{R}}_{800}+{\mathrm{R}}_{670}+0.16)}}$	[20,22]
14	Normalized difference vegetation index (NDVI [670,800])	(R800 − R670)/(R800 + R670)	[23]
15	Normalized Difference ND [705,750]	(R750 − R705)/(R750 + R705)	[24]
16	NDVI [550,750]	(R750 − R550)/(R750 + R550)	[25]
17	Optimized soil adjusted vegetation index (OSAVI)	(1 + 0.16)(R800 − R670)/(R800 + R670 + 0.16)	[22]
18	Optimized nonlinear vegetation index (ONLI)	1.05(0.6×R7982 − R728)/ (0.6*R7982 + R728 + 0.05)	[26]
19	Photon radiance index (PRI2)	(R531 − R570/(R531 + R570)	[27]
20	Ratio spectral index (RSI)	R760/R730	[28]
21	Red-edge chlorophyll index (CIre)	R780/R710 − 1	[16]
22	Structure Insensitive Pigment Index (SIPI)	(R800 − R445)/(R800 − R680)	[29]
23	Simple ratio index (SR) SR [800,680]	R800/R680	[17]
24	Triangular vegetation index (TVI)	0.5 [120(R750 − R550) − 200(R670 − R550)]	[30]
25	Transformed chlorophyll absorption in reflectance index (TCARI)	3 [(R700 − R670) − 0.2(R700 − R550)(R700/R670)]	[31]
26	TCARI/OSAVI	${\displaystyle \frac{3\left[\right({\mathrm{R}}_{700}-{\mathrm{R}}_{670})-0.2({\mathrm{R}}_{700}-{\mathrm{R}}_{550}\left)\right({\mathrm{R}}_{700}/{\mathrm{R}}_{670}\left)\right]}{(1+0.16)({\mathrm{R}}_{800}-{\mathrm{R}}_{670})/({\mathrm{R}}_{800}+{\mathrm{R}}_{670}+0.16)}}$	[22,31]
27	Vogelmann red edge index 1 (VREI1)	R740/R720	[32]

R_λ is reflectance at λ wavelength.

, were selected for LAI estimation. The PLSR-VIP was used for ranking these vegetation indices in order of their importance. The measured LAI was taken as an independent variable, and the Extreme Gradient Boosting (Xgboost) machine learning model was deployed for LAI prediction using these optimal vegetation indices. The optimal parameters, such as n_estimators, learning_rate, and max_depth, were selected using the RandomizedSearchCV function in Python 3.11.3 . The optimized model was validated against the in situ LAI measurements using root mean square error (RMSE), mean absolute error (MAE), and R-Square (R²). A 70/30 train/test data splitting was used for model training and validation. Finally, the validated model was applied to the pre-processed imagery for generating a spatial map showing the variability of LAI values.

3. Results

3.1. Optimizing Vegetation Indices Using PLSR-VIP

After pre-processing, the hyperspectral imagery was used for generating the spectral indices. The PLSR-VIP scores obtained for each spectral index were illustrated in a bar diagram as shown in Figure 3.

On applying a threshold of 1, twelve indices were finally selected as feature vectors for regression analysis. Among 27 indices, only 12 were strongly related to the measured LAI as indicated by the PLSR-VIP scoring approach. They are CIgreen, SIPI, AIVI, MRENDVI, TCARI/OSAVI, ONLI, RSI, MTCI2, VREI1, SR [800,680], NDVI [550,750], and MTVI2. The maps corresponding to these optimal indices are shown in Figure 4. All index values show a strong relationship with the N treatments.

3.2. LAI Mapping and Validation

The 12 characteristic indices and Xgboost model were deployed to develop an LAI prediction model. On validating against the in situ measurements, high prediction accuracies in terms of R² of 0.71, RMSE of 0.52, and MAE of 0.44, respectively, were obtained. The optimized values obtained for ‘learning_rate’, ‘max_depth’, and ‘n_estimators’ were 0.14, 1, and 87, respectively. The scatter diagram with validation results is shown in Figure 5a. The LAI generated by applying the optimized model to the 12 characteristic indices of UAV hyperspectral imagery is shown in Figure 5b. The LAI map was able to showcase the distinct variation in LAI with respect to N treatments. Most of the plots under the N0 treatment showed lower LAI values. Thus, the predicted map provides a clear visualization of spatial variation in LAI.

4. Discussion

Many studies have reported characteristic bands for LAI estimation using multiple feature selection models, but they varied with respect to regression modelling and variable selection approaches [33]. Optimized spectral indices consisting of red edge bands were proved to be crucial in monitoring the LAI of wheat [34]. More accurate LAI estimations were reported for vegetation indices using near-infrared bands. Compared to the red band, blue and green bands show a strong contribution to LAI estimation [2]. The indices optimized through PLSR-VIP were also dominated by these LAI-sensitive bands and thereby demonstrated the accurate estimation of LAI. Moreover, SR [800,680], MTVI2, and CIgreen, belonging to the optimized indices, have already proven their ability to estimate LAI accurately [1,2,33]. These highlight the importance of using a threshold of 1 for selecting these optimal indices, which were also reported earlier for VIP scores in LAI estimation [35]. The optimized indices and the generated LAI map show a similar trend in the LAI distribution with respect to ground truth values. The experimental plots corresponding to low-nitrogen treatments consist of many red-coloured pixels, suggesting lower LAI values and thereby a strong variation in estimated LAI with respect to N treatments was obtained. Since LAI is a strong indicator of crop health [11], these maps can be utilized for fertilizer management.

In the present study, Xgboost delivers an accurate prediction for wheat LAI using characteristic indices obtained from the PLSR-VIP selection approach. The model resulted in an R² value of 0.71, which is comparable with previous studies on LAI estimation using multiple machine learning models [4,36,37]. Among them, a few studies reported the outperformance of XGboost over other machine learning models [37,38,39]. In addition to XGboost, other advanced regressors like K-neighbours, extra trees, random forest, and support vectors also outperformed with higher R² value accuracy [40]. However, a comparison of multiple machine learning models and feature selection techniques will be performed in the future to select the best one for a more accurate prediction of LAI. Moreover, more data points from multiple growth stages and cropping seasons will be considered in the future for training ML models, to improve the prediction accuracy.

5. Conclusions

The selection of optimal vegetation indices was successfully achieved through PLSR-VIP scoring for predicting the LAI of wheat crops. A total of 27 indices was reduced to an optimal 12 characteristic indices sensitive to the LAI variability of wheat fields. These indices were used as input feature vectors for the Xgboost regression model and achieved a good prediction accuracy (R² of 0.71, RMSE of 0.52, and MAE of 0.44). The high-resolution LAI map generated from the UAV hyperspectral data using this optimized model offers a non-destructive, rapid, and accurate alternative to conventional approaches.