Machine learning techniques for accessing the chlorophyll content and fluorescence in maize based on UAV multispectral sensor

Abstract

Canopy chlorophyll content (CCC) and chlorophyll fluorescence (ChlF) are key indicators of crop physiological status and productivity. Advances in unmanned aerial vehicle (UAV)-based multispectral sensing, integrated with machine learning (ML), have opened new possibilities for precise and non-destructive monitoring of these traits. This study evaluated the performance of thirty vegetation indices (VIs) and six ML algorithms for estimating CCC and ChlF in maize during kharif season and rabi season at Tamil Nadu Agricultural University, Coimbatore, India. Regression analyses revealed that red-edge indices, such as RECI, NDRE, MTCI, and LCI, consistently outperformed greenness indices. Among them, RECI and NDRE achieved the highest R² values (0.893 for CCC and 0.829 for ChlF) and the lowest RMSE values (0.058 for CCC and 0.021 for ChlF) during kharif season. In contrast, during rabi season, MTCI for CCC and LCI for ChlF showed the best performance, with R² values of 0.783 and 0.883, and RMSE values of 0.081 and 0.018, respectively. Among machine learning algorithms, Partial Least Squares Regression (PLSR) achieved the highest predictive accuracy for CCC (R² = 0.891, RMSE = 0.063, MAE = 0.045) and ChlF (R² = 0.889, RMSE = 0.017, MAE = 0.013) during kharif season. However, Random Forest (RF) outperformed during rabi season, achieving the highest accuracy for CCC (R² = 0.623, RMSE = 0.083, MAE = 0.066) and ChlF (R² = 0.867, RMSE = 0.015, MAE = 0.011), respectively. In contrast, K-Nearest Neighbours (KNN) consistently underperformed in predicting both CCC and ChlF across kharif season and rabi season. Overall, this study demonstrated the potential of integrating UAV-based multispectral vegetation indices (VIs) with ML algorithms for accurate and scalable estimation of maize CCC and ChlF under varying seasonal conditions, highlighting their applicability for precision crop monitoring and sustainable agricultural management. However, further validation across different crop seasons and multiple locations remain limited.

Graphical Abstract

Introduction

Chlorophyll, the primary green pigment present in plant leaves, plays a fundamental role in photosynthesis by converting solar energy into biochemical energy required for carbon assimilation and plant growth [56, 75, 85]. As a critical biochemical indicator, chlorophyll content is strongly associated with crop productivity, nutrient dynamics, photosynthetic efficiency, and overall plant physiological status [80, 89, 124]. In addition to chlorophyll content, chlorophyll fluorescence (ChlF), particularly the maximum quantum efficiency of photosystem II represented by Fv/Fm, has emerged as a reliable non-destructive indicator for evaluating photosynthetic performance and detecting environmental stress in crops. The Fv/Fm parameter reflects the efficiency of energy transfer within photosystem II (PSII) and provides important insights into plant responses to stresses such as drought, nutrient deficiency, salinity, and temperature fluctuations [131]. Therefore, simultaneous assessment of canopy chlorophyll content (CCC) and chlorophyll fluorescence is essential for understanding crop physiological health and improving precision crop management practices. Conventional methods for determining chlorophyll content and chlorophyll fluorescence mainly involve destructive laboratory-based techniques, including solvent extraction, spectrophotometric analysis, and pulse-amplitude modulated (PAM) fluorometry. Although these methods provide accurate quantification, they are destructive, labour-intensive, time-consuming, and unsuitable for large-scale field monitoring [92, 110]. Moreover, PAM fluorometers require leaf dark adaptation prior to measurement, making the process difficult and impractical for high-throughput agricultural applications. Consequently, the increasing demand for rapid, non-invasive, and scalable crop monitoring techniques has accelerated the adoption of remote sensing technologies in precision agriculture.

The integration of remote sensing technologies into precision agriculture facilitates efficient management of agricultural inputs, optimization of resource utilization, and improvement of food security under changing climatic conditions [38, 90, 125]. In recent years, unmanned aerial vehicles (UAVs) equipped with multispectral sensors have gained considerable attention as effective tools for crop monitoring because of their high spatial resolution, temporal flexibility, rapid deployment, and cost-effectiveness [23, 134]. Vegetation indices (VIs) derived from unmanned aerial vehicle (UAV)-based multispectral imagery effectively capture crop spectral responses associated with chlorophyll concentration, biomass accumulation, and physiological stress [89, 122]. In particular, red-edge and near-infrared spectral bands have demonstrated high sensitivity for estimating canopy chlorophyll content and photosynthetic activity in crops [17, 118]. Furthermore, remote sensing approaches for chlorophyll fluorescence retrieval have opened new opportunities for large-scale and non-destructive monitoring of crop physiological conditions and photosynthetic performance [143].

Recent advances in UAV-based remote sensing integrated with artificial intelligence (AI) and machine learning (ML) techniques have significantly enhanced the monitoring of crop physiological and biochemical traits under field conditions. UAV-mounted multispectral and hyperspectral sensors provide high spatial and temporal resolution data for estimating chlorophyll content, nitrogen status, biomass, stress responses, and crop productivity in precision agriculture [20, 66]. ML approaches such as Multiple Linear Regression (MLR), Support Vector Regression (SVR), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) have demonstrated strong capability in modeling nonlinear relationships between spectral reflectance and vegetation biochemical parameters [8, 66]. Several studies have successfully estimated CCC using UAV multispectral imagery, particularly utilizing red-edge and near-infrared spectral regions because of their high sensitivity to chlorophyll dynamics [66, 67]. Similarly, ChlF and solar-induced fluorescence (SIF) have emerged as reliable indicators of photosynthetic activity and plant stress due to their close association with photosystem functioning and energy dissipation processes [86]. Recent reviews have further highlighted the growing role of AI and ML algorithms in fluorescence retrieval, reconstruction, and crop physiological monitoring applications [123]. However, despite substantial progress in UAV remote sensing and AI-driven crop monitoring, important research gaps still remain unresolved. Most previous studies have primarily focused on the estimation of individual physiological traits such as chlorophyll content, nitrogen concentration, biomass, or stress indicators independently [66, 135]. Studies integrating UAV-based multispectral sensing with AI techniques for simultaneous estimation of both CCC and ChlF remain comparatively limited. Furthermore, many investigations rely predominantly on hyperspectral datasets, which are computationally intensive, expensive, and less feasible for large-scale operational agricultural applications [135]. In addition, challenges related to model generalizability, overfitting, independent validation, and integration of multiple physiological indicators under varying field conditions continue to be major concerns in ML-based crop monitoring studies [66, 135].

Despite recent advances in UAV-based remote sensing and machine learning, studies on the simultaneous estimation of CCC and ChlF using UAV-based multispectral imagery remain limited. Rather than proposing a new machine learning methodology, this study provides a practical comparative evaluation of multiple ML algorithms for the simultaneous estimation of CCC and ChlF in maize across the kharif and rabi seasons using UAV-based multispectral data, thereby supporting the selection of suitable predictive models for precision crop monitoring under field conditions.

Material and methods

Location of the study area and studied field

The field experiment was conducted during the kharif season and rabi season of 2023–2024 in a maize-growing plot located at the Eastern Block Farm of Tamil Nadu Agricultural University (TNAU), Coimbatore. The geographical location of the study site is presented in Fig. 1. The farm is situated in the Western Agroclimatic Zone of Tamil Nadu, positioned at a latitude of 11°N and a longitude of 77°E, with an elevation of 426.7 m above mean sea level. The study was conducted using a randomized block design comprising fourteen treatments with three replications. The hybrid maize variety CoH(M) 6, released by TNAU, was used as the test crop. A total of 126 geo-tagged sampling points were collected during each of the kharif and rabi seasons, resulting in 252 observations that were utilized for model development, training, and evaluation.

Fig. 1.

Location of the experimental farm

Meteorological observations during the experimental period were collected from the Agro-Climate Research Centre, TNAU, Coimbatore (Fig. 2). The average maximum and minimum temperatures recorded during the 24th and 51 st standard meteorological weeks were 34.3 °C and 21.7 °C, respectively. The total rainfall received throughout the cropping period amounted to 5.81 cm, with the majority (67.36%) falling in November. The highest weekly precipitation, measuring 158 mm, occurred in the 45th standard week. Relative humidity levels varied, with a peak mean of 92.10% noted in the 45th week and the lowest mean of 44.80% during the 24th week.

Fig. 2.

Variation in weather parameters during crop growing period

For soil characterization, random surface samples were collected from a depth of 0–30 cm at the experimental site and combined to form a composite sample. Analysis revealed that the soil had a sandy clay loam texture with an alkaline pH of 8.27 and a high salinity level, indicated by an electrical conductivity of 0.17 dS m⁻¹. The organic carbon content was moderate at 4.68 g kg⁻¹. Nutrient profiling showed that the soil was deficient in available nitrogen (238 kg ha⁻¹) and phosphorus (22.97 kg ha⁻¹), while the available potassium content was comparatively high at 602.43 kg ha⁻¹.

Data collection

Image acquisition

A DJI Phantom 4 (Da-JiangInnovations, Shenzhen, China) Multispectral imaging system (specifications provided in Table 1) was integrated with a quadcopter drone for data collection during the study. This UAV platform was selected for its vertical take-off and landing (VTOL) capability, which makes it particularly suitable for confined agricultural environments. Flight operations were automated using predefined waypoints to ensure consistent image acquisition. The mission parameters—including waypoints, altitude, flight path, and speed- were programmed using DJI Ground Station Pro 2.0 (DJI GS Pro), a dedicated flight planning software.

Table 1.

Description of flight information and UAV sensor specifications

Platform	:	Quadcopter
Flight speed	:	4 m s−1
Altitude	:	30 m
Overlap rate	:	75%
Ground Sample Distance (GSD)	:	1.52 cm per pixel (per band)
Multispectral sensor	:	DJI- Phantom 4
Image spectral band	:	Blue, Green, Red, Red edge, Near IR
Characteristic wavelength	:	Name	Central wavelength (nm)	Wavelength range (nm)
		Blue (B)	450	± 16
		Green (G)	560	± 16
		Red (R)	650	± 16
		Red edge (RE)	730	± 16
		Near IR (NIR)	840	± 26
Image resolution (pixels)	:	1600 × 1300

Aerial imaging missions were carried out under clear sky conditions between 11:00 a.m. and 12:00 noon to minimize variability due to solar angle and atmospheric interference. Multispectral image data were collected using the UAV’s onboard camera system and subsequently downloaded for processing. Georeferencing of the captured imagery was performed using ground control points (GCPs), which served as spatial reference markers to enhance the positional accuracy of the data.

Radiometric calibration was conducted using a calibrated reflectance panel (CRP) placed within the survey plot. The CRP was carefully positioned on clean, unobstructed ground and imaged before each UAV flight according to the calibration manual. The multispectral camera was oriented to fully capture and illuminate the panel, ensuring consistent reflectance across all spectral bands. These pre-flight CRP images were used to normalize the reflectance values in the multispectral dataset, enabling accurate representation of ground surface conditions. Post-flight processing included radiometric corrections based on CRP readings to account for fluctuations in ambient light and atmospheric variability. The software applied corrections using metadata on solar intensity and atmospheric properties, thus compensating for potential distortions caused by variable illumination or airborne particles. This ensured that the final reflectance outputs from the imagery accurately reflected on-ground spectral characteristics.

Ground-based data collection

The tasselling and silking stages in maize represent the most critical periods for acquiring physiological data. During these stages, ground truth measurements were collected for essential parameters such as CCC and ChlF. At the same time, UAV-acquired multispectral images were utilized to derive and validate various VIs. To maintain measurement accuracy and ensure reliable validation, ten data points were recorded per plot for each parameter under observation.

Canopy Chlorophyll Contents (CCC)

To estimate CCC, 150 mg of leaf tissue was sampled from the composite leaves of each treatment. The samples were immersed in 50 mL of 80% acetone in conical flasks and kept in dark conditions for 24 h to allow complete pigment extraction. Following the extraction period, the chlorophyll solution was filtered using Whatman No. 1 filter paper, and the absorbance of the filtrate was measured at 645 nm and 663 nm using a calibrated colourimeter. The concentrations of total chlorophyll were calculated and expressed in mg g⁻¹ fresh weight using the equations described by Mahantheshwara et al. 73:

1

where,

D - 645 = Optical density at 645 nm

D - 663 = Optical density at 663 nm

V = Final volume of 80 % acetone chlorophyll extract in ml

W = Fresh weight in g of the corresponding number of fresh leaves used in the extraction of chlorophyll

These calculations provided a quantitative assessment of chlorophyll pigments in maize foliage, which were subsequently used for validating UAV-derived VIs and ML-based predictions.

Chlorophyll fluorescence (ChlF)

The fluorescence meter [Plant PAM-210 (Teaching PAM), Heinz Walz, Germany] was used to quantify ChlF. The essential fluorescence characteristics were measured: Fo (initial fluorescence), Fv (variable fluorescence), Fm (maximal fluorescence), and the ratio of Fv/Fm. Fv/Fm = Variable fluorescence/Maximum fluorescence Eq. 1 The proportion of quantum yield in response to a high degree of photosynthesis is depicted by the Fv/Fm ratio [76].

Data processing

Image mosaicking

The multispectral imagery collected during the UAV flights was processed using Pix4Dmapper 4.5.6 software. Following georeferencing, the raw data underwent processing and analysis to generate an orthomosaic. Multiple overlapping images were stitched together to create this orthomosaic, ensuring high positional accuracy. Reflectance values within the multispectral dataset were standardized by applying calibration derived from the calibrated reflectance panel images acquired during the survey. The overall workflow of multispectral image processing is presented in Fig. 3.

Fig. 3.

Flowchart of multispectral image processing

Vegetation indices estimation

VIs served as the primary analytical tool for interpreting aerial imagery. Initially, processed UAV images were imported into ArcGIS 10.6 to generate VIs. The processed spectral data were subsequently used to compute various VIs based on established formulas. VIs are essential for evaluating crop growth vigour, yield potential, and nutrient status. To distinguish maize from bare soil, pixels exclusively containing maize or soil were identified. Reflectance analysis at 720 nm revealed a clear separation, with soil exhibiting lower reflectance values. A threshold of 4% reflectance was applied, classifying pixels above this value as vegetation and those below as soil. Only maize pixels were retained for subsequent VI calculations, which included indices such as DVI, ENDVI, ExG, ExGR, ExR, GCI, GDVI, GNDVI, GOSAVI, GRVI, GSAVI, IPVI, LCI, MSR, MTCI, NDI, NDRE, NDVI, NGBDI, NGRDI, NPCI, OSAVI, RECI, REOSAVI, RGI, SIPI, SRI, TNDVI, VARIg, and VARIr (Table 2). These indices were particularly valuable for predicting parameters such as leaf area index and CCC. Using ground truth coordinates, the pixel-by-value tool in ArcGIS 10.6 was applied to extract VI values from the spectral data. The retrieved VI dataset was then subjected to statistical analysis.

Table 2.

Vegetation indices were used in this study

Sl. No	Vegetation index	Formulations	Applications	References
1	Difference Vegetation Index (DVI)	DVI = NIR–R	Prediction of green LAI of crop canopies, Comparison of different reflectance indices for vegetation analysis	[51, 136, 141]
2	Enhanced Normalized Difference Vegetation Index (ENDVI)		Estimation of canopy chlorophyll content	[84]
3	Excess Green (ExG)		Estimation of canopy chlorophyll content	[84]
4	Excess Green Red (ExGR)	ExGR = ExG-ExR	Estimation of canopy chlorophyll content	[84]
5	Excess Red (ExR)		Estimation of canopy chlorophyll content	[84]
6	Green Chlorophyll Index (GCI)		Estimation of leaf area index and green leaf biomass	[44, 136]
7	Green Difference Vegetation Index (GDVI)	GDVI = NIR-G	Estimate productivity and assess phylogenetic heritability	[51, 61]
8	Green Normalized Difference Vegetation Index (GNDVI)		Estimation of vegetation fraction, Estimation of productivity, assessment of phylogenetic heritability, discrimination of field-grown olive cultivars	[51, 63]
9	Green Optimized Soil Adjusted Vegetation Index (GOSAVI)		Extraction of land cover information, estimation of chlorophyll content	[39, 52]
10	Green Red Vegetation Index (GRVI)		Estimation of canopy chlorophyll content	[84]
11	Green Soil Adjusted Vegetation Index (GSAVI)		Estimation of canopy chlorophyll content	[21]
12	Infrared percentage vegetation index (IPVI)		Estimation of chlorophyll	[43]
13	Leaf Chlorophyll Index (LCI)		Estimation of canopy chlorophyll content	[84]
14	Modified Simple Ratio (MSR)		Estimation of canopy chlorophyll content	[25]
15	MERIS Terrestrial Chlorophyll Index (MTCI)		Estimation of canopy chlorophyll content	[32, 33]
16	Normalized Difference Index (NDI)		Vine diseases detection	[133]
17	Normalized Difference Red Edge Index (NDRE)		Detect Differences in Vegetation and Chlorophyll Content., Stress Detection, Growth monitoring	[11, 14, 77]
18	Normalized Difference Vegetation Index (NDVI)		Estimation of vegetation biomass	[19]
19	Normalized Green and Blue Difference Index (NGBDI)		Estimation of chlorophyll	[50]
20	Normalized Green Red Difference Index (NGRDI)		Monitoring of crop biomass	[41, 51]
21	Normalized Pigment Chlorophyll Ratio Index (NPCI)		Estimation of plant biomass, pigment composition and photosynthetic efficiency	[35, 94, 95]
22	Optimized Soil-Adjusted Vegetation Index (OSAVI)		Estimation of chlorophyll content	[101]
23	Red Edge Chlorophyll Index (RECI)		Estimation of chlorophyll content	[72, 147]
24	Red edge optimized soil Regulating Vegetation Index (REOSAVI)		Estimation of chlorophyll content	[59]
25	Red Green Index (RGI)		Vine diseases detection	[133]
26	Structure Insensitive Pigment Index (SIPI)		Estimation of chlorophyll content, Mapping the sugarcane	[106]
27	Simple Ratio Index (SRI)		Evaluation of the saturation property of vegetation indices	[22, 117]
28	Transformed Normalized Difference Vegetation Index (TNDVI)		Estimation of chlorophyll content	[119]
29	Green Band Atmospheric Impedance Vegetation Index (VARIg)		Estimation of chlorophyll content	[40]
30	Red-Band Atmospheric Impedance Vegetation Index (VARIr)		Estimation of chlorophyll content	[40]

Statistical analysis

Statistical validation of the results was performed using R Studio (version 4.2.2). Pearson’s correlation analysis was first conducted to identify the vegetation index (VI) most strongly associated with the field-measured CCC and ChlF values, as indicated by the correlation coefficient (r), calculated using Eq. (2). Model performance was evaluated by calculating the coefficient of determination (R²), the root mean square error (RMSE) and the mean absolute error (MAE). In the regression analysis, VIs served as independent (predictor) variables, while field-measured CCC and ChlF values acted as dependent (predicted) variables. A higher R² and lower RMSE indicated greater predictive capability of the model. The R², RMSE and MAE metrics were derived using Eqs. (3) and (4), respectively.

2

3

4

5

In this case, P_i and Q_i stand for the estimated and measured values, respectively. In the meantime, the average estimated and measured values were represented by the symbols P and Q, respectively. The sample number is denoted by the variable n.

Machine learning model algorithms

Various ML algorithms, including RF, PLSR, SVM, XGB, KNN and ANN, were applied to model nonlinear relationships between selected VIs and maize CCC as well as ChlF values. In this procedure, CCC and ChlF were considered as dependent variables, while multispectral VIs were used as independent variables. The dataset was divided into two subsets, with 70% allocated for training to calibrate the models and 30% reserved for testing to evaluate predictive performance [62, 103]. This division of data reduces the risk of overfitting and improves the ability of the models to generalize to unseen data [103]. The optimal number of factors was determined through validation approaches such as K (5) -fold cross-validation. In the present study, multiple independent calibration–validation combinations were tested, and the best-performing model was selected based on scatter plots, the highest coefficient of determination (R²) and the lowest root mean square error (RMSE).

Random Forest (RF)

Random forest is developed by Breiman [16], is an ensemble learning algorithm that combines the strength of decision trees with the bagging techniques [12, 88, 100]. RF is recognized for its robustness in the presence of noisy data and its distinct advantage in curbing overfitting. A forest of decision trees is created, with each tree independently predicting outcomes using only a subset of the data. The final prediction was then determined by a majority vote among all the independent trees. The model was fitted using the randomForest package in R and was highly effective for addressing high-dimensional data in remote sensing and artificial intelligence applications. It can handle thousands of input variables without overfitting, making it ideal for feature selection. Additionally, RF captures complex, non-linear relationships between features and the target variable, enhancing its utility in these fields [24, 145].

The RF regression model was implemented using the randomForest package in R with K (5)-fold cross-validation to improve model robustness and reduce overfitting risk. Hyperparameter optimization included 500 decision trees (ntree = 500), while the optimal number of variables randomly selected at each split (mtry) was automatically determined during model tuning. Variable importance analysis was incorporated to identify the relative contribution of vegetation indices. The ensemble bagging approach, together with random feature selection, enhanced prediction stability and reduced model variance.

Partial Least Squares Regression (PLSR)

Partial least squares regression is a multivariate statistical technique that integrates features of principal component analysis (PCA) and multiple regression, making it particularly effective when predictors are highly collinear or exceed the number of observations [2, 130]. PLSR reduces the dimensionality of predictor variables into orthogonal latent components that maximize covariance with the response variable, thereby enhancing predictive accuracy compared to traditional regression [37]. In agricultural and remote sensing applications, PLSR has been widely employed for analyzing multispectral and hyperspectral reflectance data, enabling robust estimation of crop traits, photosynthetic parameters, and VIs from noisy and high-dimensional datasets [81, 130]. In this study, PLSR was implemented in R using the pls package to model the relationship between reflectance spectra and crop physiological variables.

The PLSR model was implemented using the pls package in R with K (5)-fold cross-validation to determine the optimum number of latent variables (ncomp) that minimized prediction error. Predictor variables were centred and scaled prior to model development to improve numerical stability. The optimal number of latent components was selected based on the lowest cross-validation RMSE, enabling efficient handling of multicollinearity while improving model robustness and predictive accuracy.

Support Vector Machine (SVM)

Support vector machine includes a variation called support vector regression that is intended for regression problems. It is particularly good at estimating continuous variables such as biophysical parameters and approximating functions [18]. The radial basis kernel function (RBF) was used for SVR in this work, and model tuning was used to identify the ideal values for gamma (γ) and the regularisation parameter (C). SVM is often used in remote sensing for both classification and regression tasks and is implemented using libraries such as Libsvm. It provides flexibility in selecting alternative kernel functions and fine-tuning parameters for the best possible model performance [42, 145].

The SVM regression model was implemented using the caret and e1071 packages in R with K (5)-fold cross-validation. A radial basis function (RBF) kernel was adopted, and all predictor variables were centred and scaled prior to model development. Hyperparameter optimization was performed using tuneLength = 10, allowing automatic optimization of the regularization parameter (C) and kernel parameter (γ) to maximize predictive performance while minimizing overfitting.

Extreme Gradient Boosting (XGB)

Chen and Guestrin's [26] XGB is an advanced and effective ensemble learning technique. By adding second derivatives via the Taylor expansion, it greatly improves upon conventional gradient boosting techniques and can better capture complex connections in data. Kearns and Valiant's [55] definitions of strong and weak learnable issues served as inspiration for XGB, which combines several weak learners through additive learning to produce a reliable and accurate prediction model. The XGB package in R was used to build this method, offering a full feature set that includes significant feature selection analysis based on metrics including gain, cover, and frequency. In general, XGB was quite good at managing intricate data patterns and obtaining excellent prediction performance in different ML [69].

The XGB regression model was implemented using the xgboost package in R with K(5)-fold cross-validation to improve model robustness and reduce overfitting risk. Hyperparameter optimization included a learning rate (eta) of 0.05, maximum tree depth (max_depth) of 4, minimum child weight of 3, subsampling ratio (subsample) of 0.7, and column sampling ratio (colsample_bytree) of 0.7. L1 (reg_alpha = 0.1) and L2 (reg_lambda = 1.0) regularization parameters, along with gamma (gamma = 0.1), were incorporated to control model complexity and improve generalizability. Early stopping rounds of 10 were applied during training to prevent overfitting.

K-Nearest Neighbours (KNN)

K-nearest neighbours is a non-parametric, instance-based learning algorithm that classifies or predicts outcomes based on the proximity of a query point to its k nearest neighbours in the feature space [7, 30]. Unlike parametric models, KNN makes no prior assumptions about the underlying data distribution, which makes it flexible for diverse datasets. The algorithm relies on distance metrics, commonly Euclidean distance, to identify neighbours and infer responses through majority voting (classification) or averaging (regression) [36]. In agricultural and remote sensing studies, KNN has been applied for vegetation classification, soil property estimation, and prediction of crop traits from spectral reflectance data due to its simplicity and robustness in handling nonlinear relationships [12]. In this study, KNN was implemented in R using the caret package to evaluate its performance in predicting crop physiological parameters from multispectral reflectance data.

The KNN regression model was implemented using the caret package in R with K (5)-fold cross-validation. Predictor variables were centred and scaled before model fitting to ensure equal contribution of all vegetation indices during Euclidean distance calculations. The optimum number of nearest neighbours (k) was automatically selected through model tuning (tuneLength = 10) to achieve the best prediction accuracy while minimizing model variance and overfitting.

Artificial Neural Networks (ANNs)

Artificial neural networks are computational models inspired by the biological nervous system, designed to learn and represent complex non-linear relationships between inputs and outputs [13, 46]. They are composed of layers of interconnected neurons with adjustable weights, which are optimized through iterative learning algorithms such as backpropagation [102]. Due to their capacity to manage high-dimensional, noisy, and collinear datasets, ANNs are increasingly applied in agriculture and remote sensing for tasks such as crop classification, yield forecasting, and prediction of physiological traits from spectral reflectance [10, 108]. In this study, ANN modeling was conducted in R using the neuralnet package to establish predictive relationships between multispectral reflectance and crop variables, highlighting its potential as a reliable ML tool in agronomic prediction.

The ANN regression model was implemented using the caret package in R with K (5)-fold cross-validation to improve model robustness and generalization. Hyperparameter optimization included automatic tuning of the number of hidden neurons (size) and weight decay (decay) using repeated model training. Predictor variables were centred and scaled before model fitting, and the optimized network architecture was selected based on the minimum cross-validation RMSE to reduce overfitting and improve predictive performance.

Results

Correlation analysis between VIs, CCC, and ChlF

The correlation analysis between VIs, CCC, and ChlF during kharif season and rabi season is presented in Figs. 4 and 5.

Fig. 4.

Correlation matrix of vegetation indices against canopy chlorophyll content and chlorophyll fluorescence during kharif season

Fig. 5.

Correlation matrix of vegetation indices against canopy chlorophyll content and chlorophyll fluorescence during rabi season

The correlation analysis between thirty VIs, CCC, and ChlF was presented in Figs. 4 and 5. During the kharif season, CCC exhibited very strong correlations with NDRE (r = 0.95), LCI (r = 0.94), RECI (r = 0.94), and MTCI (r = 0.89). Moderate correlations were observed with GNDVI (r = 0.75), while IPVI, SRI, and TNDVI showed significant correlations (r = 0.59). For ChlF, strong correlations were recorded with LCI, NDRE, and RECI (r = 0.91 each), followed by MTCI (r = 0.84), GNDVI (r = 0.72), IPVI and TNDVI (r = 0.60), and SRI (r = 0.59). In the rabi season, CCC showed strong correlations with MTCI (r = 0.88), RECI (r = 0.88), NDRE (r = 0.87), LCI (r = 0.85), GNDVI (r = 0.75), GCI (r = 0.55), IPVI (r = 0.52), MSR (r = 0.52), NDVI (r = 0.52), SRI (r = 0.52), and TNDVI (r = 0.52), respectively. The same indices followed a similar trend for ChlF, recording correlation coefficient values were 0.92, 0.93, 0.94, 0.94, 0.84, 0.55, 0.72, 0.72, 0.72, 0.71, and 0.72, respectively.

Linear regression analysis between VIs and CCC, and VIs and ChlF

The regression analysis between individual VIs and CCC, and VIs and ChlF during kharif season and rabi season is presented in Tables 3 and 4.

Table 3.

Regression equations of different VIs against canopy chlorophyll content and chlorophyll fluorescence during kharif season

Sl. No	Vegetation index	Kharif season
		Canopy chlorophyll content			Chlorophyll fluorescence
		Regression equation	R2	RMSE	Regression equation	R2	RMSE
1	DVI	Y = 1.025x + 1.962	0.00005	0.177	Y = 0.577x + 0.751	0.0002	0.050
2	ENDVI	Y = 1.719x + 0.731	0.096	0.168	Y = 0.537x + 0.371	0.116	0.047
3	EXG	Y = 0.096x + 1.917	0.004	0.176	Y = 0.054X + 0.732	0.014	0.050
4	EXGR	Y = 0.0916X + 1.923	0.011	0.176	Y = 0.040X + 0.740	0.025	0.049
5	EXR	Y = 1.970–0.300X	0.024	0.175	Y = 0.759–0.111X	0.042	0.049
6	GCI	Y = 29.81X + 2.16	0.058	0.172	Y = 7.811X + 0.801	0.050	0.049
7	GDVI	Y = 7.880X + 1.916	0.002	0.177	Y = 2.424X + 0.741	0.002	0.050
8	GNDVI	Y = 1.923X + 0.823	0.564	0.117	Y = 0.522X + 0.446	0.519	0.035
9	GOSAV	Y = 1.841X + 1.892	0.005	0.176	Y = 0.536X + 0.735	0.005	0.050
10	GRVI	Y = 1.920–0.302X	0.024	0.175	Y = 0.740–0.112X	0.041	0.049
11	GSAVI	Y = 1.909 + 3.181X	0.003	0.177	Y = 0.956X + 0.740	0.003	0.050
12	IPVI	Y = 2.388–0.0245X	0.352	0.142	Y = 0.689X + 0.182	0.366	0.040
13	LCI	Y = 1.366 + 3.384X	0.882	0.061	Y = 0.928X + 0.592	0.826	0.021
14	MSR	Y = 1.580 + 0.298X	0.360	0.141	Y = 0.085X + 0.646	0.368	0.040
15	MTCI	Y = 1.395 + 1.548X	0.784	0.082	Y = 0.417X + 0.603	0.709	0.027
16	NDI	Y = 1.919 + 0.302X	0.024	0.175	Y = 0.112X + 0.740	0.041	0.049
17	NDRE	Y = 1.349 + 5.177X	0.893	0.058	Y = 1.412X + 0.589	0.829	0.021
18	NDVI	Y = 1.170 + 1.194X	0.352	0.142	Y = 0.345X + 0.527	0.366	0.040
19	NGBDI	Y = 2.510–1.135X	0.083	0.169	Y = 0.877–0.253X	0.051	0.049
20	NGRDI	Y = 1.919 + 0.302X	0.024	0.175	Y = 0.112X + 0.740	0.041	0.049
21	NPCI	Y = 2.320–0.938X	0.141	0.164	Y = 0.859X-0.270	0.145	0.046
22	OSAVI	Y = 1.948 + 0.249X	0.0001	0.177	Y = 0.218X + 0.747	0.001	0.050
23	RECI	Y = 1.413 + 2.023X	0.893	0.058	Y = 0.552X + 0.606	0.830	0.021
24	REOSAVI	Y = 1.948 + 0.215X	0.0001	0.177	Y = 0.188X + 0.747	0.001	0.050
25	RGI	Y = 2.111–0.196X	0.028	0.174	Y = 0.810–0.071X	0.045	0.049
26	SIPI	Y = 3.040–0.935X	0.237	0.154	Y = 1.075–0.277X	0.259	0.043
27	SRI	Y = 1.634 + 0.061X	0.344	0.143	Y = 0.018X + 0.662	0.349	0.040
28	TNDVI	Y = 2.512X-0.746	0.347	0.143	Y = 0.726X-0.027	0.361	0.040
29	VARIg	Y = 0.270X + 1.915	0.030	0.174	Y = 0.096X + 0.739	0.047	0.049
30	VARIr	Y = 0.677X + 1.654	0.204	0.158	Y = 0.199X + 0.665	0.219	0.044

Table 4.

Regression equations of different VIs against canopy chlorophyll content and chlorophyll fluorescence during rabi season

Sl. No	Vegetation index	Rabi season
		Canopy chlorophyll content			Chlorophyll fluorescence
		Regression equation	R2	RMSE	Regression equation	R2	RMSE
1	DVI	Y = 2.329–54.389X	0.263	0.149	Y = 0.855–15.833X	0.242	0.046
2	ENDVI	Y = 3.430–1.985X	0.078	0.167	Y = 0.915–0.223X	0.011	0.053
3	EXG	Y = 2.528–1.026X	0.276	0.148	Y = 0.867–0.214X	0.130	0.049
4	EXGR	Y = 2.311–0.536X	0.210	0.154	Y = 0.814–0.100X	0.080	0.051
5	EXR	Y = 0.983X + 2.061	0.121	0.163	Y = 0.138X + 0.763	0.026	0.052
6	GCI	Y = 50.187X + 2.344	0.305	0.145	Y = 15.304X + 0.864	0.308	0.044
7	GDVI	Y = 2.305–55.783X	0.204	0.155	Y = 0.847–16.088X	0.184	0.048
8	GNDVI	Y = 2.390X + 0.342	0.566	0.114	Y = 0.815X + 0.195	0.714	0.028
9	GOSAV	Y = 2.315–8.397X	0.195	0.156	Y = 0.849–2.403X	0.173	0.048
10	GRVI	Y = 0.922X + 2.211	0.109	0.164	Y = 0.120X + 0.782	0.020	0.052
11	GSAVI	Y = 2.308–19.139X	0.201	0.155	Y = 0.848–5.506X	0.180	0.048
12	IPVI	Y = 4.753X-2.304	0.273	0.148	Y = 1.991X-1.040	0.520	0.037
13	LCI	Y = 2.524X + 1.252	0.723	0.091	Y = 0.847X + 0.509	0.883	0.018
14	MSR	Y = 0.278X + 1.406	0.275	0.148	Y = 0.115X + 0.515	0.518	0.037
15	MTCI	Y = 1.023X + 1.394	0.783	0.081	Y = 0.322X + 0.569	0.843	0.021
16	NDI	Y = 2.211X-0.922	0.101	0.165	Y = 0.782–0.120X	0.020	0.052
17	NDRE	Y = 3.549X + 1.301	0.752	0.087	Y = 1.163X + 0.531	0.877	0.019
18	NDVI	Y = 2.377X + 0.073	0.273	0.148	Y = 0.995X-0.045	0.520	0.037
19	NGBDI	Y = 3.108–2.402X	0.498	0.123	Y = 1.057–0.647X	0.392	0.041
20	NGRDI	Y = 2.211–0.922X	0.109	0.164	Y = 0.782–0.120X	0.020	0.052
21	NPCI	Y = 2.273–1.226X	0.162	0.159	Y = 0.870–0.488X	0.279	0.045
22	OSAVI	Y = 2.344–9.554X	0.260	0.150	Y = 0.858–2.766X	0.237	0.046
23	RECI	Y = 1.168X + 1.419	0.758	0.085	Y = 0.378X + 0.572	0.870	0.019
24	REOSAVI	Y = 2.344–8.236X	0.260	0.150	Y = 0.858–2.384X	0.237	0.046
25	RGI	Y = 0.723X + 1.544	0.113	0.164	Y = 0.099X + 0.692	0.023	0.052
26	SIPI	Y = 6.595–4.397X	0.243	0.151	Y = 2.670–1.826X	0.454	0.039
27	SRI	Y = 0.044X + 1.560	0.269	0.149	Y = 0.018X + 0.579	0.507	0.037
28	TNDVI	Y = 5.39GX-4.177	0.272	0.148	Y = 2.261X-1.826	0.519	0.037
29	VARIg	Y = 2.151–0.535X	0.058	0.169	Y = 0.762–0.034X	0.003	0.053
30	VARIr	Y = 0.828X + 1.445	0.077	0.167	Y = 0.440X + 0.471	0.235	0.046

Regression analysis was performed between individual VIs and CCC, and as well as ChlF for both kharif season and rabi season. During the kharif season, the strongest coefficient of determination (R²) values for CCC were obtained with RECI (0.893), NDRE (0.893), MTCI (0.784) and GNDVI (0.564), accompanied by the lowest RMSE values of 0.058, 0.058, 0.082 and 0.117, respectively. For ChlF, higher R² values were recorded with RECI (0.830), NDRE (0.829), LCI (0.826), MTCI (0.709) and GNDVI (0.519), while the lowest RMSE values were 0.021 each for RECI, NDRE and LCI, 0.027 for MTCI and 0.035 for GNDVI. During rabi season, the VIs MTCI, RECI, NDRE, LCI and GNDVI exhibited higher R² values of 0.783, 0.758, 0.752, 0.723 and 0.566 with corresponding RMSE values of 0.081, 0.085, 0.087, 0.091 and 0.114, respectively, for CCC. Similarly, for ChlF, higher R² values were observed with LCI (0.883), NDRE (0.877), RECI (0.870), MTCI (0.843), GNDVI (0.714), IPVI (0.520), NDVI (0.520), TNDVI (0.519) and SRI (0.507), with respective RMSE values of 0.018, 0.019, 0.019, 0.021, 0.028, 0.037, 0.037, 0.037 and 0.037.

Model fitting for CCC and ChlF Using ML Algorithms

The fitting performance of different ML algorithms for CCC and ChlF estimation during kharif season and rabi season is presented in Fig. 6 and 7.

Fig. 6.

Different machine learning algorithms for model performance of canopy chlorophyll content during kharif season and rabi season (a. RF, b. PLSR, c. SVM, d. XGB, e. KNN and f. ANN)

Fig. 7.

Different machine learning algorithms for model performance of canopy chlorophyll fluorescence during kharif and rabi season (a. RF, b. PLSR, c. SVM, d. XGB, e. KNN and f. ANN)

The RF model demonstrated superior fitting performance for CCC, recording an R² value of 0.982 with RMSE and MAE values of 0.023 and 0.016, respectively, during the kharif season, and an R² value of 0.963 with RMSE and MAE values of 0.036 and 0.029, respectively, during rabi season, indicating strong model calibration and reliability. In contrast, the KNN model exhibited the weakest fitting performance, with an R² value of 0.799, RMSE of 0.085, and MAE of 0.066 during kharif, and an R² value of 0.761 with RMSE and MAE values of 0.096 and 0.075, respectively, during rabi, highlighting its limited capability in modelling the CCC dataset. Among the remaining models, SVM, XGB, and ANN showed moderate to good fitting performance. SVM achieved R² values of 0.965 and 0.875 during kharif season and rabi season, respectively, with corresponding RMSE values of 0.032 and 0.067, and MAE values of 0.021 and 0.048, respectively. Similarly, XGB recorded R² values of 0.927 and 0.853, RMSE values of 0.067 and 0.079, and MAE values of 0.048 and 0.058 during kharif season and rabi season, respectively. ANN also demonstrated satisfactory fitting accuracy, with R² values of 0.899 and 0.839, RMSE values of 0.055 and 0.073, and MAE values of 0.040 and 0.057 during kharif season and rabi season, respectively. PLSR exhibited comparatively lower, yet acceptable, fitting performance, recording R² values of 0.896 and 0.826, RMSE values of 0.055 and 0.076, and MAE values of 0.040 and 0.059 during kharif season and rabi season, respectively.

For ChlF, RF again demonstrated the highest fitting efficiency, recording an R² value of 0.960 with RMSE and MAE values of 0.010 and 0.007, respectively, during kharif season, and an R² value of 0.980 with RMSE and MAE values of 0.008 and 0.006, respectively, during rabi season, confirming its high reliability and robustness. Among the other models, XGB showed good fitting performance during kharif season, with an R² value of 0.908, RMSE of 0.016, and MAE of 0.010. In contrast, SVM performed comparatively better during rabi season, recording an R² value of 0.941 with RMSE and MAE values of 0.014 and 0.009, respectively. ANN also exhibited consistent fitting accuracy across both seasons, with R² values of 0.874 and 0.903, RMSE values of 0.017 and 0.017, and MAE values of 0.013 and 0.014 during kharif season and rabi season, respectively. PLSR demonstrated moderate fitting performance in both seasons, recording an R² value of 0.812 with RMSE and MAE values of 0.021 and 0.016 during kharif season, and an R² value of 0.889 with RMSE and MAE values of 0.019 and 0.015, respectively, during rabi season. Conversely, KNN exhibited the weakest fitting performance during kharif season, with an R² value of 0.691, RMSE of 0.029, and MAE of 0.013. During rabi season, XGB and KNN recorded comparatively lower fitting accuracies, with R² values of 0.798 and 0.843, RMSE values of 0.051 and 0.024, and MAE values of 0.042 and 0.020, respectively.

Prediction accuracy of CCC and ChlF values of maize using ML algorithms

The prediction accuracy of different ML algorithms for CCC and ChlF estimation during kharif season and rabi season is presented in Fig. 8 and 9.

Fig. 8.

Different machine learning algorithms for prediction of canopy chlorophyll content during kharif season and rabi season (a. RF, b. PLSR, c. SVM, d. XGB, e. KNN and f. ANN)

Fig. 9.

Different machine learning algorithms for prediction of chlorophyll fluorescence during kharif season and rabi season (a. RF, b. PLSR, c. SVM, d. XGB, e. KNN and f. ANN)

The predictive performance of six ML algorithms, namely RF, PLSR, SVM, XGB, KNN, and ANN, were evaluated for estimating CCC and ChlF. Among the evaluated models, PLSR demonstrated the highest predictive accuracy for CCC during the kharif season, with an R² value of 0.891, along with the lowest RMSE and MAE values of 0.063 and 0.045, respectively. This was followed by ANN and XGB, which recorded R² values of 0.888 and 0.876, respectively, with corresponding RMSE and MAE values of 0.064 and 0.046, and 0.081 and 0.052, respectively. During rabi season, SVM and XGB exhibited comparatively superior predictive performance, achieving R² values of 0.623 and 0.609, respectively, with lower RMSE and MAE values of 0.083 and 0.066, and 0.084 and 0.058, respectively. In contrast, KNN exhibited relatively poor predictive performance during both kharif season and rabi season, with R² values of 0.745 and 0.435, RMSE values of 0.100 and 0.100, and MAE values of 0.066 and 0.078, respectively.

For ChlF prediction, PLSR achieved the highest predictive accuracy during the kharif season, recording an R² value of 0.889 along with the lowest RMSE and MAE values of 0.017 and 0.013, respectively. ANN, RF, and XGB also demonstrated strong predictive capability, with R² values of 0.852, 0.838, and 0.838, respectively, accompanied by RMSE values of 0.020, 0.021, and 0.022, and MAE values of 0.016, 0.017, and 0.017, respectively. Conversely, KNN and SVM showed comparatively lower prediction accuracy, with R² values of 0.715 and 0.707, respectively, along with higher RMSE values of 0.030 and 0.029, and MAE values of 0.020 for both models. During the rabi season, RF outperformed the other algorithms in predicting ChlF, recording the highest R² value of 0.867 and the lowest RMSE and MAE values of 0.015 and 0.011, respectively. The lowest predictive performance was observed with KNN, which recorded an R² value of 0.620, RMSE of 0.025, and MAE of 0.021. This was followed by XGB, SVM, ANN, and PLSR, which achieved R² values of 0.632, 0.701, 0.819, and 0.825, respectively, with corresponding RMSE values of 0.032, 0.023, 0.017, and 0.016, and MAE values of 0.024, 0.016, 0.011, and 0.011, respectively.

Principal component analysis of VIs under seasonal variability

Principal Component Analysis (PCA) was performed to evaluate the relationships among VIs and to assess seasonal variability between kharif (K) and rabi (R) seasons is presesned in Fig. 10. The scree plot (Fig. 10d) indicated that the first two principal components (PC1 and PC2) explained 83.3% of the total variance, where PC1 contributed 59.8% and PC2 accounted for 23.5% of the variance. This indicates that most of the variability present in the multispectral vegetation indices was effectively represented by the first two components.

Fig. 10.

Principal Component Analysis (PCA) of vegetation indices under kharif (K) season and rabi(R) season showing (a) PCA individual plot, b PCA variable plot, c PCA biplot, and d PCA scree plot illustrating variance explained by principal components

The PCA variable plot (Fig. 10b) revealed strong positive associations among several red-edge and chlorophyll-sensitive indices, including RECI, NDRE, GNDVI, LCI, and MTCI. These indices showed high positive loadings mainly along PC2, indicating their significant contribution toward explaining chlorophyll-related variability. Similarly, indices such as GRVI, RGI, EXR, GCI, SIPI, and NPCI showed strong contributions along the positive side of PC1. In contrast, indices including OSAVI, DVI, EXG, REOSAVI, VARIg, and NGRDI were positioned on the negative side of PC1. The PCA biplot (Fig. 10c) and individual plot (Fig. 10a) showed moderate separation between kharif and rabi observations. The kharif season exhibited relatively compact clustering, whereas rabi season observations were more widely dispersed, indicating greater variability during the rabi season. Partial overlap between the seasonal clusters suggested that some spectral characteristics remained common across seasons.

Discussion

Linear regression analysis between VIs and CCC, and VIs and ChlF

Red-edge VIs (RECI, NDRE, MTCI, LCI) outperformed greenness VIs because the red-edge zone (∼700–740 nm) is most sensitive to chlorophyll variation while avoiding red-band saturation at high LAI [91, 92]. This spectral region tracks shifts in slope and position linked to chlorophyll concentration, improving linearity and dynamic range for CCC and ChlF. Studies in maize and related crops confirm that red-edge indices are stronger predictors of CCC and physiological status than traditional VIs, remaining robust across phenophases and stress conditions [107, 109].

The red-edge spectral region, spanning approximately 680–760 nm between red and NIR wavelengths [140], is highly sensitive to chlorophyll-driven variations in leaf and canopy structure. This region exhibits a sharp increase in reflectance due to strong red-light absorption by chlorophyll and enhanced NIR reflectance from leaf mesophyll, making it ideal for monitoring CCC, canopy structure, and photosynthetic function [17, 91, 92]. The red-edge position (REP), defined as the wavelength of maximum slope between red and NIR reflectance, shifts toward longer wavelengths with increased chlorophyll or water content and toward shorter wavelengths under stress or senescence, reflecting physiological and structural changes in leaves [64, 74, 114].

Red-edge VIs such as RECI, NDRE, MTCI, and LCI exploit this sensitivity, consistently yielding higher predictive accuracy (R²) and lower RMSE than traditional greenness indices when estimating CCC and ChlF in maize [91]. Their superiority arises from several factors: red-edge light penetrates deeper into dense canopies compared to red or blue light, maintaining sensitivity during mid to late phenophases when chlorophyll accumulation is high, and minimizing saturation effects commonly observed in conventional VIs at moderate-to-high LAI [28, 83, 91]. Additionally, the combination of red-edge and NIR reflectance enhances the linearity and dynamic range of VIs, enabling robust estimation across diverse canopy structures and environmental conditions [5, 92, 113, 150].

Model training performance and fitting efficiency of AI algorithms

Among the tested algorithms, RF delivered the most consistent training accuracy. RF showed strong performance due to its ensemble nature, ability to resist overfitting, and capacity to identify the most relevant input variables for model simplification and accuracy improvement [1, 114]. Its robustness in handling complex datasets has also been confirmed in salinity mapping and crop chlorophyll estimation [47]. Similarly, XGB achieved higher fitting accuracy by incorporating L1 and L2 regularization into its objective function, effectively controlling overfitting and improving model generalization in noisy agricultural datasets [98, 99].

Support vector machine algorithm also performed well, supported by their ability to minimize errors on unseen data without assumptions on data distribution, while deriving optimal hyperplanes for nonlinear problems [9, 15, 82]. This efficiency in handling high-dimensional datasets and small training samples explains its competitive accuracy in crop trait modelling [71, 98]. ANN provided moderate performance by effectively modelling nonlinear relationships, though their accuracy was sometimes constrained by dataset size and training complexity [87]. By contrast, KNN produced the weakest fits, with high variability and limited robustness, echoing earlier reports of its sensitivity to data distribution and outliers [104].

Predictive capability and comparative accuracy of ML algorithms

Among the evaluated ML algorithms, PLSR recorded the highest R² values along with the lowest RMSE and MAE values, [27, 127, 138], and recent UAV-based crop monitoring studies [4, 58]. The comparatively superior performance of PLSR may be attributed to the relatively smaller dataset size and the high multicollinearity among VIs derived from multispectral bands. Since several VIs share overlapping spectral information, PLSR effectively reduces dimensionality and extracts latent variables strongly associated with chlorophyll dynamics, thereby improving prediction stability under correlated spectral conditions [4, 81, 115, 151]. Although the comparative evaluation was based on R², RMSE, and MAE, no formal statistical significance tests were performed among the evaluated machine learning models. Therefore, the observed differences should be interpreted as comparative performance trends rather than statistically significant superiority [79, 97]. PCA was performed primarily to evaluate multicollinearity and the relationships among vegetation indices rather than for dimensionality reduction [145]. Although several VIs exhibited strong correlations due to overlapping spectral information, all indices were retained to comprehensively compare the predictive performance of different ML algorithms [149]. Moreover, ensemble-based algorithms such as RF are relatively robust to correlated predictors, while PLSR effectively addresses multicollinearity by projecting correlated variables into orthogonal latent components. Similar observations have also been reported in recent UAV-based crop monitoring studies, where linear latent-variable approaches demonstrated robust performance in handling redundant multispectral features and limited training datasets [92]. In contrast, advanced nonlinear models generally require larger and more diverse datasets for stable generalization and may become more sensitive to overfitting under limited sample conditions. Similar observations were reported by Zhou et al. [148], who highlighted the effectiveness of PLSR in integrating reflectance and fluorescence information for crop trait estimation. The relatively limited dataset compared with the number of VIs and ML algorithms evaluated may also have influenced model stability and prediction uncertainty [121]. Although five-fold cross-validation and hyperparameter optimization reduced the risk of overfitting, larger multi-season datasets would further improve model robustness and generalizability [60].

Random Forest demonstrated comparatively superior performance over XGB, ANN, and SVM models for CCC and ChlF, likely due to its strong capability to handle nonlinear relationships, high-dimensional datasets, and multicollinearity among VIs [132, 139]. Since many UAV-derived vegetation indices share overlapping spectral information, RF effectively reduces overfitting by constructing multiple decision trees using random subsets of variables and samples, thereby improving model stability and generalization ability under varying field conditions [99, 105]. In contrast, SVM performance was highly sensitive to kernel selection and parameter tuning [57, 141, 142]. The ensemble-based architecture of RF also provides greater robustness against noise, seasonal variability, and heterogeneous canopy conditions [96, 116, 144], which may explain its improved predictive accuracy during the rabi season. Similar findings have been reported in recent UAV-based crop monitoring studies, where RF consistently achieved reliable performance under complex agricultural datasets due to its resistance to noise, ability to capture nonlinear spectral–physiological interactions, and reduced sensitivity to correlated input variables [48, 134].

In contrast, KNN exhibited comparatively lower predictive performance, likely due to its sensitivity to high-dimensional data, spectral redundancy, and noise. In high-dimensional feature spaces, distance metrics become less discriminative, thereby reducing model efficiency [120]. Furthermore, smaller k values may increase overfitting risk, whereas larger k values can lead to underfitting [45]. The computational complexity associated with neighbour-based searching may also reduce model efficiency compared to tree-based and kernel-based approaches [54, 70]. Higher predictive accuracy for canopy chlorophyll estimation was achieved by XGB and ANN due to their ability to capture complex, non-linear relationships between multispectral VIs and CCC. XGB uses gradient boosting to iteratively refine predictions, enhancing feature interactions and improving generalization, which is particularly effective for high-dimensional spectral data [49]. Similarly, ANN can model intricate non-linear spectral–chlorophyll dynamics through its layered structure and adaptive learning, offering advantages over linear models [6].

The comparatively lower predictive accuracy observed between two seasons may be attributed to seasonal variability in canopy architecture, illumination conditions, and crop physiological responses. Variations in plant structure, leaf orientation, canopy density, and soil background effects can alter spectral reflectance patterns and weaken the relationship between VIs and chlorophyll-related traits, thereby affecting ML model performance [3, 93, 110, 111]. In addition, environmental variability such as temperature, radiation, and moisture conditions may increase spectral heterogeneity and reduce model generalizability across seasons [78, 126].

Seasonal variability and multicollinearity of VIs based on PCA

The high cumulative variance explained by PC1 and PC2 demonstrates that the selected VIs effectively captured the major spectral variability associated with maize canopy characteristics [29]. The strong clustering of RECI, NDRE, GNDVI, LCI, and MTCI confirms the close relationship among red-edge-based VIs due to their sensitivity to chlorophyll concentration and reduced saturation effects under higher canopy density [65]. Their close positioning also indicates substantial multicollinearity resulting from overlapping spectral information derived from similar wavelength regions [53, 137]. The strong contribution of pigment-sensitive indices such as SIPI, NPCI, GCI, and GRVI suggests their effectiveness in representing canopy pigment composition and physiological status [128]. Conversely, indices positioned on the negative side of PC1 were more associated with structural and soil-background-related spectral responses. The PCA biplot further demonstrated seasonal variability in spectral behaviour between kharif season and rabi season. The comparatively wider dispersion observed during the rabi season indicates increased canopy heterogeneity and environmental variability, which may have contributed to the relatively lower prediction accuracy of machine learning models during the rabi season [31, 146]. Variations in canopy structure, illumination conditions, and crop growth dynamics across seasons can substantially influence spectral reflectance and vegetation index responses [34, 68].

Furthermore, the observed multicollinearity among VIs supports the superior performance of PLSR in the present study, as PLSR efficiently handles highly correlated predictors by extracting latent variables and reducing dimensional redundancy [112, 129]. Overall, the PCA results highlighted the dominance of red-edge vegetation indices in explaining chlorophyll variability and confirmed the significant influence of seasonal variability on spectral responses and model performance.

Future use

Future studies should focus on larger multi-season datasets, advanced deep learning models, and integration of environmental variables to improve model robustness and generalizability across diverse agroecosystems. UAV-based multispectral imaging combined with AI techniques also holds strong potential for real-time crop monitoring, early stress detection, yield forecasting, and precision nutrient and water management. Furthermore, integration with IoT-based sensors and satellite observations could enhance large-scale operational applicability and support digital agriculture and sustainable crop management systems.

Novelty

This experiment is novel in applying a wide set of 30 VIs alongside various ML algorithms across two contrasting seasons, kharif and rabi, to evaluate canopy chlorophyll dynamics in maize. The dual focus on CCC and ChlF, coupled with UAV-based red-edge indices, provides a comprehensive framework for linking spectral information with photosynthetic efficiency. Moreover, the comparative evaluation of classical (PLSR, MLR) and advanced ML algorithms (RF, XGB, ANN) under seasonal variability contributes to improving the understanding of model robustness and applicability for precision crop monitoring under diverse field conditions.

Limitations

Despite the promising results, the study has certain limitations. The analysis was limited to maize under specific agro-climatic conditions, which may restrict broader applicability without validation across diverse crops and environments. UAV data acquisition is also influenced by weather conditions and operational constraints. Additionally, some ML models showed near-perfect training performance, indicating possible overfitting due to limited sample size and complex spectral relationships. Therefore, further multi-season and multi-location validation is required to improve model robustness and generalizability. Although five-fold cross-validation improved model robustness, independent validation using datasets from different years, seasons, or locations was not performed. Additionally, UAV-derived multispectral data may exhibit spatial autocorrelation, which could reduce the independence between training and testing datasets and potentially lead to optimistic estimates of prediction accuracy. Therefore, further validation using spatially independent, multi-season, and multi-location datasets is required to improve model generalizability and transferability.

Conclusion

This study demonstrated the effectiveness of integrating UAV-based multispectral sensing with ML techniques for the estimation of CCC and ChlF in maize under contrasting kharif season and rabi season. The results confirmed that red-edge VIs, particularly RECI, NDRE, MTCI, and LCI, provided superior sensitivity to chlorophyll dynamics compared to conventional greenness indices due to their reduced saturation effects and stronger association with canopy physiological characteristics. Among the evaluated models, PLSR exhibited superior prediction performance during the kharif season, while RF performed comparatively better during rabi season. Advanced algorithms such as XGB and ANN were effectively captured nonlinear spectral–chlorophyll relationships, whereas KNN showed comparatively lower performance under high-dimensional spectral conditions. The study further revealed that seasonal variability in canopy architecture, environmental conditions, and spectral heterogeneity influenced model generalizability, resulting in comparatively lower predictive accuracy during the rabi season. Nevertheless, the integration of UAV-based multispectral sensing and machine learning provides a promising framework for estimating CCC and ChlF under the experimental conditions evaluated in this study, with potential applications in real-time crop physiological monitoring, precision nutrient management, yield forecasting, and sustainable agricultural decision-making. However, as the investigation was conducted at a single location using a relatively limited dataset, additional validation across multiple seasons, locations, and environmental conditions is required before broader operational implementation.

Authors’ contributions

**Pradosh Kumar Parida**: Writing – review & editing, Writing – original draft, Visualization, Validation, Supervision, Software, Resources, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. **Somasundaram Eagan**: Writing – review & editing, Visualization, Validation, Supervision, Resources, Project administration, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. **Krishnan Ramanujam, Radhamani Sengodan**: Writing – review & editing, Visualization, Validation, Supervision, Methodology, Conceptualization. **Sivakumar Uthandi**: Writing – review & editing, Visualization, Validation, Supervision. **Naba Kishor Parida, Santosh Ganapati Patil, Nitish Kumar Jena, Manjushree Singh, Shri Rangasami Silambiah Ramasamy:** Writing – review & editing, Visualization, Validation.

Funding

This research received no external funding.

Data availability

The data generated and analysed during this study are included in the article. Additional data will be made available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

All authors have read and approved the final manuscript for publication.

Competing interests

The authors declare no competing interests.