How Do Spectral Scales and Machine Learning Affect SPAD Monitoring at Different Growth Stages of Winter Wheat?

Abstract

Chlorophyll serves as a crucial indicator for crop growth monitoring and reflects the health status of crops. Hyperspectral remote sensing technology, leveraging its advantages of repeated observations and high-throughput analysis, provides an effective approach for non-destructive chlorophyll monitoring. However, determining the optimal spectral scale remains the primary bottleneck constraining the widespread application of hyperspectral remote sensing in crop chlorophyll estimation: excessively fine spectral scale readily introduces redundant information, leading to dramatically increased data dimensions and reduced computational efficiency; conversely, overly coarse spectral scale risks losing critical spectral features such as absorption peaks and reflection troughs, thereby compromising model accuracy. Therefore, establishing an appropriate spectral scale that effectively preserves spectral feature information while maintaining computational efficiency is crucial for enhancing the accuracy and practicality of chlorophyll remote sensing estimation. To address this, this study proposes a three-dimensional analytical framework integrating “spectral scale—machine learning algorithm—crop growth stage” to systematically solve the scale optimization problem. Ground-truth measurements and hyperspectral data from five growth stages of winter wheat in Fengyang County, Anhui Province, were collected. Spectral bands sensitive to chlorophyll were analyzed, and four modeling methods—Ridge Regression (RR), K-Nearest Neighbors (KNN), Random Forest (RF), and Support Vector Regression (SVR)—were employed to integrate data from different spectral scales with respective bandwidths of 2, 3, 5, 7, 10, 20, and 50 nanometers (nm). The results evaluated the response characteristics of raw band reflectance to chlorophyll values and its impact on machine learning-based chlorophyll estimation across different spectral scales. Results indicate: (1) Canopy spectra significantly correlated with winter wheat chlorophyll primarily reside in the red and red-edge bands; (2) For single-scale analysis, larger spectral scales (10, 20 nm) enhance monitoring accuracy compared to 1 nm high-resolution data, while medium and small scales (5, 7 nm) may degrade accuracy due to redundant noise introduction. (3) Integrating growth stages, spectral scales, and machine learning revealed optimal monitoring accuracy during the jointing and heading stages using 1–5 nm spectral scales combined with the KNN algorithm. For the booting, flowering, and grain filling stages, the highest accuracy was achieved using 20–50 nm spectral scales combined with either the KNN or RF algorithm. The results indicate that high-precision chlorophyll inversion for winter wheat does not rely on a single fixed model or scale, but rather on the dynamic adaptation of the “scale-model-growth stage” triad. The proposed systematic framework not only provides a theoretical basis for chlorophyll monitoring using multi-platform remote sensing data, but also offers methodological support for future crop-sensing sensor design and data processing strategy optimization.

1. Introduction

Wheat is China’s second-largest staple crop and ranks among the world’s most widely distributed, extensively cultivated, and highest-yielding food crops. China’s wheat production constitutes a vital component of national food security [1], and ensuring stable and high yields has remained a key research focus. Chlorophyll is a crucial pigment enabling crops to absorb solar radiation for photosynthesis [2,3,4]. Chlorophyll content is intrinsically linked to leaf photosynthetic capacity; its concentration, distribution, and variation directly or indirectly reflect plant photosynthetic ability, nutrient status, and vegetation productivity. It serves as a key indicator determining crop growth status, yield levels, and quality [5,6,7]. Consequently, timely and accurate acquisition of chlorophyll content forms the foundation for monitoring crop growth and predicting yield [8,9]. Traditional chemical methods for monitoring chlorophyll content rely on destructive measurements, which are time-consuming, labour-intensive, and limited in scope, making large-scale monitoring and rapid application challenging [10,11,12]. There is therefore an urgent need for a low-cost, efficient, and accurate monitoring method.

In recent years, remote sensing technology has demonstrated extensive applications and promising prospects in crop monitoring and forecasting due to its advantages of large-area simultaneous observation and instantaneous imaging [13,14]. Satellite-derived spectral imagery enables wide-area monitoring but is susceptible to cloud cover, features long revisit periods, and exhibits low spectral and spatial resolution [15]; Unmanned aerial vehicle (UAV) remote sensing, as a significant method for monitoring crop growth, typically employs multispectral sensors at relatively low cost. However, the spectral bands captured are limited in their relevance to crop development. Both approaches suffer from constraints in spectral scale, compromising the accuracy of retrieved variables and hindering the timely and effective detection of early signals of crop stress [16]. Hyperspectral remote sensing technology offers high spectral scale and detailed data. Compared to traditional panchromatic and multispectral remote sensing techniques, it features multiple imaging bands capable of capturing diverse spectral characteristics associated with crop biochemistry and physiological traits [17]. This enables the identification of sensitive bands closely related to crops, facilitating precise and efficient monitoring of spectral changes in crops [18].

Currently, extensive research employs specific spectral band information to construct vegetation indices (VIs) for monitoring crop physiological and biochemical trait parameters, enabling qualitative and quantitative assessments of vegetation vitality and growth dynamics [19]. However, during the integration of multiple spectral bands into VIs, some detailed information from the original bands may be lost [20]. Furthermore, certain VIs exhibit heightened sensitivity only to specific crop types, growth stages, environmental conditions, and monitoring scenarios. They may fail to accurately reflect crop physiological states for other vegetation types or differing ecological environments. Studies such as that by Thenkabail et al. [21] indicate that VIs readily exhibit saturation during later crop growth stages, rendering them ineffective for estimating crop traits. Given these considerations, raw data modelling approaches significantly reduce computational costs and enhance efficiency by selecting characteristic bands strongly correlated with target variables as independent variables. Compared to VIs, raw spectral data minimises error propagation and avoids saturation effects [22].

Direct utilisation of reflectance information from raw spectral bands avoids loss of original spectral characteristics. However, hyperspectral data, comprising hundreds or thousands of narrow spectral bands, faces challenges of data redundancy and the curse of dimensionality [20]. Machine learning algorithms possess the capability to handle complex relationships between predictors and target variables, demonstrating excellent redundancy mitigation effects when addressing multimodal, high-dimensional data. Compared to traditional regression algorithms, they offer significant improvements and have been widely applied in crop phenotyping monitoring [23,24]. Shen et al. [25] monitored winter wheat SPAD values using ground-based hyperspectral data and measured agronomic data. Their findings indicated that partial least squares regression combined with specific preprocessing methods enhanced the accuracy and stability of SPAD predictions. Bhadra et al. [26] combined multiple machine learning algorithms with derivative processing to estimate sorghum chlorophyll content, concluding that support vector machine regression models delivered optimal performance. Xu et al. [27] employed four machine learning algorithms alongside feature selection methods to monitor liquorice chlorophyll content using UAV hyperspectral data. Their findings indicated that the combination of genetic algorithms and random forest models yielded the most favourable results. Zhu et al. [28] employed three machine learning algorithms alongside multivariate linear regression for monitoring maize and wheat LCC, observing superior estimation performance from SVM and RF. Whilst machine learning methods demonstrate robust nonlinear fitting and feature learning capabilities in chlorophyll inversion, spectral scale directly influences both the strength of correlation between feature bands and chlorophyll content, and the model’s feature extraction capacity.

As a primary parameter influencing spectral information, variations in spectral scale hinder direct comparison and fusion of data from different sources [29], thereby limiting assessments of model cross-platform applicability. Jiang’s [30] study employed PROSPECT simulations with stepwise bandwidths from 5 nm to 35 nm, finding that indices such as VIUPD, NDVI, and SRI achieved chlorophyll monitoring accuracies with R² > 0.88 at a 10 nm spectral resolution. However, Murugan’s [31] research suggested that data with spectral bandwidths below 8 nm yield superior accuracy in estimating chlorophyll a in aquatic environments. Dian [32] proposed that employing spectral scales between 10 and 30 nm yields superior results for chlorophyll inversion at leaf or canopy levels, a scale already comparable to satellite or UAV multispectral data. In summary, although studies have explored optimal scales, they face challenges in determining suitable scales. For instance, Jiang and Dian both set bandwidths in 5 nm increments, while Murugan used multiples of 2 nm. These scale settings are disconnected from actual sensors, featuring a wide range of scale choices that are poorly configured and fail to account for the characteristics of different sensors. Furthermore, they neglect the dynamic changes throughout the entire crop growth period. At different growth stages, vegetation exhibits changing relationships between spectral characteristics, chlorophyll content, and other biochemical/structural parameters. Consequently, the optimal spectral scale may not remain constant throughout the growth cycle. Thus, determining the appropriate scale warrants further investigation. Currently, research systematically examining how spectral scales themselves influence crop chlorophyll inversion accuracy remains scarce. Although studies have assessed differences in how various spectral scales affect crop biochemical parameters, these analyses primarily rely on vegetation indices. Few investigations utilise single-band reflectance data to evaluate SPAD response characteristics across different spectral scales, particularly lacking systematic coupling with machine learning methods for quantitative evaluation across distinct growth stages.

This study utilised winter wheat as its subject, employing 1 nm resolution canopy reflectance data acquired via hyperspectral-ground object spectrometers. Raw spectral data were resampled to seven scales—2, 3, 5, 7, 10, 20, and 50 nm—using Gaussian spectral response functions. Four machine learning methods—ridge regression (RR), K-Nearest Neighbours (KNN), Random Forest (RF), and Support Vector Regression (SVR)—were used to establish chlorophyll content monitoring models for winter wheat across its growth stages. The research framework is illustrated in Figure 1. Compared with existing studies, the innovation of this research lies in the first systematic construction of a three-dimensional analytical framework integrating “spectral scale-growth stage-machine learning algorithm.” This framework aims to: (1) systematically reveal the influence patterns of spectral scale variations on SPAD values; (2) deeply explore the accuracy differences of SPAD inversion models under the coupling of different spectral scales and machine learning algorithms; (3) Dynamically explore the growth stage specificity of optimal “scale-model” combinations. By elucidating the dynamic evolution of optimal spectral scales across growth stages and their coupling mechanisms with different algorithms, this study aims to provide a theoretical basis for high-precision monitoring of wheat chlorophyll using multi-platform remote sensing data, and to offer references for future sensor design and data processing strategy optimization.

2. Materials and Methods

2.1. Experimental Design

The study area is located in Shanhou Village, Fengyang County, Anhui Province (32°53′32″ N, 117°33′55″ E), situated within the Jianghuai Plain and belonging to the hilly terrain of the Jianghuai watershed (Figure 2). The region has an average annual temperature of 14.9 °C and average annual precipitation of 904.4 mm. The study area comprised 45 plots, with five nitrogen gradient treatments (N0: 0 kg/ha, N1: 90 kg/ha, N2: 180 kg/ha, N3: 270 kg/ha, N4: 360 kg/ha), each replicated three times. To mitigate the potential influence of single-variety specificity on the generalisability of assessment results, three representative winter wheat varieties were selected: V1: high-gluten Huai Mai 44, V2: high-gluten Yan Nong 999, and V3: low-gluten Ning Mai 13. Each plot measured 6 m in length and 4.4 m in width, with a plant spacing of 30 cm. Seeds were sown using a drill method. Sample data were collected at different growth stages: Jointing stage (T1: 15 March 2024), heading stage (T2: 9 April 2024), full heading stage (T3: 17 April 2024), flowering stage (T4: 25 April 2024), and grain filling stage (T5: 7 May 2024).

2.2. Data Acquisition

2.2.1. SPAD Data Collection

Using the TYS-4N Plant Nutrient Analyser (Zhejiang Topyun Agricultural Technology Co., Ltd., Hangzhou, China), obtain Soil and Plant Analysis Development (SPAD) data for canopy leaves across five growth stages. Within each plot, ten plants representative of the overall growth condition of the winter wheat were selected. Measurements were taken at the midpoint of the second leaf (the second leaf from the base) on each plant, avoiding the main leaf vein. The average of the ten samples was recorded as the SPAD value for that plot. The second-from-top leaf occupies the mid-canopy region, exhibiting stable physiological conditions. This position avoids the excessive sensitivity of the flag leaf and the premature senescence of basal leaves, thereby reliably reflecting the wheat’s nutritional status.

2.2.2. Canopy Spectral Data Acquisition

Spectral data acquisition was performed using the ASD FieldSpec HandHeld2 handheld ground-based spectrometer (Analytical Spectral Devices, Boulder, CO, USA). Spectral measurements spanned 325 to 1075 nm with 1 nm resolution. Experiments were conducted between 10:00 AM and 1:00 PM local time under clear, windless conditions. Calibration was performed using a standard white plate prior to data collection, followed by subsequent calibrations every three plots to mitigate environmental noise. During measurements, three representative sampling points exhibiting uniform growth conditions were selected per plot. Data were collected once at each point to generate three spectral curves, yielding a total of nine spectral curves per plot. The average of these curves served as the raw spectral data for each plot.

2.3. Data Processing

2.3.1. Resampling of Canopy Hyperspectral Data

Spectral resolution refers to a sensor’s ability to distinguish between adjacent wavelength reflectance signals, typically expressed as bandwidth (Full Width at Half Maximum, FWHM). Significant variations exist in spectral scale characteristics across different remote sensing platforms and optical sensors, directly impacting the accuracy and reliability of crop biochemical parameter monitoring. Ground-based hyperspectral instruments typically feature concentrated spectral scale, with FWHM ranging between 3 nm and 10 nm [33,34,35]. As a vital method for monitoring crop growth, UAV remote sensing, constrained by payload limitations, predominantly employs broadband multispectral sensors. Examples such as the Parrot Sequoia+ and MicaSense Altum typically feature FWHM > 10 nm [36,37]. In contrast, UAV hyperspectral sensors generally exhibit FWHM values ranging from 1 nm to 10 nm [38,39], with certain models displaying lower FWHM values. For instance, the Cubert UHD 185 achieves 25 nm at 850 nm [40]. Satellite remote sensing offers the advantage of large-area observation with a broad spectral scale range. Most sensors exhibit FWHM values between 2.5 nm and 50 nm [41,42,43], with some high-resolution sensors achieving FWHM as low as 0.1 nm and low-resolution sensors reaching up to 1000 nm. Considering the characteristics of each sensor, this study employed Gaussian distribution spectral response functions in Python (3.12.4) to resample raw 1 nm spectral reflectance data at seven spectral scale intervals: 2, 3, 5, 7, 10, 20, and 50 nm. This approach aims to simulate a continuous resolution range from fine (approaching ground/UAV hyperspectral) to coarse (covering UAV multispectral and typical satellite sensors) to quantitatively reveal the influence patterns of spectral scale variations on the estimation accuracy of winter wheat SPAD values. This provides direct evidence for scale matching and model transfer among multi-source remote sensing data.

Table 1. Resampled spectral scale and corresponding number of bands.

Spectral Scale (nm)	Number of Bands	Spectral Scale (nm)	Number of Bands
1	501	7	71
2	251	10	51
3	167	20	26
5	101	50	11

presents the resampled spectral scales and corresponding spectral band counts.

$${\mathrm{\rho }}_{\mathrm{j}}=\underset{{\mathrm{\lambda }}_{\mathrm{s}}\left(\mathrm{j}\right)}{\overset{{\mathrm{\lambda }}_{\mathrm{e}}\left(\mathrm{j}\right)}{\int }}\mathrm{\rho }\left(\mathrm{\lambda }\right){\mathrm{S}}_{\mathrm{j}}\left(\mathrm{\lambda }\right){\mathrm{d}}_{\mathrm{\lambda }}|\underset{{\mathrm{\lambda }}_{\mathrm{s}}\left(\mathrm{j}\right)}{\overset{{\mathrm{\lambda }}_{\mathrm{e}}\left(\mathrm{j}\right)}{\int }}{\mathrm{S}}_{\mathrm{j}}\left(\mathrm{\lambda }\right){\mathrm{d}}_{\mathrm{\lambda }}$$

(1)

Here, ρ_j denotes the resampled canopy reflectance for band j, ρ(λ) represents the original spectral data, λ is the wavelength (nm), λ_s(j) and λ_e(j) denote the start and end wavelengths (nm) for band j, S_j(λ) is the spectral response function, and d_λ denotes the integration step size. The Gaussian spectral response function is calculated as follows:

$${\mathrm{S}}_{\mathrm{j}}\left(\mathrm{\lambda }\right)=\mathrm{e}\mathrm{x}\mathrm{p}\left\{-{\left({\displaystyle \frac{\mathrm{\lambda }-{\mathrm{c}}_{\mathrm{j}}}{\frac{{\mathrm{F}\mathrm{W}\mathrm{H}\mathrm{M}}_{\mathrm{k}}}{2\sqrt{\mathrm{l}\mathrm{n}2}}}}\right)}^2\right\}$$

(2)

where c_j denotes the centre wavelength of band j (nm), and FWHM_k (full width at half maximum) represents the simulated spectral scale, σ = ${\displaystyle \frac{{\mathrm{F}\mathrm{W}\mathrm{H}\mathrm{M}}_{\mathrm{k}}}{2\sqrt{\mathrm{l}\mathrm{n}2}}}$ is the standard deviation.

2.3.2. Feature Band Selection

To reduce the dimensionality of hyperspectral data and select feature bands most correlated with winter wheat SPAD values, this study employed Pearson correlation coefficient analysis for feature selection. Pearson correlation analysis was conducted between reflectance data at different spectral scales and SPAD values, identifying the top 20 bands with the highest correlation coefficients at each scale. At the 50 nm spectral scale, only the top 11 bands were selected.

2.3.3. Inversion Model Construction

This study employs a repeated random partitioning validation method, randomly dividing the entire dataset into a training set (70%) and a test set (30%), and independently repeating this process 100 times. In each iteration, the model was trained on the training set and evaluated on the corresponding test set. The final model performance was represented by the mean of the 100 repeated experiments. Four distinct machine learning algorithms were employed to comprehensively assess their coupling effects with spectral scales:

The fundamental principle of the Support Vector Regression (SVR) model is to minimise the structural risk, seeking a balance between data approximation accuracy and the complexity of the approximation function to achieve optimal model generalisation capability. As a typical application of Support Vector Machines (SVM), SVR possesses the ability to prevent overfitting, with its advantage lying in conveying biophysical variables [44].

The Random Forest (RF) algorithm is an ensemble machine learning method based on regression trees, proposed by Leo Breiman [45]. As a Bagging ensemble algorithm utilising decision trees as its fundamental building blocks, RF relies on the assumption that different independent predictions will misclassify instances in distinct regions. By combining these independent predictions, overall prediction accuracy can be enhanced, demonstrating strong performance in training and learning high-dimensional datasets such as hyperspectral data [46,47,48].

K-nearest neighbours (KNN) constitutes a non-parametric supervised learning method. Within a multidimensional feature space, for any given point, the K nearest neighbouring points are first identified. Subsequently, the predicted value is calculated and derived through weighted estimation based on the feature values of these neighbouring points. The kNN method has become highly prevalent in data mining owing to its ease of implementation and notable classification performance [49].

Ridge regression (RR) is a biased estimation regression method designed to address multicollinearity issues among independent variables in small sample data [50]. This method addresses the substantial reduction in regression accuracy and stability caused by multicollinearity among independent variables in conventional regression analysis. Ridge regression abandons the unbiased estimation of ordinary least squares, sacrificing some information. However, its estimated biased regression coefficients often more closely approximate the true situation, thereby enhancing model stability and reliability. It demonstrates favourable effects in repairing and fitting pathological data.

2.3.4. Model Evaluation Criteria

Model performance is evaluated using Root Mean Square Error (RMSE), Coefficient of Determination (R²), and Relative Percent Deviation (RPD). A higher R² value approaching 1 indicates greater predictive accuracy; a lower RMSE value signifies improved model stability. When RPD < 1.4, the model is deemed unusable; values between 1.4 and 2.0 indicate reasonable reliability; and RPD > 2.0 signifies high reliability. The calculation formulae are as follows:

$${\mathrm{R}}^2={\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{(\mathrm{y}-\overline{\mathrm{y}})}^2}{\sum _{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{y}}_{\mathrm{i}}-\overline{{\mathrm{y}}_{\mathrm{i}}})}^2}}$$

(3)

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left(\mathrm{y}-{\mathrm{y}}_{\mathrm{i}}\right)}^2}{\mathrm{n}}}}$$

(4)

$$\mathrm{RPD}={\displaystyle \frac{\mathrm{SD}}{\mathrm{RMSE}}}$$

(5)

where y and y_i denote the measured chlorophyll content and the model-inverted predicted value for the sample, respectively; $\overline{\mathrm{y}}$ and $\overline{{\mathrm{y}}_{\mathrm{i}}}$ represent the measured mean and the inverted mean value for the sample; n is the number of samples; SD denotes the standard deviation.

3. Results

3.1. Analysis of Canopy Spectral Reflectance Data

Figure 3a presents a box-and-whisker plot of SPAD values for winter wheat at different growth stages, while Figure 3b displays the normal distribution of SPAD values. Combined analysis indicates that SPAD values exhibit an overall trend of initial increase followed by decline as growth progresses, peaking during the heading stage. Figure 4 depicts canopy spectral curves for winter wheat at various growth stages, with spectral reflectance ranging between 0 and 0.5. Due to chlorophyll forming absorption bands in the blue and red regions with relatively weak absorption, the curve exhibits a distinct reflection peak in the green light band (550–600 nm), characteristic of the typical vegetation “green peak” [51]. An absorption trough appears around 680 nm, corresponding to chlorophyll’s strong absorption zone. Reflectance rises sharply between 680 nm and 750 nm, known as the “red edge” effect [52]. Canopy reflectance stabilises within the 750–900 nm range. To validate the spectral integrity after resampling, we compared the curves of typical objects before and after resampling (S1), confirming that their primary spectral features (such as red edges and absorption troughs) were fully preserved without significant reflectance distortion. Before utilizing raw spectral data for modeling, we performed rigorous band filtering to control noise. Due to known severe atmospheric interference and low sensor signal-to-noise ratios in the 325–400 nm and 900–1075 nm bands, these bands were excluded from analysis. Ultimately, all processed input features comprised spectral reflectance data within the 400–900 nm range.

3.2. Correlation Analysis

3.2.1. Single-Band Correlation Analysis at Different Growth Stages

Figure 5 illustrates the variation in correlation coefficients between single-band reflectance and SPAD values across five growth stages as a function of wavelength. Results indicate that the correlation between single-band reflectance and chlorophyll content exhibits consistent trends across all stages. Within the visible spectrum, correlation coefficients predominantly remain negative, with negative correlations within the blue light region (450–495 nm) progressively strengthening as wavelength increases. correlation coefficients fluctuated within the green light region (495–570 nm), with variations in amplitude and position across different growth stages; coefficients in the red light region (620–700 nm) exhibited stable changes, transitioning from negative to positive correlation in the red edge region (700–750 nm) before stabilising in the near-infrared region (760–900 nm). Within the visible spectrum, the maximum positive correlation between leaf SPAD values and spectral reflectance occurred at 750 nm, with a correlation coefficient of 0.696. Within the near-infrared spectrum, the maximum positive correlation between leaf SPAD values and leaf spectral reflectance occurred at 813 nm, with a correlation coefficient of 0.780. As indicated by the yellow data in the figure, the characteristic bands are distributed across different wavelength ranges. The bands exhibiting strong correlations are primarily concentrated between 570 nm and 710 nm, with correlation coefficients ranging from −0.9 to −0.3.

Figure 6 presents the characteristic bands identified at different spectral scales across growth stages alongside their corresponding reflectance values. Combining Figure 4 and Figure 5, the five growth stages exhibit three distinct trends at the 1 nm spectral scale: the jointing stage shows the weakest negative correlation, with correlation coefficients around −0.40, and characteristic bands primarily concentrated in the red light band; The booting stage and flowering stage exhibited similar concentrations of characteristic bands within the red and red-edge bands, though the negative correlation for booting was marginally stronger than for flowering, with correlation coefficients around −0.75; The correlation coefficients for the heading stage and filling stage fell within the same range, both exceeding −0.88. Characteristic bands for the heading stage primarily concentrated in the yellow and red-edge bands, while those for filling mainly focused on the red and red-edge bands.

3.2.2. Correlation Analysis Across Different Spectral Scales at Various Growth Stages

As shown in Figure 6, with increasing sampling width, the characteristic bands corresponding to each spectral scale expanded uniformly in both directions from the 1 nm characteristic band; Except for the heading stage, characteristic bands for all other stages predominantly resided within the red light region. As spectral scale increased, these bands expanded towards the green and blue light regions (leftward) and the red edge and near-infrared regions (rightward). Reflectance values were ranked sequentially as follows: heading stage, filling stage, booting stage, flowering stage, and jointing stage.

Table 2. Maximum correlation coefficients between single-band spectral reflectance and SPAD values across eight spectral scales following feature screening. Note: MAX represents the maximum correlation coefficient.

Spectral Scale(nm)	Jointing		Booting		Heading		Flowering		Filling
	Band	Max	Band	Max	Band	Max	Band	Max	Band	Max
1	412	−0.520	615	−0.746	701	−0.889	698	−0.732	651	−0.877
2	412	−0.429	696	−0.744	702	−0.887	642	−0.731	642	−0.875
3	411	−0.408	696	−0.743	702	−0.887	642	−0.730	651	−0.875
5	640	−0.404	695	−0.741	700	−0.885	695	−0.730	695	−0.874
7	644	−0.403	616	−0.735	700	−0.885	644	−0.729	693	−0.874
10	640	−0.403	620	−0.734	700	−0.885	640	−0.729	650	−0.873
20	640	−0.401	620	−0.734	580	−0.884	640	−0.728	640	−0.873
50	650	−0.398	650	−0.728	600	−0.883	650	−0.726	650	−0.872

presents the maximum correlation coefficients between SPAD readings and winter wheat spectral reflectance at different spectral scales, along with their corresponding spectral bands. The findings indicate that the absolute values of maximum correlation coefficients across the five growth stages remain comparable under varying spectral scales, consistent with Cho’s [53] research conclusions. As growth stages progress, the optimal spectral regions consistently reside within the red and red-edge bands. However, differences exist in the specific spectral scales: at 1 nm, 2 nm, 3 nm, and 10 nm scales, the maximum correlation band shifts from the red to the red-edge band before reverting to the red; at 5 nm, it shifts from red to red-edge; at 7 nm, it exhibits a cyclical pattern of red light-red edge-red light-red edge. The 20 nm scale follows a red-yellow-red distribution, whilst the 50 nm scale concentrates entirely on the 650 nm wavelength within the red light region. Across different spectral scales, the core band centre wavelengths primarily cluster around 700 nm in the red edge and 650 nm in the red light region, with the 20 nm and 50 nm scales focusing exclusively on the 650 nm band.

3.3. Machine Learning Model Estimation at Different Spectral Scales

This study employed feature bands selected via Pearson correlation analysis as model input variables. Combining four machine learning algorithms (RR, RF, KNN, SVR), a total of 192 SPAD estimation models were constructed across different growth stages. Model evaluation metrics comprised R², RMSE, and RPD. Figure 7 presents the test set modelling results for the four machine learning models under varying spectral scales and growth stages. Overall, model accuracy for the entire growth period, booting stage, flowering stage, and filling stage generally increased with higher spectral scale, whereas accuracy declined during the jointing stage and heading stage. Models for the heading and filling stages exhibited higher and more consistent accuracy, with R² values ranging from 0.70 to 0.95. Models for the jointing stage demonstrated the lowest accuracy, with R² consistently below 0.50. The estimation accuracy of models for the entire growth period, boot stage, and flowering stage showed greater variability, with R² values ranging from 0.4 to 0.9.

Comparing the estimation performance of four models for SPAD across different growth stages and spectral scales. The RR model demonstrated optimal performance across all scales during the entire growth period and most scales during the booting stage, with model accuracy significantly improving as the spectral scale increased. Optimal model accuracy was achieved at spectral scales of 20 nm and 10 nm, yielding R² values of 0.774 and 0.622, respectively—representing increases of 0.108 and 0.034 compared to the 1 nm resolution. The RF model demonstrated superior performance during the filling stage, achieving an R² of 0.939—an improvement of 0.217 over the 1 nm spectral scale. Its RPD of 3.46 significantly exceeded the reliability threshold. The model also performed well at larger scales (10–50 nm), with R² values exceeding 0.91 and RPD exceeding 2.98. The KNN model demonstrated significant advantages during the heading and flowering stages (R² > 0.83). Although the overall mean R² of the flowering stage KNN model did not exceed that of the RR model, its RMSE and RPD data were superior. The SVR model demonstrated the weakest overall performance (mean R² = 0.546), with the best R² across all growth stages reaching only 0.565. During the grain filling stage, the RMSE was as high as 7.328, indicating that under the current experimental setup and data conditions, the predictive efficacy of the SVR model is suboptimal.

Comparing SPAD estimation performance across different growth stages using varying spectral scales and models (Figure 8). Throughout the entire growth period, the RR model at the 20 nm spectral scale demonstrated optimal estimation performance (R² = 0.77, RMSE = 3.42, RPD = 2.06), followed by the RR model at the 50 nm spectral scale. The optimal model for both the jointing and heading stages was the KNN model at the 1 nm spectral scale. Despite consistent spectral scale and inversion models, significant differences emerged due to distinct growth stages, with R² values at jointing being lower than those at heading (0.435). The optimal models for the heading and filling stages were both the RF model at a 20 nm spectral scale, yielding inversion accuracies of R² = 0.617 and 0.939, RMSE = 2.189 and 2.725, and RPD = 1.596 and 3.457, respectively. Although the R² at 20 nm during the heading stage was lower than that at 10 nm (0.622), it exhibited the lowest RMSE and highest RPD, making it the optimal choice when considering overall accuracy. The optimal model for the flowering stage was the KNN model at a 50 nm spectral scale, achieving an R² of 0.837. Compared to the original 1 nm data, this resulted in an R² improvement of 0.09, an RMSE reduction of 3.10, and an RPD increase of 1.33.

In summary, different models are suited to different growth stages, with distinct spectral scales matching specific growth stage characteristics. SPAD inversion must adhere to the triadic relationship of “scale-model-growth stage”.

4. Discussion

4.1. Correlation Analysis Between Canopy Spectral Data and SPAD at Different Growth Stages

Models for estimating SPAD values in winter wheat based on hyperspectral remote sensing data have yielded favourable results in previous studies and gained widespread application [54,55]. Consequently, in-depth analysis of the correlation between canopy spectral data and SPAD values aids in comprehensively understanding winter wheat growth conditions. This study conducted correlation analyses between spectral data and winter wheat SPAD values across different growth stages, revealing variations in the results (Figure 3). The overall trend in spectral data correlation with SPAD values across growth stages was upward. The weakest negative correlation with SPAD was observed during the jointing stage, while correlations strengthened during the booting and heading stages. The highest correlation in the visible spectrum was recorded during heading, with a subsequent decline during flowering. Correlation increased again during filling, reaching its peak in the near-infrared spectrum. The primary reasons are as follows: during the jointing stage, the plant canopy is not yet closed, with low coverage, leading to soil background noise affecting the spectral signal. During the booting and heading stages, the canopy gradually closes, reducing the influence of the soil background and enhancing the correlation. During flowering, floral organs exhibit high reflectance, potentially interfering with and attenuating chlorophyll absorption signals. At filling, leaf senescence occurs with chlorophyll degradation, and nitrogen and carbohydrates are transferred from leaves to grains [56]. Leaf cell structures dehydrate and shrivel, reducing scattering [57], thereby decreasing reflectance and increasing the correlation between canopy spectra and SPAD readings. As shown in Figure 5, the sensitive bands across all five growth stages are concentrated in the red and red-edge regions, primarily due to chlorophyll’s strong absorption of red light. This aligns with findings from Buscaglia [58] and Zhao [59]. Although this study successfully revealed the spectral response patterns of winter wheat SPAD values, the robustness of the model is inevitably influenced by specific years and natural conditions. Therefore, the universality of these conclusions across years and regions requires further validation. Future research will focus on collecting multi-year, multi-location data and integrating environmental variables into the model framework to develop a monitoring model with enhanced spatio-temporal generalization capabilities.

4.2. Interaction Between Spectral Scale and Model Performance

4.2.1. Evaluation of SPAD Model Performance at Different Spectral Scales

Spectral scale conversion aims to adapt spectral wavelength resolution or band response characteristics to suit different spectrometers or specific application requirements [60]. Utilising Gaussian response functions or sensor band characteristics, high-resolution spectral data can be converted into equivalent band reflectance for the target platform. This preserves core spectral features while matching the sensor’s data format. For instance, Li et al. [61] simulated the wide-band spectral reflectance characteristics of GF-1 satellite data based on measured winter wheat hyperspectral data, constructing a remote sensing monitoring model for winter wheat leaf SPAD values at different growth stages using GF-1 data. Z. Oumar et al. [62] resampled field-spectrometer (Fieldspec 3 Pro FR) measurements to the 10 nm calibration bands of the Hyperion sensor, successfully evaluating their predictive capability for Thaumastocoris peregrinus damage. Spectral scale conversion reduces data dimensions, decreases computational complexity and processing time, thereby enhancing analytical efficiency [63].

As discussed in Section 2.3, model accuracy (R² and RPD) for the entire growth period, booting stage, flowering stage, and filling stage generally exhibited a trend of first decreasing and then increasing with increasing spectral scale, whereas RMSE showed the opposite pattern. Across these four stages, model accuracy changed slowly within the first five spectral scales, with a trough occurring around 3–7 nm and a marked inflection point near 10 nm, followed by accelerated changes thereafter. Model accuracy (R² and RPD) for the jointing and heading stages exhibited an overall declining trend, with an upward peak around 10 nm, while RMSE showed the opposite variation. Overall, compared to 1 nm high-resolution data, larger spectral scales (10, 20 nm) enhanced the accuracy of SPAD monitoring for winter wheat, whereas medium and small resolutions (5, 7 nm) potentially reduced accuracy. This may occur because large-scale (10, 20 nm) spectral scale enhances signal-to-noise ratio (SNR) through spectral band aggregation, whereas extremely large-scale (>20 nm) resolution utilises only specific bands within the solar spectrum, potentially causing spectral information loss when integrating broader bandwidths [64]; Although medium- and small-scale (5, 7 nm) resolutions show improvement, they fail to effectively filter noise, leading to reduced inversion accuracy. This conclusion aligns with the findings of Thenkabail et al. [65], where excessively narrow bandwidths (<5 nm) significantly impair SNR, while overly broad bandwidths (>30 nm) result in loss of specific narrow-band information. Thenkabail et al. [21] found that for five crops, nine growth stages, and multiple traits, the vast majority of bandwidths closely associated with crop characteristics fell between 10 nm and 20 nm. Furthermore, Broge et al. [66] indicated that the optimal narrow bandwidth is approximately 15 nm.

This study employed only the 1–50 nm scale, with some sensors achieving spectral resolution of 100 nm. Future experimental designs should be more comprehensive, taking into account the characteristics of different sensors and platforms. Sampling should be performed based on their respective spectral response functions to achieve more accurate data simulation.

4.2.2. Performance Evaluation of Machine Learning Algorithms in SPAD Estimation

Previous studies have demonstrated the significant potential of machine learning in predicting chlorophyll content in winter wheat leaves [67,68,69]. This research employed four algorithms—RR, RF, KNN, and SVR—to construct SPAD prediction models for five growth stages of winter wheat based on eight spectral scale feature bands. Results indicate that RR, KNN, and SVR models achieved the highest R² (0.89, 0.86, 0.88) and RPD (2.94, 2.57, 1.32) at 10 nm during the filling stage. The RF model attained the highest R² (0.94) and RPD (3.46) at 20 nm during this stage.

A comprehensive comparison of 576 evaluation metrics across different spectral scales, machine learning algorithms, and growth stages identified optimal modelling approaches and accuracies for each scale and stage (

Table 3. The best model under different spectral scales and growth stages.

Spectral Scale (nm)	Entire Period			Jointing			Booting			Heading			Flowering			Filling
	R2	RMSE	RPD	R2	RMSE	RPD	R2	RMSE	RPD	R2	RMSE	RPD	R2	RMSE	RPD	R2	RMSE	RPD
1	0.67	4.24	1.66	0.41	5.73	1.21	0.59	2.56	1.37	0.85	2.06	2.49	0.62	3.74	1.45	0.75	4.80	1.96
2	0.67	4.24	1.66	0.37	5.81	1.19	0.58	2.60	1.34	0.83	2.19	2.34	0.63	3.73	1.45	0.75	4.80	1.96
3	0.68	4.20	1.67	0.34	5.89	1.18	0.58	2.63	1.33	0.82	2.24	2.29	0.63	3.73	1.45	0.75	4.76	1.98
5	0.69	4.10	1.71	0.29	6.00	1.15	0.56	2.33	1.50	0.82	2.25	2.27	0.63	3.73	1.45	0.77	4.57	2.06
7	0.72	3.93	1.79	0.23	6.28	1.10	0.56	2.33	1.50	0.82	2.25	2.28	0.62	3.86	1.41	0.75	4.77	1.97
10	0.76	3.53	1.99	0.30	5.97	1.16	0.62	2.54	1.38	0.82	2.25	2.28	0.63	3.91	1.39	0.93	2.75	3.42
20	0.77	3.41	2.06	0.19	6.39	1.08	0.62	2.19	1.60	0.80	2.31	2.21	0.82	3.17	1.71	0.94	2.72	3.46
50	0.77	3.49	2.01	0.18	6.41	1.08	0.60	2.37	1.47	0.78	2.41	2.12	0.84	2.33	2.33	0.91	3.17	2.98

Note: Grey cells denote the RR model, green cells denote the KNN model, and blue cells denote the RF model. Bolded text indicates the optimal model under different growth stages.

). Results indicate KNN demonstrated superior and consistent performance, achieving the highest accuracy in 21 out of 48 models (43.75%) across scales and stages. The predictive accuracy of the KNN model remained above 0.70 in most cases, reaching a maximum of 0.90. This indicates that the SPAD values predicted by the KNN model exhibit a high correlation with actual values, rendering the predictions reliable. As an instance-based algorithm, KNN relies on the similarity of neighbouring samples in the feature space for prediction, effectively capturing the non-linear relationships between local samples in hyperspectral data [70]. Second was the RR model, achieving 15 best accuracies (31.25%). It demonstrated strong generalisation capability and stability, likely because RR is a linear regression model with L2 regularisation, making it less prone to overfitting and effectively addressing multicollinearity issues in high-dimensional spectral data [71]. Although the RF algorithm demonstrated favourable predictive accuracy at certain large scales (e.g., 10 nm, 20 nm) with R² exceeding 0.92, it exhibited severe overfitting, resulting in unstable model performance and limited adaptability to the complex and variable conditions encountered in field environments. This may stem from excessive model complexity, leading to overfitting of training data details and noise, thereby compromising generalisation capabilities [72]. Compared to the former model, the SVR model yields overall lower inversion results. This may be attributable to the parameters (C, γ) and kernel function not being optimally tuned, thereby failing to effectively capture the complex relationships within the data [73]. System hyperparameter optimization is an essential approach to mitigate overfitting and enhance model generalization capabilities. Future research will prioritize hyperparameter tuning for systems, employing advanced strategies such as grid search or Bayesian optimization to identify optimal parameter configurations for different algorithms and scenarios. This aims to achieve more stable and robust inversion models. Simultaneously, effectively applying deep learning methods like convolutional neural networks (CNNs) for spectral analysis to enhance model performance and complex data processing capabilities will also be a key focus of future research [74].

When further exploring the application of spectral scales and machine learning methods, ensuring the quality and accuracy of raw data is essential. Hyperspectral data features numerous bands with complex structures and highly correlated information across bands, making it prone to data collinearity and redundancy issues [75]. Appropriate data preprocessing methods, such as first-order derivatives and SG smoothing, must be employed to enhance data precision.

5. Conclusions

This study addressed the core issue of spectral scale variations across multi-platform remote sensing data affecting crop physiological parameter inversion accuracy. By integrating Gaussian resampling techniques with machine learning algorithms, it systematically elucidated the influence mechanism of spectral scale on SPAD inversion for winter wheat. Key conclusions are as follows:

(1)

The correlation between winter wheat canopy spectral data and SPAD values exhibits distinct developmental stage dynamics, with sensitive spectral bands concentrated in the 570–710 nm range. The highest correlation was observed during the heading stage (r = −0.89), while the lowest was recorded during the jointing stage (r = −0.41), indicating that changes in vegetation physiological processes directly influence spectral response intensity.

(2)

Spectral resolution systematically influences model performance. Compared to 1 nm raw data, larger scales like 10 nm and 20 nm effectively suppress noise interference and enhance model robustness. Conversely, intermediate scales such as 5 nm and 7 nm reduce model accuracy due to information redundancy and noise amplification effects.

(3)

This study first clarifies the dynamic adaptation rules among “scale-model-growth stage”: During the jointing and heading stages where vegetative growth and reproductive growth coexist, the 1–5 nm fine scale combined with the KNN algorithm achieves optimal monitoring. In contrast, during the booting stage and the flowering and grain-filling stages dominated by reproductive growth, the 20–50 nm spectral scale combined with either the KNN or RF algorithm proves more suitable.

Spectral scale conversion achieves data comparison and fusion by standardizing multi-source spectral response functions, effectively resolving spectral baseline shifts between sensors. This provides foundational support for cross-platform data fusion and interoperability, playing a crucial role in collaborative satellite-UAV remote sensing analysis to significantly enhance data reliability and model generalization capabilities for large-scale agricultural monitoring [76,77]. Research indicates that when using UAVs for crop monitoring, sensors should be dynamically selected based on growth stages: hyperspectral sensors should be prioritized in early stages to preserve fine spectral information, while multispectral sensors with matched band centers and widths (e.g., 20–50 nm) can be used in later stages to optimize cost and accuracy. Beyond UAV-based near-surface remote sensing, our framework integrates with satellite-based spaceborne remote sensing to achieve precise, timely estimation of multi-scale crop traits. For instance, the Gaofen-5 satellite (VNIR: 5 nm, SWIR: 10 nm) enables direct application of research outcomes to guide satellite data preprocessing and monitoring algorithm selection. Future work will validate the optimal models developed in this study across ecological regions using China’s typical land cover spectral database and multi-regional independent datasets, advancing their integration into operational crop growth monitoring systems.