Monitoring Chlorophyll Content of Brassica napus L. Based on UAV Multispectral and RGB Feature Fusion

Abstract

Accurate prediction of chlorophyll content in Brassica napus L. (rapeseed) is essential for monitoring plant nutritional status and precision agricultural management. The current study focuses on single cultivars, limiting general applicability. This study used unmanned aerial vehicle (UAV)-based RGB and multispectral imagery to evaluate six rapeseed cultivars chlorophyll content across mixed-growth stages, including seedling, bolting, and initial flowering stages. The ExG-ExR threshold segmentation was applied to remove background interference. Subsequently, color and spectral indices were extracted from segmented images and ranked according to their correlations with measured chlorophyll content. Partial Least Squares Regression (PLSR), Multiple Linear Regression (MLR), and Support Vector Regression (SVR) models were independently established using subsets of the top-ranked features. Model performance was assessed by comparing prediction accuracy (R² and RMSE). Results demonstrated significant accuracy improvements following background removal, especially for the SVR model. Compared to data without background removal, accuracy increased notably with background removal by 8.0% (R²p improved from 0.683 to 0.763) for color indices and 3.1% (R²p from 0.835 to 0.866) for spectral indices. Additionally, stepwise fusion of spectral and color indices further improved prediction accuracy. Optimal results were obtained by fusing the top seven color features ranked by correlation with chlorophyll content, achieving an R²p of 0.878 and an RMSE of 52.187 μg/g. These findings highlight the effectiveness of background removal and feature fusion in enhancing chlorophyll prediction accuracy.

1. Introduction

Brassica napus L. (rapeseed) is a critical oilseed crop, and effective monitoring of its field growth status is essential to ensure edible oil security [1]. Chlorophyll (Chl), a vital pigment involved in photosynthesis within plant leaves [2], not only directly influences photosynthetic efficiency but also closely correlates with crop nutritional status, physiological traits, and final yield. Previous studies have demonstrated that chlorophyll content can effectively reflect plant nitrogen status [3,4] for guiding precise nitrogen fertilizer application and optimizing crop growth conditions [5]. Additionally, the relationship between chlorophyll content and crop physiological traits has been widely validated [6,7]. Therefore, accurate monitoring of chlorophyll content is crucial for precise assessment of crop growth conditions.

Traditional methods for chlorophyll measurement mainly include laboratory chemical analysis [8] and portable chlorophyll meter methods, like SPAD-502 [9]. Chemical analysis is time-consuming, labor-intensive, and unable to provide timely field data. The latter, while enabling rapid measurement, is limited to the single-leaf scale, making it unlikely for efficient monitoring at the regional scale [10,11]. In recent years, with the rapid development of remote sensing technology, non-destructive and efficient remote sensing techniques for chlorophyll content monitoring have become widely adopted [12,13]. Unmanned aerial vehicle (UAV)-based remote sensing platforms, due to their high flexibility, convenience, and high spatial resolution, have emerged as an important tool in agricultural monitoring, demonstrating significant advantages in the rapid and non-destructive detection of crop physiological parameters [14].

Although UAV remote sensing has made progress in chlorophyll monitoring for crops such as rice [15,16], wheat [4,17,18], cotton [19], soybean [20], and peanut [21], previous studies often overlook the influence of background information on spectral feature extraction and remote sensing accuracy. To improve the sensitivity of spectral features to leaf biochemical parameters, various vegetation indices have been developed to reduce background interference [22,23]. For example, the soil-adjusted vegetation index reduces soil background effects by including soil adjustment coefficients [24]; indices like the optimized soil-adjusted vegetation index and modified chlorophyll absorption ratio index have been effective in accurately estimating chlorophyll content at the canopy level [14,25]. However, relying on a single type of feature for chlorophyll monitoring may not fully capture the actual physiological status of crops, potentially reducing model accuracy.

This study aims to use UAV-based RGB and multispectral sensors to investigate how image segmentation and background removal affect chlorophyll detection accuracy. Detection machine learning models for chlorophyll content in rapeseed leaves at seedling, bolting, and initial flowering stages will be developed, examining the influence of color and spectral features. Specifically, this study will (1) analyze dynamic changes in rapeseed canopy chlorophyll content and vegetation coverage under different nitrogen fertilizer levels; (2) evaluate the impact of background removal on improving chlorophyll prediction accuracy; and (3) assess the estimation accuracy of color and spectral features for canopy chlorophyll content, exploring improvements achieved by integrating these features.

2. Materials and Methods

2.1. Sample Preparation

This study selected six rapeseed cultivars (genotypes), including QinYou 908 (QY908), ZheDa 619 (ZD619), ZheDa 635 (ZD635), ZheDa 649 (ZD649), ZheDa Fense (ZDFS), and ZheDa Juhuang (ZDJH). The selected varieties include widely cultivated, resistant, and adaptable main varieties (like QY908 and ZD619), high-oil-quality varieties (like ZD635 and ZD649), and innovative materials with different flower colors (like ZDFS and ZDJH), thus ensuring diversity and representativeness to enhance model generalization and robustness.

The field experiment was conducted from October 2023 to May 2024 at Hangzhou Raw Seed Growing Farm, Zhejiang Province, China (30°22′54.19″ N, 119°55′29.49″ E). The local climate is subtropical monsoon, with an average temperature of 17.8 °C and annual precipitation of 1050–1250 mm. The previous crop was rice, and the soil contained organic matter at 32.3 g/kg, total nitrogen at 1.63 g/kg, available phosphorus at 25.86 mg/kg, and available potassium at 121 mg/kg. Seedlings of the six rapeseed varieties were transplanted on 18 November 2023 to field plots, each covering 8 m² with a planting density of 9 plants/m². Six nitrogen fertilizer levels (N0, N1, N2, N3, N4, and N5), corresponding to 0, 60, 120, 180, 240, and 300 kg/ha, respectively, were applied in a randomized block design with three replicates, totaling 108 plots (Figure 1). Urea (46% N) was applied in two splits: 60% on the day of transplanting and 40% on 16 December 2023. Equal amounts of phosphate fertilizer (90 kg/ha, P₂O₅ 12%), potassium fertilizer (150 kg/ha, K₂O 60%), and boron fertilizer (15 kg/ha, B 11%) were applied as base fertilizer to support rapeseed growth.

2.2. UAV Data Collection and Processing

In this study, a DJI Mavic 3 Multispectral UAV platform (M3M, DJI Innovation Technology Co., Ltd., Shenzhen, China) equipped with multispectral sensors was used for data acquisition (Figure 2). M3M is equipped with 4 multispectral sensors and 1 visible light lens. It can obtain single-channel images in the green (Green, G), red (Red, R), red edge (RedEdge, RE), and near-infrared (Near Infrared, NIR) bands, as well as high-precision visible light images. Flight operations were conducted between 10:00 and 14:00 under sunny, windless conditions on 27 January 2024 (seedling stage), 3 March 2024 (bolting stage), and 22 March 2024 (initial flowering stage). Each flight was performed at an altitude of 12 m and a speed of 1 m/s, with the camera oriented vertically to capture orthophotos of the rapeseed canopy. The overlap and side overlap rates were set at 70% and 80%, respectively, to ensure good image stitching quality. To minimize this variability, two 25 × 25 cm diffuse reflectance reference panels (SphereOptics GmbH, Herrsching, Germany) were used to calibrate spectral images. Before each flight, two reference panels (models SG20081913 and SG20081921), with reflectance values of 80% and 30%, respectively, were placed on open ground near the experimental area. The UAV hovered at a height of 1.5 m above these panels to capture calibration images. DJI Terra software (DJI Innovation Technology Co., Ltd., Shenzhen, China) was used to process and stitch the captured images, enabling calibration, alignment, and generation of orthophotos of the experimental site.

To visually demonstrate the effects of different nitrogen treatments and growth stages on canopy coverage, representative RGB images of rapeseed cultivar QY908 at the seedling, bolting, and initial flowering stages under six nitrogen treatments (N0–N5) are presented in Figure 2b. Clear differences can be observed in canopy coverage as the plants developed and nitrogen supply increased. Specifically, during the bolting and initial flowering stages, higher nitrogen treatments (N4 and N5) promoted significantly greater canopy closure and biomass accumulation compared to low nitrogen treatments (N0 and N1), which displayed notably sparse vegetation coverage. These visible differences further represent the effectiveness of using UAV-based multispectral remote sensing to monitor the chlorophyll status of rapeseed.

2.3. Measurements of Chl Content

Leaf samples were collected using a three-point sampling method. Eight rapeseed plants with similar growth status were selected from the front, middle, and back areas of each plot, respectively. The newest fully expanded leaves from each plant were collected, mixed, and used as a single sample. A total of 324 samples were collected at each growth stage, resulting in 972 samples throughout the entire experiment. After UAV data acquisition, each leaf was placed in a labeled and sealed bag and temporarily stored in an icebox. Samples were quickly transported to the laboratory at Zhejiang University (Hangzhou, China) for chlorophyll and nitrogen content analysis. Leaf discs were collected using a hole punch with a diameter of 0.6 cm and immersed in 4 mL of 95% ethanol and stored in darkness for approximately 48 h until the discs became white, indicating complete chlorophyll extraction. Absorbance measurements of the extracted solutions were conducted at wavelengths of 470, 649, and 665 nm using a spectrophotometer (Epoch, BioTek Instruments, Winooski, VT, USA). Chlorophyll content was calculated based on previously published equations [26]. For nitrogen content determination, the total nitrogen content of the leaf samples was determined by the Kjeldahl method and was calculated and expressed as g/kg dry weight.

2.4. Data Analysis Methods

2.4.1. Image Feature Extraction

To fully utilize the multispectral imagery from the UAV platform and construct robust models, this study extracted color and spectral features related to chlorophyll content in rapeseed leaves. Background pixels, such as soil and crop shadows in original images, could negatively impact the selection of sensitive features for physical and chemical parameters. Thus, it was essential to remove non-crop pixels from the region of interest. Meyer et al. [27] demonstrated that the Excess Green minus Excess Red (ExG-ExR) vegetation index effectively differentiates vegetation from the background. Considering the multispectral sensor characteristics and image resolution used in this study, pixels with ExG-ExR values greater than 0.02 were classified as rapeseed pixels. Firstly, a canopy mask was created using RGB imagery, matching the region based on geographic coordinates. Then, vegetation pixels were separated from background pixels through threshold segmentation (Figure 3). Finally, reflectance values for visible RGB bands and multispectral R, B, RedEdge, and NIR bands within each plot were extracted. Three types of reflectance were calculated: Average reflectance of all pixels (AllPix), average reflectance of rapeseed pixels (GreenPix), and background pixel reflectance (BgPix).

Various color indexes (

Table 1. Color index established based on color features.

Color Index	Formula	Reference
R	R	/
G	G	/
B	B	/
NRI	R/(R + G + B)	[28]
NGI	G/(R + G + B)	[28]
NBI	B/(R + G + B)	[28]
G/R	G/R	[28]
G/B	G/B	[28]
R/B	R/B	[28]
GRRI	R/G	[29]
VIGreen	(G − R)/(G + R)	[29]
VIB,R	(B − R)/(B + R)	[29]
VIG	2G × (R − B)/(R + B)	[29]
VIR	2R × (G − B)/(G + B)	[29]
VARI	(NGI − NRI)/(NGI + NRI − NBI)	[30]
GLI	(2 × NGI ‒ NBI − NRI)/(2 × NGI + NBI + NRI)	[30]
GLI_ORI	(2 × NGI − NBI − NRI)/(−NBI − NRI)	[30]
MGRVI	(NGI2 − NRI2)/(NGI2 + NRI2)	[29]
RGBVI	(NGI2 − NBI × NRI)/(NGI2 + NBI × NRI)	[29]
NDYI	(NGI − NBI)/(NGI + NBI)	[29]
CIVE	0.441 × NRI − 0.811 × NGI + 0.385 × NBI + 18.78745	[29]
VEG	NGI/(NRI0.667 × NBI(1−0.667))	[29]
IPCA	0.994 × |NRI − NBI| + 0.961 × |NGI − NBI| + 0.914 × |NGI − NRI|	[29]
H	Arccos (0.5 × [(R − G) + (R − B)]/[(R − G)2 + (R − B)(G − B)]0.5)	[31]
S	1 − (3 × [min (R,G,B)])/(R + G + B)	[31]
I	(R + G + B)/3	[31]
L*	116 × (0.299R + 0.587G + 0.114B)1/3 ‒ 16	[29]
a*	500 × [1.006 × (0.607R + 0.174G + 0.201B)1/3 − (0.299R + 0.587G + 0.114B)1/3]	[29]
b*	200 × [(0.299R + 0.587G + 0.114B)1/3 − 0.846 × (0.066G + 1.117B)1/3]	[29]

Note: R, G, and B represent the reflectance values of the red band, green band, and blue band in an RGB image. * in L, a, and b denotes values in the CIELAB color space, where L indicates lightness, a represents the green–red axis, and b represents the blue–yellow axis.

) and vegetation indexes (

Table 2. Vegetation indices established based on spectral characteristics.

Vegetation Index	Formula	Reference
Green	R560	/
Red	R650	/
RedEdge	R730	/
NIR	R860	/
NDVI	(R860 − R650)/(R860 + R650)	[32]
GNDVI	(R860 − R560)/(R860 + R560)	[32]
CIgreen	R860/R560 − 1	[33]
CIre	R860/R730 − 1	[33]
NDRE	(R860 − R730)/(R86 + R730)	[34]
OSAVI	(R860 − R650)/(R860 + R650 + 0.16)	[35]
SAVI	(1 + L) × (R860 − R650)/(R860 + R650 + L)	[36]
RVI	R860/R650	[37]
DVI	R860 − R650	[37]
MCARI	[(R730 − R650) − 0.2 × (R730 − R560)](R730/R650)	[38]
TVI	0.5 × (120 × (R860 − R560) – 200 × (R650 − R560))	[39]
MSAVI	(2 × R860 + 1 – sqrt ((2 × R860 + 1)2 − 8 × (R860 − R650)))/2	[40]
NRI	(R560 − R650)/(R560 + R650)	[41]
MTCI	(R860 − R730)/(R730 − R650)	[42]

Note: R₅₆₀, R₆₅₀, R₇₃₀, and R₈₆₀ represent the reflectance values of the green band, red band, red-edge band, and near-infrared band, respectively; L is the soil adjustment coefficient, and in this paper, it is set to 0.5.

) commonly used in published literature were computed from these extracted reflectance values. In this study, 29 color indexes and 18 vegetation indexes were utilized as the basis for subsequent research.

2.4.2. Predictive Model Construction

Partial Least Squares Regression (PLSR), Support Vector Regression (SVR), and Multiple Linear Regression (MLR) were utilized to estimate leaf Chl content based on spectral and color features extracted from UAV images.

PLSR is widely used in regression modeling for high-dimensional datasets. It is capable of fitting the linear regression relationship between variables and Chl content values [43]. The key hyperparameter optimized for PLSR was the number of latent variables, which was selected through a grid search approach ranging from 1 to 50. The optimal latent variable number was determined based on minimizing the root mean square error during 10-fold cross-validation on the calibration dataset.

SVR is a powerful machine learning algorithm known for its strong generalization ability, which is effective in solving high-dimensional problems. SVR maps variables and target values to a high-dimensional feature space through a nonlinear transformation, where a linear decision function is constructed to perform the regression [44]. The Radial Basis Function (RBF) kernel was selected for its superior ability to capture nonlinear relationships between spectral and color indices and chlorophyll content. Two crucial hyperparameters were optimized for the SVR model: The regularization parameter (c) can control the trade-off between model complexity and prediction error, and the kernel parameter (g) can influence the radius of influence of individual training samples. These two hyperparameters were optimized using a systematic grid search; the range for c was set from 10⁻⁵ to 10⁵, and for g from 10⁻⁶ to 10¹.

MLR is a statistical technique that uses several explanatory variables to predict the outcome of a response variable. The goal of MLR is to model the linear relationship between the independent variables and Chl content [45]. Unlike PLSR and SVR, MLR is deterministic and does not require hyperparameter tuning. The regression coefficients in the MLR model were directly estimated from the calibration dataset via ordinary least squares, providing a baseline comparison against the more complex regression models.

To rigorously evaluate model accuracy and predictive performance, the dataset was divided into a calibration set (70%) and a prediction set (30%). All hyperparameter optimizations and model training processes utilized a 10-fold cross-validation strategy on the calibration set, ensuring robustness and generalizability of the developed models.

2.4.3. Software and Model Evaluation

All data processing, model construction, and analysis in this study were performed on a workstation running the Windows 10 operating system. The hardware was configured with an Intel^® Core™ i9-12900K CPU, an NVIDIA^® GeForce^® RTX™ 3090 Ti GPU, and 128 GB of RAM. The predictive models were implemented in Python v3.9, primarily utilizing the scikit-learn library.

The Pearson correlation coefficient (r) was used to calculate the correlation between variables. The value of r ranges from −1 to 1, where a value of 1 indicates a perfect positive correlation, −1 indicates a perfect negative correlation, and 0 indicates no linear correlation between the variables. Additionally, one-way analysis of variance (ANOVA) was employed as a statistical method to compare the means among multiple groups. The performance of the developed models was evaluated using the coefficient of determination (R²) and the root mean square error (RMSE). To comprehensively assess model accuracy and robustness, these metrics were calculated for the calibration set (R²_c, RMSE_c) and the prediction set (R²_p, RMSE_p).

3. Results

3.1. Analysis of Chl Content

The changes in canopy leaf chlorophyll content for six rapeseed varieties under different nitrogen treatments across the seedling, bolting, and initial flowering stages are presented in Figure 4. As shown in the figure, the experimental plots with no nitrogen application (N0) exhibited the lowest chlorophyll content across all growth stages. Conversely, as the nitrogen level increased from N1 to N5, the chlorophyll content showed a generally monotonic upward trend. Notably, during the bolting and initial flowering stages, the chlorophyll content corresponding to the N4 and N5 treatments was significantly higher than that of the N0 to N3 treatments. Furthermore, for the majority of the varieties, the difference in chlorophyll content between the N4 and N5 treatments was not significant, suggesting that the chlorophyll in the canopy leaves was nearly saturated once the nitrogen application reached the N4 level.

A comparison across the three growth stages reveals that chlorophyll content accumulated significantly as the plants developed. During the seedling stage, chlorophyll content was approximately 400–500 μg/g. Upon entering the bolting stage, the content for all treatments was significantly higher than in the seedling stage, with most varieties under the N4 and N5 treatments reaching 550–650 μg/g. At the initial flowering stage, chlorophyll content increased further, with levels in the N4 and N5 treatments commonly exceeding 750 μg/g and reaching 800 μg/g in some varieties. ANOVA indicated that the differences among nitrogen treatments were not yet significant during the seedling stage. However, during the bolting and initial flowering stages, the gap between treatments progressively widened, with a particularly significant difference observed between low-nitrogen (N0) and high-nitrogen (N4, N5) treatments. This suggests that the dependence of chlorophyll content on nitrogen supply substantially increases from the bolting stage onward.

Meanwhile, although the overall trends were similar, there were notable differences among the varieties in the magnitude of their response to nitrogen application. The ZDJH variety had relatively lower chlorophyll content during the seedling stage, whereas ZD649 and ZD635 were relatively higher. At the bolting stage, ZD619 showed the most drastic increase in chlorophyll. While ZD649 and ZD635 also showed significant increases, the relative gap was smaller. In contrast, QY908 and ZDFS exhibited a more linear response, with a continuous increase in chlorophyll from the N2 to N5 treatments, indicating a relatively steady rate of nitrogen uptake and utilization in their later growth stages. Overall, the dynamic changes in chlorophyll content among the different varieties under varying nitrogen treatments were significant.

3.2. Analysis of Fraction Vegetation Coverage

Figure 5 displays the temperature variation curve for the Hangzhou region during the experimental period. In the initial stage following transplanting, a large diurnal temperature range was observed. The daily maximum temperatures were mostly between 10 and 20 °C, whereas the nightly minimum temperatures often dropped below 0 °C, with extreme lows reaching approximately −5 °C.

This period coincided with the seedling sampling stage, during which the low-temperature environment restricted the growth and development of the seedlings, keeping both the number and size of their leaves at a low level. Consequently, the overall Fraction of Vegetation Coverage (FVC) during this stage was low. As shown in Figure 6a, the FVC during the seedling stage fluctuated around 15% under all nitrogen treatments, and the differences among these treatments were not significant. This suggests that the prevailing light and temperature conditions were insufficient to support large-scale leaf expansion, and therefore the promotional effect of nitrogen on canopy expansion had not been fully realized. This also explains why the differences in chlorophyll content among the various treatments were not significant during the seedling stage.

As the season progressed into the bolting stage, the temperature rose significantly, and the diurnal temperature range narrowed. The daily maximum temperature was mostly between 15 and 25 °C, while the nightly minimum temperature stabilized between 5 and 10 °C. This relatively warm environment with sufficient sunlight provided favorable conditions for photosynthesis in the rapeseed plants, accelerating the synthesis of chlorophyll in the leaves and promoting the gradual closure of the canopy. Figure 6b shows that while the median FVC for the no-nitrogen (N0) plots was approximately 15%, the FVC for all nitrogen-treated plots increased substantially. By this stage, nitrogen had sufficiently accumulated within the plants, and leaf number and leaf area index were initially established. However, the differences among the nitrogen levels were particularly significant. The medium-to-high nitrogen groups exhibited a wider interquartile range and a higher concentration of data points, indicating that an adequate nitrogen supply significantly promoted the rapid expansion and coverage of the canopy. As the temperature further increased to a daily maximum of 25–30 °C and a minimum of around 10 °C, the rapeseed plants entered their peak physiological stage. Leaf function tended to stabilize, and the Leaf Area Index increased significantly. Figure 6c shows that the FVC in the no-nitrogen (N0) plots changed little. The low-nitrogen group (N1) showed some improvement in coverage compared to the bolting stage, hovering between 20% and 45%. The coverage of the medium-nitrogen groups (N2 and N3) jumped to 55–70%, while the high- and very-high-nitrogen groups (N4 and N5) were in the 70–90% range.

The vegetation coverage of the rapeseed exhibited a dynamic evolution across the seedling, bolting, and initial flowering stages, progressing from low to high and from a slow to a fast rate of increase. While many current studies on retrieving crop biophysical parameters are based on mixed-pixel analysis, this experiment successfully created significant differences in FVC among the plots through nitrogen treatments. This provides an opportunity to apply image segmentation techniques to remove the non-vegetation background and extract pure vegetation information. Consequently, this allows for a comparative evaluation of how background removal impacts the accuracy of chlorophyll estimation.

3.3. Analysis of Nitrogen Content and Spectral Reflectance

To investigate the relationship between spectral reflectance, nitrogen content, and chlorophyll content in rapeseed leaves, the leaf nitrogen contents of different rapeseed cultivars under varying nitrogen fertilizer treatments were measured at three key growth stages (Figure 7).

Across all growth stages and cultivars, leaf nitrogen content generally increased with higher nitrogen fertilizer application, demonstrating a clear positive correlation. At the seedling stage (SS), the response with significant differences was primarily observed between no nitrogen (N0) and the highest nitrogen treatments (N4 and N5). During the bolting stage (BS) and the initial flowering stage (IFS), the differences in nitrogen content became more pronounced, particularly between treatments with low nitrogen supply (N0, N1) and those with moderate to high nitrogen supply (N3, N4, and N5). Notably, at higher nitrogen application rates (N4 and N5), leaf nitrogen content tended to plateau, indicating potential nitrogen saturation at these levels. These clear differences in leaf nitrogen content under distinct fertilizer treatments provide the biochemical basis for variations observed in spectral reflectance and chlorophyll content across the canopy, and the effectiveness of indices derived from UAV imagery for chlorophyll monitoring can be further validated and understood.

Figure 8 illustrates the trends in the average reflectance of the four multispectral bands for all pixels (AllPix), rapeseed pixels (GreenPix), and background pixels (BgPix) within the experimental plots. The data are shown for the three growth stages: seedling (Figure 8a), bolting (Figure 8b), and initial flowering (Figure 8c), across six different nitrogen treatment levels.

Across all growth stages, the reflectance of GreenPix was consistently higher than that of AllPix in all bands, with the most significant differences observed in the RedEdge and NIR bands. Overall, the GreenPix data exhibited the lowest reflectance in the Red band, followed by a rapid increase through the RedEdge and NIR bands, reflecting the high scattering of near-infrared light by the internal structure of living leaves. In the AllPix data, when vegetation coverage was low (e.g., during the seedling stage under low-nitrogen treatments), the reflectance across the bands more closely resembled that of BgPix. As the rapeseed plants grew, particularly with increasing nitrogen levels, the AllPix spectrum gradually converged towards the GreenPix spectrum, and the gap between them in the Red, RedEdge, and NIR bands narrowed significantly.

When integrating these reflectance results with the measured nitrogen (Figure 7) and chlorophyll content (Figure 4), clear physiological insights can be drawn. At the seedling stage, where nitrogen treatments had a relatively moderate impact on nitrogen and chlorophyll content, differences in spectral reflectance among nitrogen treatments were also relatively minor. As the plants entered the bolting and initial flowering stages, nitrogen treatments exerted more pronounced effects on leaf nitrogen accumulation and chlorophyll synthesis, correspondingly amplifying spectral differences between treatments, especially in the RedEdge and NIR bands. Specifically, higher nitrogen availability increased both leaf nitrogen and chlorophyll content, resulting in decreased reflectance in visible bands, due to greater chlorophyll absorption, and increased reflectance in RedEdge and NIR bands, owing to enhanced internal leaf scattering from improved leaf structure and density. Under excessive nitrogen conditions (N4, N5), while nitrogen content continued to increase slightly or plateaued, chlorophyll content also approached saturation, causing the differences in reflectance to stabilize across higher nitrogen treatments.

3.4. Correlation Analysis Between Feature Parameters and Chl Content

Pearson correlation analysis was conducted to evaluate the relationship between chlorophyll content and the various color and spectral indices derived from both GreenPix and AllPix data across the three key growth stages of rapeseed. As shown in Figure 9, the color indices derived from GreenPix exhibited a strong correlation with chlorophyll content. One reason for this is that the background interference was effectively removed, allowing the indices to more accurately reflect the pigment information of the plants themselves. Among them, the Hue index (H) showed the highest correlation (r = 0.87, p < 0.01), followed by the GRRI (r = −0.85, p < 0.01) and VARI (r = 0.85, p < 0.01) indices. Additionally, the correlations for the MGRVI and VIGreen indices were also high (both r = 0.84, p < 0.01).

The correlation analysis for the AllPix-derived color indices (Figure 10) indicated that, compared to GreenPix, the correlations with chlorophyll content were relatively lower, though they remained highly significant. The G/R index had the highest correlation (r = 0.74, p < 0.01), followed by the VI_Green and MGRVI indices (both r = 0.73, p < 0.01). The correlations for the GRRI and VARI indices were comparatively lower (r = −0.73 and r = 0.70, respectively), but they also reached a highly significant level (p < 0.01). This result suggests that the background information within the AllPix data may have masked or weakened the sensitivity of the color indices to changes in the plants’ chlorophyll content.

The vegetation indices derived from the GreenPix data generally showed high correlations with chlorophyll content (Figure 11). Among them, the NIR band and the DVI index had the highest correlation with chlorophyll content (r = 0.90, p < 0.01), followed by TVI (r = 0.89, p < 0.01). Furthermore, the RedEdge, MSAVI, and SAVI indices also exhibited strong correlations. The CI_green (r = 0.87, p < 0.01) and GNDVI (r = 0.87, p < 0.01) indices were also highly correlated, indicating the suitability of combining the green and near-infrared bands for chlorophyll retrieval.

The spectral features extracted from the AllPix data also showed high overall correlations with chlorophyll content (Figure 12). The MCARI index had the highest correlation (r = 0.90, p < 0.01), followed by DVI and CI_green (r = 0.89, p < 0.01), which suggests that these indices maintained strong sensitivity even with the inclusion of background pixels. However, compared to the GreenPix results, the overall correlations for the AllPix indices were somewhat lower. This was particularly evident for the NIR band, where the correlation coefficient for AllPix was 0.85, noticeably lower than the 0.90 for GreenPix. This suggests that the presence of background information may have reduced the sensitivity of the NIR band to chlorophyll content to some extent. Additionally, the correlations for some indices, such as CI_re and NDRE, were slightly lower in the AllPix data than in the GreenPix data, indicating that interference from background noise weakened the performance of these indices.

3.5. Prediction of Chl Content Based on Color Indices

Prediction models for rapeseed leaf chlorophyll content were developed separately using the color index datasets derived from GreenPix and AllPix. For each dataset, the 29 color features were first ranked in descending order based on their correlation coefficients with chlorophyll content. Subsequently, subsets of the top 5, top 10, top 15, top 20, and top 25 features, as well as the complete set of all features, were used as inputs to build prediction models. The performance of PLSR, SVR, and MLR models is summarized in

Table 3. Prediction results of chlorophyll content in rapeseed leaves based on color characteristics.

Number of Features	Source Dataset	Model	Calibration Set		Prediction Set
			R2c	RMSEc	R2p	RMSEp
5	GreenPix	PLSR	0.764	72.782	0.747	75.314
		SVR	0.764	72.765	0.760	73.458
		MLR	0.764	72.782	0.747	75.314
	AllPix	PLSR	0.620	92.455	0.595	95.385
		SVR	0.644	89.510	0.626	91.637
		MLR	0.620	92.455	0.595	95.385
10	GreenPix	PLSR	0.775	71.049	0.734	77.305
		SVR	0.781	70.173	0.765	72.556
		MLR	0.777	70.803	0.739	76.563
	AllPix	PLSR	0.662	87.127	0.631	91.047
		SVR	0.715	79.992	0.673	85.605
		MLR	0.680	84.791	0.651	88.543
15	GreenPix	PLSR	0.775	71.068	0.734	77.229
		SVR	0.781	70.178	0.761	73.232
		MLR	0.783	69.886	0.742	76.154
	AllPix	PLSR	0.664	86.918	0.635	90.543
		SVR	0.713	80.349	0.679	84.813
		MLR	0.690	83.464	0.664	86.826
20	GreenPix	PLSR	0.776	70.964	0.736	76.982
		SVR	0.783	69.760	0.761	73.278
		MLR	0.784	69.623	0.747	75.365
	AllPix	PLSR	0.662	87.099	0.631	91.019
		SVR	0.713	80.279	0.681	84.552
		MLR	0.697	82.577	0.658	87.667
25	GreenPix	PLSR	0.776	70.929	0.736	76.993
		SVR	0.783	69.891	0.761	73.246
		MLR	0.787	69.253	0.742	76.138
	AllPix	PLSR	0.663	87.089	0.631	90.988
		SVR	0.712	80.389	0.683	84.334
		MLR	0.701	81.998	0.541	101.514
All	GreenPix	PLSR	0.776	71.017	0.739	76.502
		SVR	0.783	69.898	0.763	72.991
		MLR	0.788	68.960	0.743	76.016
	AllPix	PLSR	0.662	87.208	0.630	91.065
		SVR	0.711	80.604	0.683	84.383
		MLR	0.704	81.526	0.591	95.861

Note: RMSE_c and RMSE_p (μg/g).

.

The results indicated that the prediction accuracy of models based on the GreenPix dataset was significantly higher than those based on the AllPix dataset. For instance, when using 5 features, the SVR model with the GreenPix dataset achieved a prediction accuracy (R²_p = 0.760, RMSE_p = 73.458 μg/g) that was markedly superior to the model with the AllPix dataset (R²_p = 0.626, RMSE_p = 91.637 μg/g). The AllPix data, which includes information from the soil background, likely introduces significant noise that reduces the sensitivity between the features and chlorophyll content. In contrast, the GreenPix data more purely reflects the relationship between rapeseed canopy color and its chlorophyll content. This demonstrates that removing background pixels can effectively improve the stability and accuracy of chlorophyll content prediction.

As the number of color index features increased, the prediction accuracy showed a trend of first increasing and then decreasing. The peak accuracy was achieved when the number of features was 10. At this point, the SVR model based on the GreenPix dataset performed optimally, with a prediction set R² of 0.765 and an RMSE_p that decreased to 72.556 μg/g. Although the prediction accuracy for the AllPix dataset also improved with 10 features, it remained significantly lower than that of the GreenPix model. When the number of features further increased, the prediction accuracy showed a slight decline. This phenomenon may be attributed to the introduction of weakly correlated features as the feature count surpassed a certain threshold, which increased model complexity and reduced its generalization performance.

A comparison of the three prediction models revealed that the SVR model consistently performed the best across all datasets. Its strong non-linear mapping capability provides a significant advantage in capturing the complex, non-linear relationship between the plant’s color features and its chlorophyll content. In summary, for predicting chlorophyll content based on color indices, the GreenPix dataset demonstrated superior predictive ability. Specifically, the best prediction result was achieved by using an SVR model with the top 10 most correlated features from the GreenPix dataset.

3.6. Prediction of Chl Content Based on Spectral Indices

For both the GreenPix and AllPix datasets, prediction models for rapeseed leaf chlorophyll content were established using PLSR, SVR, and MLR. The 18 extracted spectral features were first ranked in descending order based on their correlation coefficients with chlorophyll content. Subsequently, subsets of the top 3, top 6, top 9, top 12, and top 15 features, as well as the complete set of all features, were used as model inputs. The results are presented in

Table 4. Prediction results of chlorophyll content in rapeseed leaves based on spectral characteristics.

Number of Features	Source Dataset	Model	Calibration Set		Prediction Set
			R2c	RMSEc	R2p	RMSEp
3	GreenPix	PLSR	0.840	60.566	0.838	60.826
		SVR	0.851	58.121	0.832	62.113
		MLR	0.840	60.566	0.838	60.826
	AllPix	PLSR	0.794	70.642	0.795	70.098
		SVR	0.809	67.345	0.797	69.473
		MLR	0.794	70.642	0.795	70.098
6	GreenPix	PLSR	0.861	55.768	0.851	58.014
		SVR	0.862	55.605	0.845	59.264
		MLR	0.862	55.760	0.851	57.934
	AllPix	PLSR	0.819	65.191	0.820	64.617
		SVR	0.811	66.865	0.800	68.983
		MLR	0.819	65.191	0.820	64.617
9	GreenPix	PLSR	0.854	57.407	0.846	58.967
		SVR	0.870	53.929	0.858	56.374
		MLR	0.862	55.551	0.852	57.646
	AllPix	PLSR	0.820	64.954	0.825	63.566
		SVR	0.830	62.854	0.826	63.394
		MLR	0.825	63.876	0.824	63.672
12	GreenPix	PLSR	0.862	55.577	0.855	57.061
		SVR	0.873	53.275	0.862	55.534
		MLR	0.866	54.677	0.862	55.462
	AllPix	PLSR	0.825	63.839	0.828	62.829
		SVR	0.831	62.427	0.827	63.174
		MLR	0.828	63.191	0.829	62.738
15	GreenPix	PLSR	0.862	55.695	0.855	57.057
		SVR	0.868	53.953	0.866	54.622
		MLR	0.870	53.882	0.866	54.713
	AllPix	PLSR	0.825	63.807	0.827	63.094
		SVR	0.830	62.840	0.828	62.846
		MLR	0.829	62.998	0.830	62.511
All	GreenPix	PLSR	0.864	55.142	0.857	56.706
		SVR	0.877	52.382	0.866	54.597
		MLR	0.870	53.833	0.864	54.787
	AllPix	PLSR	0.824	64.018	0.820	63.593
		SVR	0.831	62.524	0.835	63.449
		MLR	0.829	62.898	0.821	62.216

Note: RMSE_c and RMSE_p (μg/g).

.

The prediction results show that the models based on the GreenPix dataset had a significantly better overall prediction accuracy than those based on the AllPix dataset, a trend similar to that observed with the color features. This further confirms that the effective removal of background pixels can significantly reduce the interference of background noise on the plant’s spectral features. This, in turn, allows for a more precise reflection of the true spectral information of the plant canopy, thereby improving the accuracy of chlorophyll prediction. Among the three models, the SVR model performed best, achieving a higher coefficient of determination and lower error on the prediction set. As the number of features increased from three to all, the model prediction accuracy showed a trend of progressive improvement. Specifically, when the feature count was increased from 3 to 6, the R²_p of the PLSR model with GreenPix data improved from 0.838 to 0.851, while the RMSE_p decreased from 60.826 to 58.014 μg/g, a clear enhancement. Interestingly, when the feature numbers were 3 and 6, the PLSR and MLR models yielded identical accuracies, which may be because the few selected features had a high linear correlation, causing both models to function as equivalent linear regressions. As more features were added (12, 15, and all), the accuracy continued to improve, though the magnitude of improvement gradually diminished and began to stabilize. The optimal prediction result was ultimately achieved using the SVR model with all spectral features from the GreenPix dataset (R²_p = 0.866, RMSE_p = 54.597).

3.7. Prediction of Chl Content Based on the Fusion of Color and Spectral Indices

To further enhance the accuracy of chlorophyll content prediction, the top 10 color features most significantly correlated with chlorophyll content were progressively fused with the baseline set of 18 spectral features to evaluate the resulting improvement in prediction performance. The prediction results for the AllPix data (

Table 5. The prediction results of chlorophyll content in rapeseed leaves based on the gradual fusion of spectral features and color features in the AllPix dataset.

Number of Add Color Features	Number of All Features	Model	Calibration Set		Prediction Set
			R2c	RMSEc	R2p	RMSEp
0	18	PLSR	0.824	64.018	0.825	63.593
		SVR	0.831	62.524	0.835	63.449
		MLR	0.829	62.898	0.821	62.216
1	19	PLSR	0.832	62.384	0.826	63.429
		SVR	0.843	59.855	0.836	61.207
		MLR	0.838	61.021	0.830	62.553
2	20	PLSR	0.832	62.287	0.827	63.040
		SVR	0.844	59.724	0.835	61.361
		MLR	0.839	60.801	0.832	62.032
3	21	PLSR	0.833	62.067	0.829	62.696
		SVR	0.844	59.668	0.834	61.502
		MLR	0.839	60.774	0.831	62.250
4	22	PLSR	0.834	61.925	0.829	62.660
		SVR	0.844	59.630	0.834	61.688
		MLR	0.839	60.774	0.831	62.246
5	23	PLSR	0.834	61.842	0.829	62.657
		SVR	0.844	59.634	0.832	61.934
		MLR	0.839	60.694	0.832	62.111
6	24	PLSR	0.834	61.813	0.829	62.690
		SVR	0.844	59.664	0.832	62.067
		MLR	0.839	60.688	0.832	61.951
7	25	PLSR	0.842	60.132	0.838	60.748
		SVR	0.847	58.969	0.837	61.020
		MLR	0.854	57.447	0.847	58.893
8	26	PLSR	0.844	59.574	0.844	59.429
		SVR	0.855	57.214	0.845	59.179
		MLR	0.854	57.424	0.847	58.869
9	27	PLSR	0.846	59.111	0.848	58.514
		SVR	0.856	56.888	0.846	59.057
		MLR	0.855	57.312	0.849	58.376
10	28	PLSR	0.846	59.210	0.848	58.631
		SVR	0.857	56.856	0.845	59.177
		MLR	0.855	57.104	0.850	58.096

Note: RMSE_c and RMSE_p (μg/g).

) show that as color features gradually fused with the spectral features, the overall prediction accuracy of the models exhibited a distinct upward trend. Initially, without any fused color features, the SVR model yielded an R²_p of 0.835 and an RMSE_p of 63.449 μg/g. As color features were incrementally added, model accuracy improved significantly, especially with the addition of the seventh color feature. At this point, the MLR model’s prediction accuracy rose to 0.847, and its RMSE_p fell to 58.893 μg/g, marking the most substantial improvement. Accuracy continued to increase slightly with the addition of more features, reaching its maximum when all 10 color features were fused, where the MLR model achieved an R²_p of 0.850 and an RMSE_p of 58.096 μg/g.

This result indicates that the appropriate fusion of color features can significantly improve the accuracy of chlorophyll content prediction from AllPix data, with the most pronounced improvement observed after adding seven color features. Furthermore, the MLR model performed optimally after feature fusion, surpassing both the PLSR and SVR models. This suggests that linear models may hold an advantage in a fused spectral and color feature context, possibly because the linear relationships within the fused feature space become more prominent.

Table 6. The prediction results of chlorophyll content in rapeseed leaves based on spectral features and gradually fused with color features in the GreenPix dataset.

Number of Add Color Features	Number of All Features	Model	Calibration Set		Prediction Set
			R2c	RMSEc	R2p	RMSEp
0	18	PLSR	0.864	55.142	0.857	56.706
		SVR	0.877	52.382	0.866	54.597
		MLR	0.870	53.833	0.864	54.787
1	19	PLSR	0.872	53.437	0.860	55.916
		SVR	0.878	52.178	0.870	53.771
		MLR	0.879	51.923	0.871	53.649
2	20	PLSR	0.872	53.353	0.860	56.018
		SVR	0.877	52.236	0.869	54.024
		MLR	0.880	51.629	0.867	54.580
3	21	PLSR	0.873	53.303	0.860	56.083
		SVR	0.877	52.383	0.869	54.160
		MLR	0.880	51.553	0.869	54.151
4	22	PLSR	0.873	53.246	0.859	56.188
		SVR	0.876	52.473	0.868	54.287
		MLR	0.881	51.539	0.869	54.093
5	23	PLSR	0.874	52.957	0.859	56.191
		SVR	0.879	51.934	0.871	53.695
		MLR	0.883	51.107	0.870	53.775
6	24	PLSR	0.874	52.992	0.859	56.167
		SVR	0.883	50.961	0.875	52.707
		MLR	0.885	50.626	0.872	53.432
7	25	PLSR	0.871	53.644	0.861	55.753
		SVR	0.884	50.685	0.878	52.187
		MLR	0.885	50.529	0.872	53.304
8	26	PLSR	0.870	53.890	0.851	57.859
		SVR	0.886	50.375	0.878	52.070
		MLR	0.886	50.270	0.872	53.370
9	27	PLSR	0.871	53.749	0.855	57.036
		SVR	0.886	50.313	0.878	52.160
		MLR	0.887	50.213	0.872	53.301
10	28	PLSR	0.869	54.155	0.855	57.149
		SVR	0.886	50.294	0.878	52.150
		MLR	0.887	50.206	0.872	53.298

Note: RMSE_c and RMSE_p (μg/g).

presents the prediction results for rapeseed leaf chlorophyll content using the GreenPix dataset, where color features were progressively fused with spectral indices. Compared to the AllPix, the GreenPix data exhibited higher prediction accuracy even before feature fusion (SVR model: R²_p = 0.866, RMSE_p = 54.597 μg/g). As color features were gradually fused, the prediction accuracy for the GreenPix data continued to improve notably. The improvement was particularly significant with the fusion of the 6th and 7th color features, and the SVR model reached its peak accuracy when the 7th color feature was fused (R²_p = 0.878, RMSE_p = 52.187 μg/g). After this point, the accuracy tended to stabilize or show slight fluctuations. The MLR model also performed exceptionally well after fusion, reaching an R²_p of 0.872 and an RMSE_p of 53.432 μg/g with the addition of six features, after which its performance remained stable. In contrast, the performance of the PLSR model was slightly weaker, with relatively limited accuracy gains from feature fusion.

To visually evaluate the predictive capability and generalization of the optimal SVR model developed from fused color and spectral features based on GreenPix data, Figure 13 presents RGB images and corresponding chlorophyll predictions for 100 randomly selected samples from the independent prediction set. The predicted chlorophyll content heatmap demonstrates good consistency with the visual assessment of plant status from the RGB images. Samples exhibiting dense vegetation and darker green foliage visually correspond to higher predicted chlorophyll content (green color), whereas plots with sparser canopy coverage, visible stress, or nitrogen deficiency correspond to lower predicted chlorophyll predictions (yellow color). These results confirm the robust predictive capability of the developed model across various canopy conditions and nitrogen treatments.

Overall, fusing color features with spectral features can significantly enhance the prediction accuracy of chlorophyll content. This improvement is especially pronounced for the AllPix data, which is heavily affected by background interference. The number of fused color features is not necessarily “the more, the better,” as for the GreenPix dataset, fusing the top seven color features was sufficient to achieve the optimal prediction result. Finally, compared to the data from all pixels, the background-removed GreenPix dataset consistently demonstrated higher prediction accuracy, confirming that the removal of background pixels is an effective step for improving the accuracy of chlorophyll prediction.

4. Discussion

The dynamic changes in chlorophyll content were strongly influenced by nitrogen availability and phenological stage in rapeseed. As a key component of chlorophyll [46], nitrogen application directly boosted pigment content, though a saturation effect was observed at higher application rates (N4 and N5), indicating an optimal level for maximizing nitrogen use efficiency without diminishing returns [47]. The impact of nitrogen was also modulated by environmental conditions; low temperatures at the seedling stage limited overall growth, making temperature the primary constraint. As temperatures rose during the bolting and initial flowering stages, the demand of rapeseed for nitrogen increased, and the differential effects of the nitrogen treatments on both FVC and chlorophyll became highly significant. This underscores that the efficacy of nutrient management is intrinsically linked to the crop’s developmental stage and environment.

Also, we confirmed that background removal significantly enhances chlorophyll estimation accuracy from UAV imagery. The marked improvement in model performance when using pure vegetation pixels (GreenPix) compared to mixed pixels (AllPix) highlights the pervasive issue of background interference, commonly known as the mixed-pixel problem in remote sensing. In high-resolution UAV imagery, especially during early growth stages or under nutrient-stressed conditions where the canopy is sparse, as reflected by low FVC plots in Figure 6, individual pixels frequently capture combined signals from vegetation and non-vegetation elements such as soil, shadows, and residues [48]. These background components introduce substantial noise and bias into reflectance data, compromising the relationship between the spectral signal and biophysical parameters like chlorophyll content. Zhao et al. [49] demonstrated that the combination of Excess Green (ExG) and Excess Red (ExR) indices effectively isolates vegetation from complex backgrounds, significantly improving segmentation accuracy and thus supporting accurate extraction of plant features. Our results quantitatively illustrate this effect: the removal of background improved the SVR model’s prediction accuracy R²_p from 0.683 to 0.763 for color indices and from 0.835 to 0.866 for spectral indices (Table 3 and Table 4). This confirms the effectiveness of the ExG-ExR thresholding method for isolating pure vegetation signals and enhancing model sensitivity.

The advantage of multispectral indices over RGB-based color indices for chlorophyll estimation is consistent with findings reported by Maimaitijiang et al. [50] for soybean nitrogen and chlorophyll-a content. In the single-feature dataset, the best spectral index model is SVR on GreenPix data, with an R²_p of 0.866, significantly outperforming the best color index model (R²_p = 0.765). This performance gap aligns with the biophysical principles underlying plant spectroscopy. Multispectral sensors capture specific, narrow spectral bands, including near-infrared and red-edge regions, known for their sensitivity to chlorophyll content [51]. Particularly, the red-edge region captures subtle shifts linked directly to chlorophyll variations, marking the transition from strong chlorophyll absorption in the red to high internal leaf scattering in the near-infrared band [52]. Consequently, multispectral vegetation indices leveraging red-edge information, like NDRE and CI_re, typically exhibit strong, robust correlations with chlorophyll concentration and overall plant health [53]. In contrast, although RGB images offer higher spatial resolution, their broad spectral bands are more susceptible to interference from illumination variability, canopy structure, and signal saturation at elevated chlorophyll levels, thereby constraining their predictive accuracy. However, RGB indices have shown robustness under certain scenarios, such as single vegetation indices in estimating maize grain yield [54]. Notably, previous research combining RGB and multispectral indices achieved promising results in rice yield estimation [55], highlighting the potential value of exploring combined data sources to further enhance chlorophyll and nutrient status predictions in rapeseed. Among the regression algorithms tested, SVR consistently outperformed PLSR and MLR when using individual feature sets, no matter whether color or spectral. The relationship between canopy spectral reflectance and its biochemical constituents is inherently complex and non-linear, influenced by the hierarchical structure of the canopy, pigment packaging within cells, and bidirectional reflectance effects [56]. Linear models like PLSR and MLR, which assume a direct linear relationship, are often insufficient to capture this complexity. SVR utilizes a kernel function to project the input data into a higher-dimensional feature space. In this transformed space, a linear relationship can be established, enabling the model to effectively capture the underlying non-linear patterns in the original data. This flexibility is the primary reason for SVRs superior performance in this context.

Fusing RGB-derived color indices with multispectral indices further enhances chlorophyll prediction accuracy. This indicates that the two data sources provide complementary, rather than merely redundant, information. For the background-removed GreenPix data, fusing the top seven color features with the full set of spectral indices increased the SVR model’s R²_p from 0.866 to a study-best of 0.878, while decreasing the RMSEp from 54.597 to 52.187 μg/g (Table 6). This synergy likely arises because the sensors capture different but complementary aspects of canopy status. The multispectral sensor provides precise, narrow-band data directly linked to chlorophyll’s absorption features, while the RGB sensor captures broad-band color and textural information that relates to overall plant vigor, senescence, and subtle color shifts not fully characterized by the limited number of multispectral bands. This finding aligns with a growing consensus that multi-sensor data fusion is a powerful strategy for improving the estimation of crop parameters. Studies have shown that fusing spectral data with textural, structural, or thermal data consistently leads to more robust models for estimating yield [57,58,59], LAI [60], and phenotype analysis [61,62]. An important finding was that the optimal fusion strategy involved selecting a subset of the most informative color features rather than all of them. For the GreenPix data, performance peaked with the addition of seven color features and then plateaued (Table 6). This illustrates the principle of model parsimony; adding features with a low signal-to-noise ratio can introduce more error than useful information, potentially leading to overfitting and reduced generalization on new data.

5. Conclusions

This study developed and validated a robust framework for the rapid, non-destructive estimation of canopy chlorophyll content in rapeseed, leading to the following conclusions: (1) The removal of background pixels using the ExG-ExR thresholding method is a critical step that significantly improves the accuracy of predictive models. Compared to models using the original mixed-pixel data (AllPix), those based on the purified vegetation dataset (GreenPix) were consistently superior. For instance, when modeling with spectral features alone, the prediction accuracy R²_p of the SVR model increased from 0.835 (AllPix) to 0.866 (GreenPix). (2) The fusion of color and spectral features further enhances prediction accuracy beyond using either feature type alone. The Support Vector Regression (SVR) model demonstrated the best performance across various datasets and feature combinations, proving effective at capturing the complex, non-linear relationships between remote sensing data and chlorophyll content. (3) The optimal approach identified in this study is the combination of the background-removed GreenPix dataset with fused color and spectral features, modeled using SVR. This method achieved the highest prediction accuracy with an R²_p of 0.878 and an RMSE_p of 52.187 μg/g. This integrated strategy provides a reliable and effective tool to support precision nitrogen management in rapeseed cultivation.