Imaging Hyperspectral Feature Fusion for Estimation of Rapeseed Pod’s Water Content and Recognition of Pod’s Maturity Level

Abstract

Imaging hyperspectral technology is becoming popular in agriculture to provide detailed information on crop growth. In this work, we propose an estimation of rapeseed pod’s water content model and identification of maturity levels (green, yellow, and full) model by using this technology. Four types of hyperspectral features are extracted—color, texture, spectral three-edge parameters, and spectral indices. By integrating these features, satisfactory results are achieved: the optimal feature combination is from spectral indices and three-edge parameters, with low RRMSE and RE for yellow maturity. Incorporating spectral indices significantly improved the pod’s water content estimation, reducing RRMSE by up to 43.30% and 30.11% in the green and full maturity stages. Random forest and support vector machine with kernel method (SVM-KM) algorithms outperformed other statistical models, with SVM-KM achieving up to 96.90% accuracy in identifying maturity levels. These findings provide valuable insights for managing rapeseed production during the pod stage.

1. Introduction

As a pivotal oilseed crop, rapeseed (Brassica napus L.) has garnered substantial research interest in agricultural sciences, particularly pertaining to growth regulation and yield/quality enhancement. The pod-filling stage is a critical determinant of seed production [1], with approximately two-thirds of the dry matter accumulation attributed to photosynthesis in the pod [2]. Notably, during the green-ripe and yellow-ripe phases of pod development, photosynthetic efficiency exhibits a strong correlation with the water status of the pod. Consequently, precise monitoring of the pod’s water content during these stages is imperative. However, as pods mature, pod shatter emerges as a prevalent issue, adversely impacting harvest efficiency and yield. Inadequate harvest timing management can result in up to 20% yield losses [3]. Since the pod’s water content and maturity level are key factors influencing pod shatter susceptibility [4,5,6,7,8], they provide valuable insights for optimizing harvest timing and mitigating losses. In summary, the capacity for rapid, non-destructive detection of rapeseed pod’s water content and accurate identification of maturity stages is crucial for maximizing yield and quality in rapeseed production systems.

Traditional desiccation-based techniques for quantifying rapeseed pod pericarp water content [9,10,11] are not only time-consuming and labor-intensive but also destructive, precluding their effective integration with remote sensing technologies for rapid and accurate assessment across large-scale cultivation areas. Conventionally, the determination of optimal harvest timing has heavily relied on empirical observations, such as the visual cue of field-wide pod yellowing resembling loquat fruits or pericarp water contents ranging between 11 and 13% [12]. However, these subjective approaches lack consistency, are challenging to master, and are antithetical to the principles of smart agriculture. Consequently, the development of rapid, non-destructive methodologies for estimating rapeseed pod pericarp water status and maturity is a critical imperative for optimizing yield and quality in rapeseed production systems.

Imaging hyperspectral technology offers a viable solution by synergistically combining imaging and spectroscopic modalities, facilitating the simultaneous acquisition of spatial and biochemical information from target specimens. This integrative approach has emerged as an invaluable diagnostic tool for plant phenotyping and growth monitoring. In recent years, numerous studies have leveraged hyperspectral imaging to assess plant physiological status through spectral information analysis [13,14,15,16,17,18], with a burgeoning focus on water content determination [19,20,21,22,23,24], such as Sun et al. [22] who developed a hyperspectral method for accurate, rapid lettuce leaf water content detection during growth. Applying the CARS-ABC-SVR model optimized using an artificial bee colony algorithm, the R² is found to be 0.9214 and the RMSE is 2.95% on the prediction set.

Spectral data are information-rich, but the strong correlations and redundancies among adjacent frequency bands may impact efficiency and accuracy [25,26,27]. Image data facilitate target recognition and estimation through color and texture features, providing multi-angular information. To obtain comprehensive and refined target information from raw imaging hyperspectral data, feature extraction and inter-class feature fusion are required. This approach overcomes single-data modality limitations, fully leverages the complementarity of multi-angular and multi-dimensional information, and enhances decision-making accuracy and reliability. Peñuelas et al. [28] constructed the Water Band Index (WBI) using the 970 nm water-sensitive waveband and the 900 nm reflectance, verifying WBI’s effectiveness in predicting leaf water stress. Ramoelo et al. [29] demonstrated the practicality of field spectroscopy for estimating nitrogen-to-phosphorus ratios in savanna grasses. Zhang et al. [30] extracted feature parameters from rapeseed canopy spectra, multispectral images, and canopy temperature, constructing an information fusion estimation model for rapeseed water content, exhibiting significantly higher accuracy than single-modality models.

In this work, we apply imaging hyperspectral features to investigate water content estimation and maturity level identification, which are of importance to rapeseed pod production management. Four types of features including color, texture, spectral three-edge parameter, and spectral index are extracted from hyperspectral data. These features are then combined intra-class and fused inter-class. Multiple machine learning models are employed to model for the purpose mentioned above.

2. Materials and Methods

2.1. Experimental Materials

For this experiment, we selected the Fengyou 520 rapeseed cultivar bred by the Hunan Crop Research Institute. The plants were planted in suitable and uniform experimental fields on 2 October 2022. With an average plant height of 167.6 cm, this cultivar belongs to the mesophytic branching type. We collected samples during three critical periods of pod development: Pod Stage I or Green Ripe Stage (where pods are initially formed and appear full, 18 April 2023, denoted as P1), Pod Stage II or Yellow Ripe Stage (where pods start to turn yellow in small areas, 28 April 2023, denoted as P2), and Pod Stage III or Full Ripe Stage (where most of the pods are yellow and maturing, starting to dry gradually, 23 May 2023, denoted as P3). To avoid weather influences like rainfall or morning and evening dew, we collected samples in the afternoon of the third consecutive sunny day when the field temperature was between 20 °C and 25 °C. Each time, we randomly selected 8 rapeseed plants from the experimental area and picked 9 pods from each plant: 1 pod each from the top 0–20 cm, 20–40 cm, and 40–60 cm sections, and 1 pod each based on color shade (greenish, yellow-green, and yellowish). Immediately after sampling, the pods were taken to the laboratory for imaging hyperspectral and water content measurement. In this way, 216 pod samples were collected across the three stages.

2.2. Data Acquisition and Processing

2.2.1. Data Acquisition

In this work, a SOC710 hyperspectral imager (wavelength 375–1041 nm, resolution 4.6875 nm, produced by Surface Optics Co., 11555 Rancho Bernardo Road San Diego, California 92127, USA) was used to measure imaging spectral data, as shown in Figure 1. The experiment was conducted in a darkroom using a 50 × 60 cm, 100 cm high mission module coated with diffuse reflective material, surrounded by a 70 W halogen light source at a 45-degree angle, and equipped with a ventilation and cooling system. The imager was placed vertically above the target with an exposure distance of 300 mm. After preheating for 15 min, the imaging spectral information of the rapeseed pods was measured. Data processing was performed using Spectral Radiance Analysis Toolkit and ENVI5.3 software, and 5 rounds of random measurements were taken in the region of interest of the pods, with the average value taken as the final spectral reflectance value of the ROI.

On the other hand, we employed a high-precision TP300D electronic balance and a 101-type electric heating blowing dry box to measure the water content of the pod pericarp. The balance has a range of 300 g, an accuracy of 0.001 g, and multiple functions such as calibration, peeling, and zero tracking. The temperature control range of the drying box is 5–250 °C, the temperature control accuracy reaches ±1 °C, and the temperature uniformity is 3.5%, which is accurate and easy to maintain. After obtaining the imaging hyperspectral data of rapeseed pods, to avoid the influence of oil volatilization in rapeseed on the measurement results, we performed seed removal on the pods before each drying. After a series of operations such as seed removal, weighing the fresh weight of the pod pericarp, drying the pod pericarp to a constant weight, and weighing the dry weight, the water content of the pod pericarp was calculated. We obtained a total of 216 data points for pod pericarp water content.

Table 1. Statistical characteristics of pod’s water content.

Period Name	Mark	No. of Samples	Min(%)	Max (%)	Mean (%)	std (%)	cv
Green maturity	P1	72	55.7281	95.7142	78.1949	9.5343	0.1219
Yellow maturity	P2	72	58.6956	88.9447	81.3432	4.2243	0.0519
Full maturity	P3	72	11.2676	50.6493	22.8161	13.1342	0.5756

Tip: std means standard deviation and cv means coefficient of variation.

lists the statistical characteristics of water content in rapeseed pod pericarp at various stages in detail. As can be seen from the table, there is little difference in the mean water content between the green and yellow maturity stages, while the mean water content during the full maturity stage is significantly reduced, and the coefficient of variation is significantly increased. Through the above process, we obtained a total of 216 spectral reflectance curves and the corresponding RGB images, as well as the pod’s water content.

2.2.2. Feature Extraction

To avoid data redundancy and improve modeling efficiency, four types of features of the imaging hyperspectral data are extracted for our consideration, namely image color, image texture, spectral three-edge (red, blue, yellow) parameters, and spectral index.

• Image color

Focusing on the region of interest after eliminating background interference in the hyperspectral images of rapeseed pods, we obtained 98 color features according to the ten-color method [31,32,33]. Those color features are labeled as “TcC-” prefix.

• Image texture

Texture features reveal changes in the image surface and the organizational characteristics of its structure. Based on the rapeseed pod images, various texture features, such as contrast, energy, correlation, entropy, third-order moments, brightness, consistence, homogeneity, and smoothness are extracted through the gray-level co-occurrence matrix [34], which are labeled as “TcT-” prefix, as listed in

Table 2. The calculation formula of image texture features.

ID	Label	Texture Name	Description or Calculation
1	TcT-Con	Contrast	${\displaystyle {\sum }_i{\displaystyle {\sum }_j{(i-j)}^{2}p(i,j)}}$
2	TcT-Ener	Energy	${\displaystyle {\sum }_i{\displaystyle {\sum }_{j}p{(i,j)}^2}}$
3	TcT-Corr	Correlation	${\displaystyle {\sum }_i{\displaystyle {\sum }_j(i-\mu )(j-\mu )p(i,j)}}/{\sigma }^2$,
4	TcT-Ent	Entropy	$-{\displaystyle {\sum }_i{\displaystyle {\sum }_{j}p(i,j)\log p(i,j)}}$
5	TcT-TOM	Third-order moments	${\left[{\displaystyle {\sum }_i{\displaystyle {\sum }_j{(p(i,j)-\mu )}^3}}/N\right]}^{1/3}$
6	TcT-Smooth	Smoothness	measuring the uniformity of image grayscale
7	TcT-Cons	Consistence	reflecting the autocorrelation of image texture.
8	TcT-Bri	Brightness	reflecting the overall brightness of the image
9	TcT-Homo	Homogeneity	${\displaystyle {\sum }_i{\displaystyle {\sum }_j\frac{p(i,j)}{1+{(i-j)}^2}}}$

Tip: p(i,j) is the frequency of the pixel (i,j) in an image. μ and σ are the mean value and standard deviation of the frequency, respectively. N is the number of image pixels.

.

• Spectral three-edge parameter

The red edge parameters refer to the reflection spectral characteristics of green plants in the red edge (680–760 nm), including red edge position, red edge amplitude, and red edge area. Additionally, the blue edge (490–530 nm) and yellow edge (560–640 nm) parameters, collectively known as spectral three-edge parameters [32,35], are related to the growth status and biomass of plants. We choose 23 three-edge parameters for our consideration, which are identified with the “Tri-” prefix, as listed in

Table 3. The calculation formula of hyperspectral three-edge parameters.

ID	Label	Description or Calculation
1	Tri-Db	The spectral maximum value of the first derivative in the blue edge (490~530 nm);
2	Tri-Bep	The bands corresponding to the maximum value of the first derivative in the blue edge;
3	Tri-SDb	Integration of the first derivative in the range of blue edge;
4	Tri-Dbmin	The minimum value of the first derivative in the blue edge;
5	Tri-NDBE	Normalized blue-edge index: (Db − Dbmin)/(Db + Dbmin);
6	Tri-Dy	The spectral maximum value of the first derivative in the yellow edge (560~640 nm);
7	Tri-Yep	The bands corresponding to the maximum value of the first derivative in the yellow edge;
8	Tri-SDy	Integration of the first derivative spectrum in the range of yellow edge;
9	Tri-Dymin	The minimum value of the first derivative spectrum in the yellow edge;
10	Tri-NDYE	Normalized yellow-edge index: (Dy − Dymin)/(Dy + Dymin);
11	Tri-Dr	The spectral maximum value of the first derivative in the red edge (680~760 nm);
12	Tri-Rep	The bands corresponding to the maximum value of the first derivative in the red edge;
13	Tri-SDr	Integration of the first derivative in the range of red edge;
14	Tri-Drmin	The minimum value of the first derivative in the red edge;
15	Tri-NDRE	Normalized red-edge index: (Dr − Drmin)/(Dr + Drmin);
16	Tri-RB	SDr/SDb;
17	Tri-RY	SDr/SDy;
18	Tri-YB	SDy/SDb;
19	Tri-NRB	SDr − SDb/(SDr + SDb);
20	Tri-NRY	SDr − SDy/(SDr + SDy);
21	Tri-NBY	SDb − SDy/(SDb + SDy);
22	Tri-Kur	Kurtosis of the first derivative curve of the red edge;
23	Tri-Ske	Skewness of the first derivative curve of the red edge.

.

• Spectral index

The crop spectral index is the feature obtained by combining spectral reflectance from sensitive wavelength regions, which is very useful for assessing and monitoring crop growth. Ratio spectral index (RSI), normalized difference spectral index (NDSI), and difference spectral index (DSI), which are most important for green crops [36], are used for our study. In addition, four other key spectral indices related to plant water content, namely plant senescence reflectance index (PSRI) [37], ripening stage of pod maturity index (RSMI) [38], structure-independent pigment index (SIPI) [39], and modified normalized difference red edge index (mNDRE) [37], are also employed. The calculation formulas are shown in Equations (1)–(7), where R_λ1, R_λ2, and R_λ3 represent the reflectance at wavelengths λ₁, λ₂, and λ₃, respectively, which are labeled as “SpI-” prefix. Considering the differences in sensitive bands of spectral indices among different crops and varieties, we perform interactive combinations of bands in the range of 375–1041 nm. According to the correlation analysis with the pod’s water content, the wavelength bands with the highest correlation coefficient are obtained at the three periods, listed in

Table 4. The top three features corresponding to the best correlation coefficient (CC) in P1.

Feature	Formula/Description	CC	Source
SpI-PSRI	(R390 − R567)/R664	0.9425 **	Ref. [37]
SpI-RSMI	1 − (0.64R643 + 0.36R669)/R662	0.9178 **	Ref. [27]
SpI-SIPI	(R490 − R512)/(R490 + R679)	0.9146 **	Ref. [39]
Tri-NDBE	$(D_b-D_{bmin})/\left(D_b+D_{bmin}\right)$	−0.7888 **	Ref. [40]
Tri-NDYE	$(D_y-D_{ymin})/\left(D_y+D_{ymin}\right)$	−0.7756 **	Ref. [40]
Tri-NDRE	$(D_r-D_{rmin})/\left(D_r+D_{rmin}\right)$	−0.7506 **	Ref. [40]
TcC-(2G-R-B)/L*	Colorimetric color F54	0.3051 **	Ref. [31]
TcC-G-R	Colorimetric color F22	0.3013 *	Ref. [31]
TcC-G/Y	Colorimetric color F41	0.2983 *	Ref. [31]
TcT-Homogeneity	Uniformity of pixel	−0.3754 **	Ref. [31]
TcT-Contrast ratio	Difference in gray level	0.3690 **	Ref. [31]
TcT-Correlation	Similarity of grayscale levels in orientation	−0.3454 **	Ref. [31]

Tip: * and ** indicate a 5% and 1% significance level, respectively.

,

Table 5. The top three features corresponding to the best correlation coefficient (CC) in P2.

Feature	Formula/Description	CC	Source
SpI-PSRI	(R505 − R633)/R1041	0.9025 **	Ref. [37]
SpI-mNDRE	(R523 − R695)/(R523 + R695 – 2 * R585)	0.8989 **	Ref. [37]
SpI-SIPI	(R504 − R633)/(R504 + R1019)	0.8966 **	Ref. [39]
Tri-NDBE	$(D_b-D_{bmin})/\left(D_b+D_{bmin}\right)$	0.8249 **	Ref. [40]
Tri-D_rmin	The minimum value of the first derivative of the spectrum in red edges (680–760 nm)	−0.8134 **	Ref. [40]
Tri-Kur	The kurtosis of the first-order derivative of the spectrum in red edges	−0.8090 **	Ref. [40]
TcC-G-R	Colorimetric color F22	0.8335 **	Ref. [31]
TcC-R-Y	Colorimetric color F97	−0.8312 **	Ref. [31]
TcC-a*/b*	Colorimetric color F67	−0.8243 **	Ref. [31]
TcT-TOM	Third-order center moment	−0.3372 **	Ref. [31]
TcT-Corr	Similarity of grayscale levels in orientation	0.2712 *	Ref. [31]
TcT-Con	Difference in gray level	−0.2379 *	Ref. [31]

Tip: * and ** indicate a 5% and 1% significance level, respectively.

and

Table 6. The top three features corresponding to the best correlation coefficient (CC) in P3.

Feature	Formula/Description	CC	Source
SpI-RSMI	1 − (0.64R890 + 0.36R967)/R914	0.7899 **	Ref. [27]
SpI-mNDRE	(R913 − R938)/(R913 + R938 – 2 * R796)	0.7742 **	Ref. [37]
SpI-PSRI	(R944 − R963)/R1034	0.7342 **	Ref. [37]
Tri-D_r	The maximum value of the first derivative of the spectrum in the red edge (680–760 nm)	0.3080 **	Ref. [41]
Tri-NDYE	$(D_y-D_{ymin})/\left(D_y+D_{ymin}\right)$	−0.2849 *	Ref. [40]
Tri-D_b	The maximum value of the first derivative of the spectrum in the blue edge (490–530 nm)	0.2294	Ref. [41]
TcC-R-Y	Colorimetric color F97	−0.4358 **	Ref. [31]
TcC-DGCV	Colorimetric color F95	0.4209 **	Ref. [31]
TcC-DGCI/V	Colorimetric color F62	0.4168 **	Ref. [31]
TcT-TOM	Third-order center moment	−0.3386 **	Ref. [31]
TcT-Smooth	Local changes in texture	−0.2751 *	Ref. [31]
TcT-Corr	Similarity of grayscale levels in orientation	−0.2646 *	Ref. [31]

Tips: * and ** indicate a 5% and 1% significance level, respectively.

.

$$\mathrm{RSI} ({\lambda }_1, {\lambda }_2)=\frac{R_{{\lambda }_1}}{R_{{\lambda }_2}}$$

(1)

$$\mathrm{NDSI} ({\lambda }_1, {\lambda }_2)=\frac{R_{{\lambda }_1}-R_{{\lambda }_2}}{R_{{\lambda }_1}+R_{{\lambda }_2}}$$

(2)

$$\mathrm{DSI} ({\lambda }_1, {\lambda }_2)=R_{{\lambda }_1}-R_{{\lambda }_2}$$

(3)

$$\mathrm{PSRI} ({\lambda }_1, {\lambda }_2, {\lambda }_3)=\frac{R_{{\lambda }_1}-R_{{\lambda }_2}}{R_{{\lambda }_3}}$$

(4)

$$\mathrm{RSMI} ({\lambda }_1, {\lambda }_2, {\lambda }_3)=1-\frac{0.64R_{{\lambda }_1}+0.36R_{{\lambda }_2}}{R_{{\lambda }_3}}$$

(5)

$$\mathrm{SIPI} ({\lambda }_1, {\lambda }_2, {\lambda }_3)=\frac{R_{{\lambda }_1}-R_{{\lambda }_2}}{R_{{\lambda }_1}+R_{{\lambda }_3}}$$

(6)

$$\mathrm{mNDRE} ({\lambda }_1, {\lambda }_2, {\lambda }_3)=\frac{R_{{\lambda }_1}-R_{{\lambda }_2}}{R_{{\lambda }_1}+R_{{\lambda }_2}-2R_{{\lambda }_3}}$$

(7)

The color and texture features in imaging hyperspectral provide information about color differences and spatial distribution of the samples. The spectral index reflects characteristics derived from combinations of different bands in hyperspectral reflectance, while spectral three-edge parameters describe the hyperspectral performance at potentially sensitive locations. Therefore, various features may contribute to better model performance, to some extent, in estimating the pod’s water content.

2.3. Feature Fusion and Modeling

2.3.1. Feature Fusion

By using raw data and its feature information as training datasets for machine learning algorithms, information fusion can be achieved, bringing many advantages such as improving estimation accuracy, enhancing feature expression, and improving model stability and interpretability. Information fusion can be classified into three types based on the level of processing: data-level fusion, feature-level fusion, and decision-level fusion [42]. In this work, a feature-level fusion strategy is adopted. Firstly, the correlation analysis is conducted between each type of feature introduced above and the pod’s water content. Three features with the highest correlation are selected for each category, followed by an intra-class combination (CB) and inter-class fusion (FS) processing of these features. CB-TcC{*,*,*}, CB-TcT{*,*,*}, CB-SpI{*,*,*}, and CB-Tri{*,*,*} represent the intra-class combinations of the four types of features, respectively. Inter-class feature fusion is represented by FS{*,*,… }; for example, FS{CB-SpI, CB-Tri} denotes the deep fusion of the spectral index and three-edge parameters, in which there are six features.

2.3.2. Modeling Methods and Processes

Random decision forest (RDF) is a highly integrated supervised learning algorithm that utilizes decision trees as estimators [43]. Decision trees are constructed from randomly selected subsets of data and features, and classification results are obtained through bagging resampling and a voting mechanism, suitable for both classification and regression problems. The RDF model is flexible in structure, strong in fault tolerance, and easy to understand and implement [44]. We use the RDF to build an estimation model of the pod’s water content, and a recognition model of the pod’s maturity levels. Meanwhile, the effects of other modeling methods, such as linear regression, support vector machines, and neural networks are compared. The modeling process is presented in Figure 2.

According to the diagram, we execute the tasks of the pod’s water estimation modeling and maturity identification modeling in four steps.

Step 1:

Significance analysis

Firstly, we conduct variance analysis on the pod’s water content samples from different periods, positions, and colors. The results show that the water content was significantly affected only by the period, while the three positions and three colors had no significant difference in their effects on water content. Based on this finding, we decided to use spectral data only for period classification and recognition.

Step 2:

Feature extraction and data fusion

Next, we select the three most correlated feature parameters from each of the four feature categories through correlation analysis with the water content of the silique pericarp. We then form intra-class combinations (CB-*{*,*,…}) and inter-class fusion combinations (FS{CB-*, CB-*…}) as inputs for subsequent models.

Step 3:

Model Construction

Based on various feature combinations, we fuse the extracted feature data using machine learning algorithms. By comparing evaluation metrics from 100 model runs, we establish a quantitative model to invert the pod’s water content. Targeting different pod’s maturity periods, we use the coefficient of variation I_o to select the most sensitive features from spectral indices, color, texture, and spectral parameters.

Step 4:

Model Validation and Optimization

Finally, we test the accuracy and stability of both the pod’s water content inversion model and the pod’s maturity recognition model by varying the number of training set samples with K-fold cross-validation.

2.3.3. Model Evaluation

In this paper, root mean square error (RMSE), relative error (RE), and relative root mean square error (RRMSE) are adopted as metrics to evaluate the modeling results. Among them, RMSE can effectively reflect the accuracy of the estimation model. The closer its value is to 0, the higher the estimation accuracy of the model. RE measures the absolute error ratio between the measured value and the true value, and a smaller RE value indicates smaller errors. RRMSE, defined as the ratio of RMSE to the average value of the target variable and expressed as a percentage, can eliminate the influence of dimensions on the evaluation results [45]. The formulas are as follows:

$$RMSE=\sqrt{\frac{1}{n}{{\displaystyle \sum }}_{i=1}^n{\left({\widehat{Y}}_i-Y_i\right)}^2}$$

(8)

$$RE=\frac{1}{n}{{\displaystyle \sum }}_{i=1}^n\left|\frac{{\widehat{Y}}_i-Y_i}{Y_i}\right|\times 100\%$$

(9)

$$RRMSE=\frac{RMSE}{\overline{Y}}\times 100\%$$

(10)

3. Results and Discussion

3.1. Significant Difference Test of Pod’s Water Content

The pod samples were picked from three growth locations (see Section 2.1) at three color levels. To detect whether there is a difference in the water content of pods under these two factors, analysis of variance (ANOVA) is conducted at the three periods (P1, P2, P3) as well as the whole period (PW) of the pods. We listed the statistical p-values in

Table 7. Significant p-value of pod’s water content for different pod positions, color combinations, and their interactions.

Period	Position	Color	Pi vs. Pj	PW
P1	0.2112 N	0.9947 N	P1 vs. P2: 0.012 *	8.1042 × 10−101 **
P2	0.9478 N	0.5625 N	P1 vs. P3: 4.5087 × 10−61 **
P3	0.9684 N	0.5373 N	P2 vs. P3: 7.1597 × 10−73 **

Tips: * and ** indicate a 5% and 1% significance level, respectively; ^N indicates no significant difference.

. Clearly, during the periods of P1, P2, and P3, both position and color have no significant effect on the pod’s water content. The potential reasons may be (1) in the P1, the vascular bundle structure inside the rape stalks is intact and elastic at this stage, allowing normal transport of water and nutrients. (2) In the P2, though the pods begin to turn yellow, the pods are not fully ripe yet. The stalks still retain their toughness and the vascular bundles function normally, thus pods in different positions and colors still have normal water absorption capabilities. (3) In the P3, the rape seeds are completely mature, and the pods turn completely yellow and gradually dry. Although the stalks show signs of aging, such as slight lignification or color changes, the water and nutrient transport capacity of the vascular bundles may be somewhat reduced. However, they are still relatively intact and not sufficient to significantly affect the water content of different pods.

ANOVA between periods shows that the pod’s water content is significantly affected by the growth stage. In this regard, we focus on the pod’s water content modeling in each period and ignore the influence of pod position and color.

3.2. Correlation Analysis

To select feature parameters to estimate the pod’s water content, we perform the correlation analysis between the pod’s water content and the various feature parameters. Figure 3a–c are heatmaps of the correlation coefficients between any two-band combinations of NDSI, DSI, and RSI spectral index (given by Equations (1)–(3)) in the range of 375–1041 nm and the pod’s water content at P1. The deeper the color, the higher the degree of correlation. The red pentagram corresponds to the maximum correlation coefficient. In addition, the other four spectral indices, PSRI, RSMI, SIPI, and mNDRE (given by Equations (4)–(7)), are also studied. Figure 3d shows the correlation coefficients between the optimal band combinations of seven spectral indices and the pod’s water content. The top three are PSRI, RSMI, and SIPI, with correlation coefficients of 0.9425, 0.9178, and 0.9146, respectively. Meanwhile, the nine spectral three-edge parameters are also performed in the correlation analysis, as shown in Figure 3e. Among them, NDBE, NDYE, and NDRE have the highest correlation with the water content, with the correlation coefficients −0.7888, −0.7756, and −0.7506, respectively. Figure 3f shows the correlation of the top six image colors and all image textures within the pod’s water content. Table 4 shows the top three feature parameters with the best correlation of the above four types of features during P1. Similarly, the top three features corresponding to the best correlations in P2 and P3 are shown in Table 5 and Table 6, respectively. In what follows, we use these feature parameters as feature combinations or fusions to establish models for estimating the pod’s water content.

3.3. Pod’s Water Content Estimation Modelling

3.3.1. Analysis and Comparison of the Estimation Model in P1

An estimation model for the pod’s water content in P1 is established using the RDF algorithm. In total, 56 samples are randomly selected as the training set, while the remaining samples serve as the test set. The pod’s water content is set as the output layer, and the combined/fused features are used as the input layer. According to Table 4, those three features in identical types are combined, and two, three, or four combinations are fused (denoted as FS{CB-*,CB-*,…}). To eliminate the influence of random factors, 100 repetitions are conducted. The results are obtained on the test set, which are listed in

Table 8. Performance of modeling on the pod’s water content with respect to various feature combinations in P1. The best result of each case is in bold.

	Feature Combination/Fusion	Training Set (N = 56)			Testing Set (N = 16)
		RRMSE (%)	RMSE	RE (%)	RRMSE (%)	RMSE	RE (%)
Single feature	CB-SpI{PSRI, RSMI, SIPI}	2.0450	0.0160	1.3618	4.8181	0.0378	3.1292
	CB-Tri{NDBE, NDYE, NDRE}	3.1269	0.0245	2.2877	6.3252	0.0493	4.6913
	CB-TcC{(2G-R-B)/L*, G-R, G/Y}	6.3703	0.0497	5.4892	12.6718	0.0993	11.0284
	CB-TcT{Homogeneity, Contrast ratio, Correlation}	6.4786	0.0506	5.7543	13.0173	0.1018	11.5910
Two features	FS{CB-TcC, CB-TcT }	5.4930	0.0429	4.7322	11.7985	0.0925	10.2612
	FS{CB-TcC, CB-SpI}	2.0670	0.0162	1.3340	4.7556	0.0373	3.0744
	FS{CB-TcC, CB-Tri}	3.0961	0.0242	2.2290	6.2319	0.0488	4.6117
	FS{CB-TcT, CB-SpI}	2.0036	0.0157	1.3177	4.6865	0.0367	3.1346
	FS{CB-TcT, CB-Tri}	2.9059	0.0227	2.1301	6.2865	0.0491	4.6520
	FS{CB-SpI, CB-Tri}	2.1186	0.0166	1.3396	4.3458	0.0340	2.8841
Three features	FS{CB-TcC, CB-TcT, CB-SpI}	2.0355	0.0159	1.3477	4.6504	0.0363	3.1407
	FS{CB-TcC, CB-TcT, CB-Tri}	2.8719	0.0224	2.1187	5.9199	0.0464	4.4864
	FS{CB-TcC, CB-SpI, CB-Tri}	2.1337	0.0167	1.3528	4.2670	0.0334	2.9191
	FS{CB-TcT, CB-SpI, CB-Tri}	2.0633	0.0161	1.3197	4.2788	0.0336	2.8983
Four features	FS{CB-TcC, CB-TcT, CB-SpI, CB-Tri}	2.0703	0.0162	1.3320	4.2565	0.0334	2.9293
Other models	Robust regression	4.5691	0.0357	2.7804	4.7751	0.0374	3.3251
	NNfit	4.8347	0.0378	3.4436	7.5282	0.0591	5.6168
	Stepwisefit regression	3.9717	0.0311	3.1847	4.9380	0.0387	3.6461
	Multivariate linear regression	3.8111	0.0298	2.9306	5.0740	0.0398	3.6875
	Support vector machine	1.1649	0.0091	1.1533	7.3547	0.0574	5.4737

. Figure 4 illustrates RRMSE, RMSE, and RE, as well as their standard deviations under the best feature combination/fusion. The results indicate that multi-category fusion is superior to single-category combination. As the number of feature categories increases, the overall trend of various indicators decreases and the standard deviation also decreases. It suggests that the accuracy and stability of the model improve as the number of feature parameter categories increases. However, the trend of model optimization slows down when the number of fusion combination categories reaches two or more. To confirm this, a T-test is implemented for them, and the statistical p-value is listed in

Table 9. T-test of statistical p-values of feature fusion in P1.

Feature Fusion	RRMSE	RMSE	RE (%)
Single feature vs. Two features	0.0262 *	0.0238 *	0.0171 *
Two features vs. Three features	0.7151 N	0.7563 N	0.7447 N
Two features vs. Four features	0.6712 N	0.7145 N	0.6688 N
Three features vs. Four features	0.9586 N	0.9611 N	0.9194 N

Tips: * is at a 5% significant level and ^N indicates no significant difference.

. The fusion effect by two types of features is significant. In contrast, there is no significant difference between more feature fusions. In this regard, we suggest the two types of feature fusion FS{CB-SpI, CB-Tri} for the pod’s water content modeling for the sake of modeling efficiency. The RRMSE, RMSE, and RE are 4.3458%, 0.0340, and 2.8841%, respectively. To test the performance of the RDF, we also apply other five statistical models for the same purpose; the results are obtained on the dual feature fusion FS{CB-SpI, CB-Tri}, also listed in Table 8.

We show the RDF model performance under the two types of features fusion FS{CB-SpI, CB-Tri} in P1 in Figure 5. The left panel illustrates the fitted and observed values of the pod’s water content on the training set. The right panels compare the observed and estimated water content values on the testing set.

To further investigate the influence of spectral index and spectral three-edge parameters on the model performance, we compare the results that include these features with those that do not, as listed in

Table 10. Influence of the spectral index and spectral three-edge parameters on model performance in P1 (the best result of each case is in bold).

		Feature	Training Set (N = 56)			Testing Set (N = 16)
			RRMSE (%)	RMSE	RE (%)	RRMSE (%)	RMSE	RE (%)
Spectral Index	Two types of features	Include	2.0631	0.0162	1.3304	4.5960	0.0360	3.0310
		Exclude	3.8317	0.0299	3.0304	8.1056	0.0635	6.5083
	Three types of features	Include	2.0775	0.0162	1.3401	4.3987	0.0344	2.9860
		Exclude	2.8719	0.0224	2.1187	5.9199	0.0464	4.4864
Three-edge parameters	Two types of features	Include	2.7069	0.0212	1.8996	5.6214	0.0440	4.0493
		Exclude	3.1879	0.0249	2.4613	7.0802	0.0555	5.4901
	Three types of features	Include	2.3563	0.0184	1.5971	4.8219	0.0378	3.4346
		Exclude	2.0355	0.0159	1.3477	4.6504	0.0363	3.1407

. It is observed that the spectral index and spectral three-edge parameters indeed have a significant positive effect on the model. By comparison, the former demonstrates a better performance than the latter.

To evaluate the stability of the pod’s water content estimation model, we vary the number of samples in the training set, i.e., the number of samples is set from 40 to 64 with random selection. The average results of the three evaluation indicators (RRMSE, RMSE, and RE) are calculated, as shown in Figure 6. It is found that there is no significant change with an increase in the training set size, indicating that the selected fused feature exhibits satisfactory stability.

3.3.2. Analysis and Comparison of Estimation Models for P2 and P3

The pod’s water content modeling in P2 and P3 is conducted similarly to P1.

Table 11. Optimal feature combination/fusion in the three pod stages.

Feature Combination	P1	P2	P3
Single-type feature	CB-SpI{PSRI, RSMI, SIPI} CB-Tri{NDBE, NDYE, NDRE}	CB-SpI{PSRI, mNDRE, SIPI} CB-Tri{NDBE, D_rmin, Kur}	CB-SpI{RSMI, mNDRE, PSRI} CB-Tri{D_r, NDYE, D_b}
Two types of features	FS{CB-SpI, CB-Tri}	FS{CB-SpI, CB-Tri}	FS{CB-SpI, CB-Tri}

lists the optimal single-feature combinations and two types of feature fusions. The optimal two types of feature fusions in P2 and P3 remain the same as in P1, i.e., FS{CB-SpI, CB-Tri}. However, the optimal feature combination differs between the three periods. For P2, the best RRMSE, RMSE, and RE are 2.6218%, 0.0212, and 1.8820%, respectively. For P3, the RRMSE, RMSE, and RE are 34.1718%, 0.0758, and 25.0430%, respectively. The poor model performance in P3 is due to the very low water content of the pods, which cannot be accurately identified by existing instruments. It is necessary to continuously improve the resolution and other performances of the experimental equipment. Additionally, further research is needed to expand the spectral band range, innovate more effective feature extraction methods, and develop new modeling approaches, which are the directions for future exploration.

3.4. Recognition Model for Pod’s Maturity Level

3.4.1. Analysis of Recognition Models

Recognizing the maturity level of rapeseed pods can help determine the growth status of rapeseed and the optimal harvest timing, preventing low grain weight and oil content caused by premature harvesting, or pod cracking, grain and oil content loss caused by late harvesting. Therefore, accurately judging the maturity level of rapeseed pods and determining the appropriate harvest time can help managers arrange farming activities reasonably and improve efficiency.

Based on the significant difference test for the pod’s water content presented in Section 3.1, it is evident that the water content is primarily influenced by the pod’s different growth stages. To intelligently distinguish the maturity level of the pods, we extract the features by referencing sensitive characteristics of entire pod growth stages from the imaging hyperspectral data first. This resulted in 151 features, namely 16 spectral indices, 27 spectral three-edge parameters, 98 image color features, and 10 image texture features for 216 pod samples. To improve efficiency during modeling, an indicator Io, similar to the variation coefficient, is proposed to screen the optimal features. On the one hand, let σ_i represent the standard deviation of feature values among 72 samples in period Pi (i = 1, 2, 3), characterizing the intra-period difference. Marking the average value over the three periods as ${〈{\sigma }_i〉}_i$. On the other hand, σ_a denotes the standard deviation of the average value of the studied feature across the three periods, indicating inter-period difference. The indicator I_o is the ratio of these two standard deviations. A smaller I_o indicates that the feature parameters selected using this variation coefficient are more effective in distinguishing between periods. The formula is as follows:

$$I_o=\frac{{〈{\sigma }_i〉}_i}{{\sigma }_a}$$

(11)

Accordingly, we screen the features with the smallest I_o from each type of feature, as listed in

Table 12. Ranking of I_o for various types of features.

Feature Type	Spectral Index	Image Color	Three-Edge Parameter	Image Texture
Feature name	RSMI(λ643, λ669, λ662)	a*/b*	NDBE	Smoothness
Io	0.4198	0.1861	0.2210	0.5271

. We combine the former three features to construct three-dimensional parameter space {SpI-RSMI, TcC-a*/b*, Tri-NDBE}, as shown in Figure 7, where each point represents a pod’s sample, and different shape symbols denote different pod growth stage. Obviously, the samples of the same period are clustered together, while those from the different periods are well separated.

Subsequently, the combined features {SpI-RSMI, TcC-a*/b*, Tri-NDBE} are used as inputs for the recognition model. Here, besides the RDF model used for the pod’s water content estimation, we employ two other prevail classification algorithms in machine learning for the recognition model, namely support vector machine and kernel method (SVM-KM) and K-nearest neighbor (KNN). For these algorithms, RDF excels in handling high-dimensional data and feature selection [46], SVM performs well in both linear and nonlinear classification problems [47,48], and KNN has advantages in certain applications due to its simplicity and instance-based learning approach [49]. We adopt K-fold cross-validation to test our model, where (K − 1)/K × 100% of the samples are set as the training set and the ones remaining are in the testing set. By setting K = 2, 4, 6, 8, and 10, the average recognition accuracy for identifying pod periods is obtained, as shown in

Table 13. Average accuracy (%) of the three models on {SpI-RSMI, TcC-a*/b*, Tri-NDBE} under K-fold cross-validation.

K-Fold	RDF	SVM-KM	KNN
2	93.38	95.76	95.2
4	93.63	96.33	95.91
6	93.89	96.36	95.94
8	93.93	96.44	96.15
10	94.29	96.90	96.24

. The best recognition effect is achieved using SVM, with 96.90% accuracy. The average modeling accuracy obtained with K = 2–10 ranges from 95% to 97%, demonstrating the good accuracy and stability of the model in recognizing pod periods.

3.4.2. Model Validation

To demonstrate the effectiveness of the indicator of I_o, we randomly choose three other features to combine the three-dimensional features for comparison, namely {SpI-SIPI, TcC-G/Y, Tri-NDRE}, {SpI-SIPI, TcC-(2G-R-B)/L*, Tri-D_rmin}, and {SpI-mNDRE, TcC-DGCV, Tri-NDYE}. Using the SVM-KM algorithm as the classifier, the modeling process is the same as above. The average accuracy is listed in

Table 14. Average accuracy (%) on different feature combinations under K-fold cross-validation.

K-Fold	{SpI-RSMI(Ⅰ), TcC-a*/b*, Tri-NDBE}	{SpI-SIPI(Ⅰ), TcC-G/Y, Tri-NDRE}	{SpI-SIPI(Ⅱ), TcC-(2G-R-B)/L*, Tri-D_rmin}	{SpI-mNDRE(Ⅱ), TcC-DGCV, Tri-NDYE}
2	95.76	92.11	94.51	93.42
4	96.33	92.09	95.11	93.06
6	96.36	91.72	95.89	93.08
8	96.44	92.37	95.63	93.19
10	96.9	92.24	96.33	93.29

. Clearly, the optimal feature parameter combination {SpI-RSMI, TcC-a*/b*, Tri-NDBE} selected by I_o has the highest accuracy in period recognition.

Finally, we investigate the impact of the number of features on the modeling results. SVM-KM and 5-fold cross-validation are used for this purpose. The number of features from 1 to 10 is varied, which is selected from the ascending order of the indicator Io. Figure 8 reports the result. The average accuracy arises sharply when the feature number is from 1 to 3; however, it remains almost unchanged when the number of features exceeds 3. Considering the modeling performance and efficiency, we suggest using the three feature combinations {SpI-RSMI, TcC-a*/b*, Tri-NDBE} for recognition modeling of the pod growth stage.

4. Discussion and Conclusions

In this paper, imaging spectrometry technology is utilized to construct inversion models for water content in rapeseed pod pericarp and recognition models for pod’s maturity. By fusing features such as image color, texture, spectral three-edge parameters, and spectral indices, it is found that the inter-class feature fusion exhibits higher accuracy and stability in modeling compared to the intra-class feature combination. The best performance is achieved with the inter-class feature fusion combination of FS{CB-SpI, CB-Tri} in all periods. Meanwhile, spectral indices significantly enhance the model’s performance. When establishing the recognition model, variance analysis is employed, and a difference index I_o is defined to filter features, resulting in the optimal feature combination of {SpI-RSMI, TcC-a*/b*, Tri-NDBE}. Finally, it is observed that the random forest algorithm performs better in modeling the estimation problems studied in this paper, while the support vector machine algorithm excels in recognition problems. These results provide meaningful references for managing photosynthetic efficiency measures during the pod stage of rapeseed and formulating and implementing low-loss harvesting and drying practices.

It is important to highlight that our focus has been on providing reliable feature selection and interpretable model construction for estimating the pod’s water content and identifying the pod’s maturity. To this end, we conducted thorough feature engineering and machine learning modeling processes.

Existing studies on rapeseed pod’s maturity level recognition or water content estimation are limited. In the existing works [37,38], the authors proposed novel spectral indices, the Canola Pod Maturity Index (CPMI) and Rape Siliques Maturity Index (RSMI), respectively. While the CPMI exhibited satisfactory model performance for canola [37], its correlation with the water content of our rapeseed pods was poor. The RSMI, proposed in [38] and designed for Chinese rapeseed pod cultivars, demonstrated a good correlation coefficient for our data. Consequently, we employed RSMI (given by Equation (5)) as a feature for our modeling. This finding highlights the significant regional differences in rapeseed materials, manifested as variations in optimal spectral indices. In light of the global importance of rapeseed materials, the models proposed in this work can provide valuable experience for subsequent related research.