Oilseed Rape Yield Prediction from UAVs Using Vegetation Index and Machine Learning: A Case Study in East China

Abstract

Yield prediction is an important agriculture management for crop policy making. In recent years, unmanned aerial vehicles (UAVs) and spectral sensor technology have been widely used in crop production. This study aims to evaluate the ability of UAVs equipped with spectral sensors to predict oilseed rape yield. In an experiment, RGB and hyperspectral images were captured using a UAV at the seedling (S1), budding (S2), flowering (S3), and pod (S4) stages in oilseed rape plants. Canopy reflectance and spectral indices of oilseed rape were extracted and calculated from the hyperspectral images. After correlation analysis and principal component analysis (PCA), input spectral indices were screened to build yield prediction models using random forest regression (RF), multiple linear regression (MLR), and support vector machine regression (SVM). The results showed that UAVs equipped with spectral sensors have great potential in predicting crop yield at a large scale. Machine learning approaches such as RF can improve the accuracy of yield models in comparison with traditional methods (e.g., MLR). The RF-based training model had the highest determination coefficient (R²) (0.925) and lowest relative root mean square error (RRMSE) (5.91%). In testing, the MLR-based model had the highest R² (0.732) and lowest RRMSE (11.26%). Moreover, we found that S2 was the best stage for predicting oilseed rape yield compared with the other growth stages. This study demonstrates a relatively accurate prediction for crop yield and provides valuable insight for field crop management.

1. Introduction

Reliable prediction of crop yield is important for farmers to determine agricultural management. An accurate crop yield prediction model can help farmers decide on what to plant, when to grow, and when to harvest. Different approaches have been developed to predict the yields of various crops [1,2,3,4,5]. Many of these studies have used traditional methods to estimate crop yield by collecting soil and climate parameters [6]. However, these methods are time-consuming and do not fully describe the spatial variability of yield [7]. In contrast to traditional methods, remote sensing techniques are relatively economical and can provide quick and convenient solutions to predicting crop yields. Recently, vegetation indices (VIs) calculated from spectral reflectance in remote sensing have been developed to predict crop yields; their use requires less labor and avoids crop destruction [8,9].

Generally, near-infrared reflectance (NIR) and the visible spectrum are considered to be related to crop yield. VIs are calculated mainly by a combination of the above reflectance in two or more different bands. The reliability and potential of many indices have been investigated in various studies to estimate the crop yield of different crops. Several VIs, such as water index (WI), soil-adjusted vegetation index (SAVI), green index (GI), and green vegetation index (GVI), have been proposed [10,11,12]. The normalized difference vegetation index (NDVI) is widely used to estimate crop yield due to the significant correlation between crop growth conditions and productivity [13].

Usually, a single index alone may not provide excellent prediction all the time and everywhere. Combining multiple indices may lead to better predictive performance. However, a large number of VIs may also make data analysis and model building complex and difficult [14]. Therefore, many researchers have investigated the optimal combination of VIs for the most accurate yield estimation. In 2022, Kayad et al. [15] used normalized difference red-edge (NDRE) and NDVI to predict maize yield, and the NDRE- and NDVI-based model’s determination coefficient (R²) was 0.85 and 0.92, respectively. VIs may contain a large amount of information, and they usually have a nonlinear relationship with crop yield. Feature filtering, such as correlation analysis, importance ranking, and principal component analysis (PCA), can be applied to help us select model input variables.

Machine learning has big advantages in processing large amounts of nonlinear data and is capable of analyzing spectral big data. Many machine learning algorithms, such as artificial neural networks (ANNs) [16], support vector machines (SVM) [17], random forests (RFs) [18], and AdaBoost [19], have successfully predicted the yields of various crops. They each have different advantages. Rajković et al. [20] found that ANNs have a high ability to model nonlinear complex relationships between dependent and independent variables. Meanwhile, support vector machines (SVMs) have the potential to solve the overfitting problem when using high-dimensional data such as spectral data [14]. Spectral indices have been used in machine learning methods for the yield estimation of rice [21], potatoes [22,23], wheat [4,24], and cotton [25]. Marques Ramos et al. [18] employed five different machine learning methods with vegetation indices to predict maize yield, and they found that RF predicted maize yield more accurately than other models.

UAVs can be equipped with many portable devices and have been widely used in various agricultural tasks. A UAV can accomplish seeding, fertilization, spraying, and crop monitoring for a low cost and taking little time [26]. UAVs mounted with spectral sensors can collect crop canopy spectral reflectance and then obtain vegetation indices through spectral analysis. Crop yields have been predicted using such VIs by many researchers [12,27]. However, there are limited studies on predicting oilseed rape yield at different growth stages using spectral VIs obtained by UAVs and machine learning methods.

Considering all the above, the aims of the present work are (1) to investigate the effect of the growth stage on oilseed rape yield model performance, (2) to compare the oilseed rape yield model performances of four different machine learning methods and reveal the best solution, and (3) to investigate the effects of different input strategies on oilseed rape model performance. Our study improves the relationships between oilseed rape yield and predictive VI variables obtained by UAV proximal sensing.

2. Materials and Methods

2.1. Study Area and Experimental Design

The test site was located at the experiment field base of the Zhejiang Academy of Agricultural Sciences, Liangzhu town, Hangzhou city, Zhejiang province, China (Figure 1). It has a monsoon climate and is on the southern edge of the northern subtropical zone. Its winter and summer are long, while spring and autumn are short, warm, and humid; the region experiences abundant sunlight and rainfall. The annual average temperature ranges from 15.3 °C to 16.2 °C, and the annual average rainfall ranges from 1150 to 1550 mm (https://upimg.baike.so.com/doc/5366320-5602031.html. In Chinese, accessed on 1 July 2024).

Details on the experimental design are shown in

Table 1. Experimental design details.

Notation	Treatment	Note
F1	N fertilizer 125 kg/ha	Repeated 12 times
F2	N fertilizer 500 kg/ha
F3	N fertilizer 750 kg/ha
NHB	Herbicide (S)-Metolachlor sprayed with 1.5 L/ha	Repeated 12 times
HB	Without herbicide (S)-Metolachlor sprayed
D1	Density of 1.5 × 105 seedlings/ha	Repeated 6 times
D2	Density of 4.5 × 105 seedlings/ha
D3	Density of 7.5 × 105 seedlings/ha
CF	Compound fertilizer (N:P2O5:K2O = 16:16:16)	Repeated 3 times
SF	Special fertilizer for oilseed rape plant

. In order to obtain gradient yield data, we investigated nitrogen fertilizer application (F1, F2, and F3), herbicide (NHB and HB), planting density (D1, D2, and D3), and oilseed rape special fertilizer application (CF and SF). The field experiment area could be seen in Figure 1. The study (the above four experiment treatments) involved a total of 84 field plots. Each field plot was 10.0 m long and 2.0 m wide. Oilseed rape seedlings with three leaves were transplanted on October 1 in 2021 at a depth of 10–15 cm. Except where otherwise stated, oilseed rape was planted with a density of 4.5 × 10⁵ seedlings/ha, compound fertilizer (N:P₂O₅:K₂O = 16:16:16) at 750 kg/ha was applied, and pre-emergence herbicide (S)-Metolachlor (1.5 L/ha) was sprayed. Oilseed rape was manually harvested and dry-weighed plot by plot on 20 May 2022. Yield was expressed as kg/ha.

2.2. Data Acquisition

We chose two UAVs as platforms to obtain the RGB and hyperspectral image collection: a DJ Phantom-4-RTK and a DJ M600 Pro (DJI Technology Co., Shenzhen, China). An aerial survey was conducted under conditions of a low wind speed and clear skies from 11:00 AM to 2:00 PM local time on February 4 (seedling, (S1), March 11 (budding, S2), April 3 (flowering, S3), and April 25 (pod, S4), 2022. The mission was planned at a flight height of 60 m and with a flight speed of 6 m/s, with a front overlap of 80% and a side overlap of 70%.

A visible and near-infrared hyperspectral imaging system, the Pika XC2 camera (Resnon Inc., Bozeman, MT, USA), with a spectral range of 400–1100 nm, a spectral resolution of 1.3 nm, 447 spectral channels, and 1600 spatial channels with a 12-bit depth, was mounted and fixed on the DJ Matrice 600 Pro (DJI, Shenzhen, China) to capture the spectral images of oilseed rape plants. The maximum payload of the UAV was 6.0 kg and that of the Pika XC2 camera was 2.51 kg. Captured images were stored in memory cards. Reflectance of the oilseed rape was extracted using the software Spectronon Pro 1.0 (Resonon Inc., Bozeman, MT, USA), and spectral VIs were calculated as shown in

Table 2. Summary of the spectral reflectance indices used in this study.

Index	Equation	Reference
NDVI	(R800 − R670)/(R800 + R670)	[28]
NDVY	(R560 − R450)/(R560 + R450)	[29]
CIre	R800/R720-1	[30]
CIgreen	R800/R550-1	[30]
RVI	R800/R670	[31]
TVI	0.5 × (120 × (R750 − R550) − 200 × (R670 − R550)	[32]
BNDVI	(R860 − R450)/(R860 + R450)	[33]
DVI	R800-R670	[34]
RDVI	[NDVI × (R800 + R670) ^ 2] ^ 0.5	[35]
NIR	R860	/

Rx refers to the leaf reflectance at a wavelength x in nanometers; NDVI is normalized difference vegetation index; NDVY is the normalized yellow difference vegetation index; CIre is the red-edge chlorophyll index; CIgreen is the green chlorophyll index; RVI is the ratio vegetation index; TVI is triangular vegetation index; BNDVI is the blue normalized difference vegetation index; DVI is the difference vegetation index; RDVI is the ratio difference vegetation index; NIR is the near-infrared reflectance.

.

2.3. Data Reduction

Hyperspectral images containing huge amounts of data may make data analysis complex and difficult. Thus, it is necessary to reduce the data and select the appropriate spectral VIs to construct the yield prediction model. We applied correlation analysis and PCA to reduce the dimension and amount of large spectral data. A two-tailed test with p < 0.05 (*) was used in the correlation analysis between reflectance, VIs, and oilseed rape yield. After correlation analysis, PCA with two principal components was conducted. The resulting scores allowed us to screen VIs with a close relationship with oilseed rape yield; these VIs were then used as model input variables.

2.4. Data Modeling

For modeling, two-thirds of the data (56 data points) were used for training, while the remaining third (28 data points) were used independently for testing. To build the oilseed rape yield prediction model, we used four machine learning regression methods: RF, multiple linear regression (MLR), and SVM. RF is a commonly used machine learning algorithm. It can combine the output of multiple decision trees to reach a single result. In the RF, the number of decision trees was set to 500, and other parameters were set to the algorithm’s default values. For MLR, “input” is the model construction method. SVM is commonly used for regression analysis due to its powerful supervised learning algorithms. The employed SVM used default parameters.

2.5. Model Performance

The flowchart of the entire process, starting from image collection to data analysis and modeling, is shown in Figure 2. Before modeling, data of all the input variables were normalized using Equation (1):

$$x_n=\frac{x_i-x_{\min }}{x_{\max }-x_{\min }}$$

(1)

Here, $x_i$ is an original value, $x_n$ is a normalized value, and $x_{\max }$ and $x_{\min }$ are the maximum and minimum values, respectively.

Figure 2. Development workflow for the yield prediction model.

Figure 2. Development workflow for the yield prediction model.

The R², root mean square error (RMSE), and relative root mean squared error (RRMSE) were used to evaluate training and testing performance (Equations (2)–(4)):

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

(2)

$$RMSE=\sqrt{\frac{1}{N}{\displaystyle \sum _{i=1}^N{(y_i-\widehat{y})}^2}}$$

(3)

$$RRMSE=\sqrt{\frac{\frac{1}{N}{\displaystyle {\sum }_{i=1}^N{(y_i-\widehat{y})}^2}}{{\displaystyle {\sum }_{i=1}^N{(y_i)}^2}}}\times 100\%$$

(4)

where $x$ and $y$ are the measured and predicted values, respectively; and $\overline{x}$ and $\overline{y}$ are the average values of measured and predicted yield, respectively. N is the number of observations.

We conducted a Pearson correlation analysis and PCA using IBM SPSS Statistics 25.0 (SPSS Inc., New York, NY, USA, 2017). The correlation, scatter, fitted line, and PCA plot were drawn by Origin 2022 (Origin Lab Corporation, Northampton, MA, USA, 2022). The machine learning algorithms were run in MATLAB R2022a (The MathWorks, Inc., Natick, MA, USA, 2022).

3. Results

3.1. Oilseed Rape Yield

The yield of oilseed rape plants following various treatments is shown in Figure 3. It can be seen that oilseed rape yield increased from F1 to F3, and the yield with F3 was the highest (3302.5 kg/ha). Regarding herbicide treatments, the yield with HB was 2930.04 kg/ha; this was higher than that seen with NHB treatment. The yield for the density of 4.5 × 10⁵ seedlings/ha (D2) was the highest among the density treatments. The average yield across all treatments was 2757.2 kg/ha; the minimum and maximum yields were 1071.96 and 4152.08 kg/ha, respectively.

3.2. Reflectance of Oilseed Rape Plants

The reflectance of oilseed rape plants at various growth stages is shown in Figure 4. There are obvious reflectance differences in visible and near-infrared wavelength across the S1, S2, S3, and S4 stages (Figure 4A). We note that the reflectance in the NIR wavelength was the highest in the S2 stage and the lowest in the S4 stage. Hence, reflectance in the NIR wavelength may be used to differentiate the growth stages of oilseed rape.

The relation between reflectance and oilseed rape yield is shown in Figure 4B. The reflectance from 442 to 713 nm at the S2 stage and from 737 to 921 nm at the S1 stage had a significant relationship (p < 0.05) with oilseed rape yield. Meanwhile, reflectance from 529 to 561 nm in the S1 stage and from 529 to 561 nm in the S2 stage had a greatly significant relationship (p < 0.01) with oilseed rape yield. This indicates that S2 has a greater effect on oilseed rape in comparison with the S1, S3, and S4 stages. Compared with the visible region, the NIR region was more sensitive, and NIR reflectance had a closer relationship with the yield. Therefore, R860 in NIR region was selected as another input variable for yield modeling.

3.3. VIs of Oilseed Rape Plants

We performed a correlation analysis to evaluate the relationships between VIs and oilseed rape yield. This analysis may be helpful to ultimately determine the variables that should be used as the inputs in yield prediction models. The results are shown in Figure 5. It can be seen that all VIs at S1; all VIs except NDVY at S2; CIre and RDVI at S3; and NDVI, CIre, and RVI at S4 had significant relationships with oilseed rape yield. NDVI and BNDVI at S2 had the highest correlation coefficients: 0.896 and 0.856, respectively. S1 and S2 had more VIs with a closer relationship to oilseed rape yield than other growth stages.

3.4. PCA of VIs

As seen from Figure 5, a large number of VIs had a significant relationship with oilseed rape yield. VIs also had significant correlations among themselves (Figure 5). The use of highly correlated variables as inputs in prediction models may cause multicollinearity problems. Meanwhile, too many data make the analysis difficult. To overcome these problems, we can reduce the number of VIs. PCA of VIs which had a very significant relationship with yield was performed. The PCA result is shown in Figure 6. VIs which were in the same quadrant with yield and the first two most closely related to yield were selected and considered as the oilseed rape yield model input variables. So, we can see that NDVI and CIre at S1, RVI and CIgreen at S2, CIre at S3, and CIre at S4 were ultimately selected.

3.5. Performance of Oilseed Rape Yield Prediction Model

The aforementioned selected VIs were used as the input variables to build oilseed rape yield models based on the RF, MLR, and SVM. For one stage, more than one model may be established because there may be more than one input variable and combinations. Hence, the models with the best R² were selected (see Figure 7, Figure 8 and Figure 9 and Table 3). During model training, the RF model had the highest R² (0.925) and lowest RMSE and RRMSE (167.8 kg/ha and 5.91%, respectively) across growth stages. For testing, the MLR model had the highest R² (0.732) and lowest RMSE and RRMSE (309.5 kg/ha and 11.26%, respectively). The RF models performed second and third best. Regarding the various oilseed rape growth stages, S2 had the highest R² and lowest RMSE and RRMSE across both model training and testing.

Table 3. Performance of models utilizing various machine learning methods.

Stage	Model	Training Performance			Testing Performance
		RMSE (kg/ha)	RRMSE (%)	R2	RMSE (kg/ha)	RRMSE (%)	R2
S1	RF	214.9	7.57	0.893	531.3	19.34	0.243
	MLR	370.7	13.05	0.618	559.6	20.37	0.297
	SVM	398.2	14.02	0.568	518.5	18.87	0.314
S2	RF	167.8	5.91	0.925	319.7	11.64	0.723
	MLR	309.7	10.90	0.733	309.5	11.26	0.732
	SVM	316.9	11.15	0.722	318.2	11.58	0.714
S3	RF	227.7	8.02	0.861	601.3	21.88	0.220
	MLR	398.6	14.03	0.558	546.6	19.89	0.268
	SVM	404.0.	14.22	0.558	577.0	21.00	0.268
S4	RF	247.4	8.71	0.842	543.0	19.76	0.227
	MLR	503.8	17.74	0.293	477.0	17.36	0.367
	SVM	529.0	18.62	0.293	525.4	19.12	0.367

Table 3. Performance of models utilizing various machine learning methods.

Stage	Model	Training Performance			Testing Performance
		RMSE (kg/ha)	RRMSE (%)	R2	RMSE (kg/ha)	RRMSE (%)	R2
S1	RF	214.9	7.57	0.893	531.3	19.34	0.243
	MLR	370.7	13.05	0.618	559.6	20.37	0.297
	SVM	398.2	14.02	0.568	518.5	18.87	0.314
S2	RF	167.8	5.91	0.925	319.7	11.64	0.723
	MLR	309.7	10.90	0.733	309.5	11.26	0.732
	SVM	316.9	11.15	0.722	318.2	11.58	0.714
S3	RF	227.7	8.02	0.861	601.3	21.88	0.220
	MLR	398.6	14.03	0.558	546.6	19.89	0.268
	SVM	404.0.	14.22	0.558	577.0	21.00	0.268
S4	RF	247.4	8.71	0.842	543.0	19.76	0.227
	MLR	503.8	17.74	0.293	477.0	17.36	0.367
	SVM	529.0	18.62	0.293	525.4	19.12	0.367

3.6. Comparisons of Predicted and Measured Yield

We compared the yield predicted through machine learning and the measured values, and the results are shown in Figure 10 and

Table 4. Data characteristics of the predicted and measured yield.

Stage		Median (kg/ha)	Average (kg/ha)	Max (kg/ha)	Min (kg/ha)	Range	Skewness	Kurtosis
	Measured	2901.3	2747.7	3701.9	1215.0	2486.8	−0.53	0.20
S1	RF	2950.9	2856.9	3313.8	1695.9	1617.9	−2.00	4.11
	MLR	2928.4	2744.0	3304.4	643.6	2660.8	−2.23	5.59
	SVM	3001.7	2871.5	3310.4	1322.2	1988.2	−2.12	4.67
S2	RF	2825.2	2676.1	3318.4	1388.7	1929.8	−2.00	4.11
	MLR	2870.2	2695.4	3338.1	1470.1	1868.0	−2.23	5.59
	SVM	2885.1	2718.4	3279.0	1591.5	1687.5	−2.12	4.67
S3	RF	2856.4	2704.8	3603.4	1477.4	2125.9	−2.00	4.11
	MLR	2779.7	2703.3	3822.5	1658.9	2163.6	−2.23	5.59
	SVM	2792.8	2705.8	3980.7	1516.0	2464.8	−2.12	4.67
S4	RF	2877.9	2761.6	3384.0	1458.5	1925.5	−2.00	4.11
	MLR	2842.7	2825.1	3401.8	2182.3	1219.5	−2.23	5.59
	SVM	2907.4	2897.8	3213.3	2546.2	667.1	−2.12	4.67

. The violin plots show that the median value of the predicted yield was near the measured value across all four stages. Except for MLR and SVM for S3, the maximum yield predicted across stages was smaller than the measured yield. Moreover, except for MLR for S1, all minimum yield values predicted through machine learning were higher than the measured values across all stages. In addition, except for MLR for S1, the yield range of values predicted through machine learning were smaller than the manually measured treatment. Finally, the skewness and kurtosis (i.e., absolute value) of values predicted through machine learning were higher than those of the measured values.

4. Discussion

In recent years, UAVs, as a near-ground-based platform to collect crop data, have been widely used in agriculture management. UAV-derived imagery data are geometrically registered, and they may have multiple sensors such as those used to collect spectral, thermal, radar, and fluorescence information. Previous studies proved that crop yield can be well predicted with UAV-collected data [36,37]. In this study, a hyperspectral imagery system was mounted on a UAV, and VIs were used to predict oilseed rape yield. R² of model training and testing reached 0.925 and 0.768, respectively. This suggests that UAVs may have a great ability to predict crop yield on a large scale.

VIs, including NDVI, are often used for predicting crop yield. Some studies indicated that NDVI had a significant relationship with yield and that crop yield could be predicted using NDVI [13,38,39]. A similar result was seen in our study. In our experiment, the correlation coefficient between NDVI and oilseed rape yield was 0.896 at S2. However, Lukas et al. [40] reported that the R² of NDVI-based oilseed rape yield was very low (only up to 0.40). In our study, NDVI was not selected as a model input variable because it had a lower correlation with oilseed rape yield than RVI and CIgreen in PCA. Other VIs may be better than NDVI in oilseed rape yield prediction models. In 2022, Lukas et al. found that R² of BNDVI and NDYI for oilseed rape yield reached 0.95 and 0.90, respectively. In our study, R² of RVI+CIgreen with an RF-based oilseed rape yield model reached 0.925. This suggests that RVI+CIgreen has great potential for predicting crop yield.

We selected RVI, CIgreen, CIre, and their combinations as model input variables. The models during training had good performance (i.e., high R² and low RMSE and RRMSE). However, the testing performance was not excellent. The reason for this may be that VI data in actual field conditions are often disturbed by background noise, image shadows, and other factors. Some researchers reported that VIs combined with crop height, leaf area index, and morphological and physiological data can improve the yield model accuracy [41,42,43,44]. Which VIs and what combinations with other data should be used to help build a better yield model need to be further investigated.

Generally, hyperspectral imagery with rich spatial and spectral information can comprise hundreds of VIs or more. These VIs may have auto-collinearity problems. This makes subsequent data analysis complex and difficult. Feature screening to reduce the number of data and increase data quality is necessary for building models. To accomplish this, we employed Pearson correlations and PCA, and we obtained three features (VIs) to keep in our models. Feature screening methods have also been used by Taşan et al. [27] for predicting eggplant yield. Other data dimensionality reduction methods have been employed by many authors [44,45,46]. Appropriate feature selection methods for predicting crop yield models require future research.

Machine learning has been widely employed in agriculture management in recent years. The RF, MLR, and SVM are the main bases for model construction. Their suitability in predicting oilseed rape yield was demonstrated in the present work. We found that, during training, the RF-based model had the highest R² and lowest RMSE and RRMSE across all four stages: 0.925, 167.8 kg/ha, and 5.91%, respectively. A similar result was obtained in a study estimating pasture aboveground biomass by Freitas et al. [46]. During testing, the MLR-based model had the best performance, with its R², RMSE, and RRMSE at 0.732, 309.5 kg/ha, and 11.26%, respectively. The RF-based models were more successful than other models in analyzing the complex relationships between oilseed rape yield and spectral indices. We also note that the SVR model had the lowest performance among the four investigated models.

We investigated oilseed rape growth from S1 to S4 for yield prediction. The S2 stage was determined as the best growth stage to predict yield. The earlier (S1) and latter stages (S3 and S4) may be unhelpful for predicting yield. However, we note that S2 is a growth stage and not a specific time point. Obtaining the best date for predicting yield with the best accuracy needs further study.

The data collection and model application need to follow certain standards. Our data include UAV-based imagery and ground-based measurements alongside various modeling approaches. Our data may thus encounter errors from uncontrollable factors. Therefore, standardized, unified, clear, and controlled procedures must be carefully developed and integrated into an intelligent and automated system. Uncertainty, transferability, and further application of the developed models should be assessed in future work.

There are some limitations to our research. First, we hypothesized that oilseed rape yield was determined only by spectral information. Crop yield prediction can be explained in part by spectral information. This may be the reason that our testing model had relatively weak performance. Second, in our study, spectral reflectance was extracted by averaging all the pixel reflectance of the sampling area without considering each pixel in the spectral image. Calculating the reflectance pixel by pixel may be a better approach [47]. Finally, our research result was based on a one-year field experiment. Two or more years of experiments may help in building a stable and universal yield prediction model.

5. Conclusions

In recent years, studies on the prediction of crop yield at various growth stages using spectral VIs captured by UAVs and machine learning methods have been scarce. Therefore, the current research investigated the impact of growth stage on the performance of oilseed rape yield models, compared the efficacy of four distinct machine learning techniques, and analyzed the influence of different input strategies on model performance.

The key findings of this investigation are as follows: (1) oilseed rape yield can be accurately predicted using spectral VIs obtained from UAVs and machine learning methods, achieving an excellent model performance, with an R² value of 0.925 and an RRMSE of 5.91%. (2) Among the four studied growth stages, the budding stage (S2) is the most optimal stage for predicting oilseed rape yield. (3) Among the four machine learning algorithms, random forest emerged as the most effective method for predicting oilseed rape yield. (4) The combination of RVI and CIgreen served as the optimal input VIs for the oilseed rape yield model.

Looking ahead to the practical application of this yield model, several critical challenges remain to be addressed. Notably, the accurate identification of specific days and time points, rather than a broad growth stage, for accurate yield prediction demands further investigation. Additionally, rigorous evaluations of the uncertainty, transferability, and wider applicability of the developed yield models are imperative. In summary, the integration of spectral sensors with UAVs presents immense potential for large-scale crop yield forecasting. Moreover, machine learning approaches, particularly RF, significantly enhance the accuracy of yield models, surpassing traditional methods such as MLR. This study contributes a relatively precise prediction model for crop yield, offering profound insights that can drive advancements in field crop management practices.