1. Introduction
Wheat (
Triticum aestivum L.), as the third largest cereal crop in the world, plays an important role in world food production and food security strategies. According to the Food and Agriculture Organization of the United Nations (FAO), more than 220 million ha are sown with wheat, with over 770 million tons of wheat being produced in 2021 [
1]. China accounts for over 23 million ha of wheat crop and is responsible for nearly 18% (about 137 million tons) of all wheat produced worldwide [
1]. Wheat is a major food source for the people in North China. The North China Plain is one of China’s primary wheat-growing areas (mainly winter wheat). However, in this region, winter wheat becomes more vulnerable to drought stress due to the continental monsoon climate, leading to decreased output. Groundwater irrigation is one main source of water supply during the winter wheat growth season of the North China Plain, but it is severely restricted according to the sustainable development policy. Thus, in order to fulfill the dual purpose of food security protection and water conservation, it has become necessary to breed drought-resistant varieties of winter wheat to maintain production. As the ultimate detection target, the crop yield becomes an important selection parameter of winter wheat breeding. However, in the empirical breeding process, yield can only be measured after the crop growth cycle, which is costly, labor-intensive, and time-consuming for the breeding researchers. Therefore, establishing a method that could estimate winter wheat yield on a field plot scale within a short time would help them in selecting drought-resistant varieties.
Satellite remote sensing data have been widely applied for non-destructive crop yield estimation over various large-scale regions (from local to national, continental, and global) since the 1970s [
2,
3]. The yield estimation models based on these data have demonstrated reasonable crop yield estimation accuracy in the large-scale regions based on satellite imagery and have been widely used due to their convenience and simplicity [
2,
3,
4,
5,
6,
7,
8]. However, for breeding researchers, the application of satellite data in yield estimation is often hampered by the high spatial heterogeneity within the small areas of breeding fields, missing data during critical crop growth stages, and high costs due to the coarse spatial resolution and fixed passing time and band setting.
In recent years, the development of sensor technologies in unmanned aerial vehicles (UAVs) has promoted their application for data acquisition [
9]. Compared to satellite remote sensing, UAV remote sensing has an improved spatial, spectral, and temporal resolution and is associated with lower costs and greater flexibility and versatility [
10], making UAVs increasingly popular in precision agriculture [
11,
12]. Previous studies have reported the relationship between crop yield and crop phenotypic parameters such as plant height, leaf nitrogen content, leaf area index, above-ground biomass, and so on [
13,
14,
15,
16]. UAV platforms have exerted a beneficial effect in the retrieval of a wide array of crop characteristics that are associated with yield [
17,
18,
19,
20], and their use can help meet breeding researchers’ requirements regarding crop yield estimation on a small plot area scale within a short amount of time.
Furthermore, vegetation indices (VIs) derived from UAV-mounted multispectral (MS) and hyperspectral (HS) sensors have been widely used to estimate crop yield [
21,
22]. Duan et al. developed a method based on VIs that were derived from UAV multispectral data to correlate with rice phenotyping and estimate grain yield [
13]. García-Martínez et al. estimated corn grain yield by combining vegetation indices, canopy cover, and plant density using multispectral and RGB images acquired via the use of UAVs [
23]. Ramos et al. proposed a random forest algorithm that performed well in tests with a ranking-based strategy that focused on predicting maize crop yield using UAV-based multispectral vegetation indices [
24]. However, the application of UAV-mounted hyperspectral sensors for agricultural monitoring is limited by the weight of imaging systems and the complexity of image processing [
25,
26]. As a result, they have rarely employed by breeding researchers in the study of crop breeding. Moreover, the VIs used in previous studies have been mostly derived from limited growth stages within the crop growing season, which may increase the risk of missing critical spectral features in other growth stages [
27,
28]. Therefore, estimating crop yield through combining UAV-mounted HS sensor-based VIs in several critical crop growing stages could be helpful to integrate critical spectral features throughout the crop growing season and lead to a higher estimation accuracy, which would be beneficial to breeding researchers.
For this study, motivated by the need to boost the efficiency of winter wheat breeding, we aimed to evaluate yield estimation among various winter wheat breeding cultivars using multi-temporal vegetation indices derived from UAV-mounted multispectral and hyperspectral sensors. To accomplish this objective, we (1) investigated the correlation between the winter wheat yield and 19 UAV-based multispectral VIs and 17 band combination types of hyperspectral data at single growth stages and multiple growth stages, respectively, and (2) identified the best VI and the best time for estimating winter wheat yield using UAV data.
2. Materials and Methods
2.1. Experimental Setup
The experimental site was set up at the Dry-Land Farming Institute of Hebei Academy of Agricultural and Forestry Sciences (DFI) at Hengshui City, Hebei Province, China (37°54′15.63″N, 115°42′29.32″E, World Geodetic System 1984) (
Figure 1). The area has a semi-arid temperate and monsoonal climate characterized by four distinct seasons, with an average yearly temperature of 13.3 °C and a yearly precipitation of 497.1 mm.
The experimental site design included eleven winter wheat cultivars: C1 (Chang8744), C2 (Shimai22), C3 (Luyuan472), C4 (Shimai15), C5 (HengH1603), C6 (Xinmai28), C7 (Jimai418), C8 (Shannong28), C9 (Nongda212), C10 (Heng4399), and C11 (Jimai22). Each cultivar was then split into seven different irrigation groups and three repeats (1.5 m × 6 m in size) according to a randomized block design. All cultivars were planted on 15 October 2020 at a density of 375 plants/m
2. A base fertilizer, pure nitrogenous fertilizer (225 kg/ha), P
2O
5 (112.5 kg/ha), and K
2O (112.5 kg/ha) were applied before sowing. No additional fertilizers were used for the growth of the winter wheat discussed in this study. The irrigation date of each irrigated sub-plot is shown in
Table 1. The irrigation volume of each time was 750 m
3/ha. The total precipitation during the 2020–2021 growing season in the site was 43.9 mm.
2.2. Data Acquisition
2.2.1. Ground Truth Data
All winter wheat cultivars were harvested on 8 July 2021. The yield of each sub-plot was weighted and normalized to a moisture content of 13% and is expressed as ‘t/ha’. According to the experimental design, 231 samples were measured, while 17 measurements were eliminated as outliers, including 6 plots of Repeat 1, 4 plots of Repeat 2, and 7 plots of Repeat 3. The statistics of measured winter wheat yield are outlined in
Table 2. The mean measured grain yield values under different irrigation groups and winter wheat cultivars are shown in
Figure 2. Winter wheat yield differences across different irrigation groups and different cultivars were assessed using a one-way analysis of variance after checking the normality assumption at 0.05 probability level (
Table 3). There were significant differences among the different irrigation groups and different winter wheat cultivars. The grain yield of irrigation group B was lower than other irrigation groups, and group A had the highest yield. The C6 cultivar showed the poorest yield, while C9 performed best.
2.2.2. Multi-Sensor UAV Data
UAV images derived from multispectral (MS) and hyperspectral (HS) sensors were employed in this study (
Figure 3). The UAV campaign was conducted under low wind speed and clear sky conditions between 10:00 a.m. and 2:00 p.m. local time to reduce the influence of atmospheric and solar radiation. The overlap percentages in the forward and lateral flying directions of both UAVs were 80% and 70%, respectively. The acquisition dates and details corresponding to the growth stages regarding both UAVs are shown in
Table 4.
The DJI P4 Multispectral (DJI Technology Co., Ltd., Shenzhen, China) was used to collect multispectral images, including 5 sensors of blue, green, red, red-edge, and near-infrared with wavelengths of 456 nm (±16 nm), 560 nm (±16 nm), 650 nm (±16 nm), 730 nm (±16 nm), and 840 nm (±26 nm), respectively. The sensors used a 1/2.9 inch complementary metal-oxide-semiconductor (CMOS). The field of view was 62.7°, the focal length was 5.74 mm, the f-number was f/2.2, and focus was kept at the infinite point (∞). The MS images were obtained at a flight height of 50 m with an accuracy of flight altitude of 0.1 m, and the corresponding ground pixel resolution was 1.6 cm. The MS orthomosaic maps were generated using Pix4D mapper (Pix4D SA, Lausanne, Switzerland).
A DJI M600 Pro (DJI Technology Co., Ltd., Shenzhen, China) equipped with a Pika L hyperspectral camera (Resonon, Inc., Bozeman, MT, USA) was used to capture the hyperspectral images discussed in the study. The hyperspectral camera has 150 bands in a spectral range of 400–1000 nm with a spectral resolution of 4 nm. The HS images were acquired at a flight height of 50 m with an accuracy of flight attitude of 0.5 m, and the corresponding ground pixel resolution was 3.0 cm. SpectrononPro (Resonon, Inc., Bozeman, MT, USA) and ENVI 5.3 (Esri Inc., Redlands, CA, USA) were used to generate HS orthomosaic maps.
2.3. Vegetation Indices Calculation
The average of a 0.8 m × 4 m image area was used for band reflectance value calculation for each sub-plot. The area was approximately in the center of each sub-plot to eliminate the effect of the marginal areas of each sub-plot and other neighboring sub-plots. The extracted images are herein referred to as UAV images.
A large number vegetation indices have been proposed for grain yield estimation. A total of 19 MS VIs that were previously used for crop yield and yield-related phenotypic characteristic estimation were calculated, with average reflectance being derived from UAV MS images [
3,
18], and the 19 indices are shown in
Table 5.
Additionally, for various spectral bands of the UAV HS sensor, the HS VIs minimized the spectral redundance, which was usually found in the hyperspectral data and also promoted the computational optimization [
29,
30]. Therefore, 17 prevalent formulas, composed of two, three, or four spectral bands using function of sum, difference, ratio, double difference, normalized difference, and hybrid, were regarded as the HS VIs and employed in the study [
31]; the 17 formulas are shown in
Table 6. Band iteration was applied to all hyperspectral bands in each formulation.
Table 5.
Multispectral vegetation indices used in the study.
Table 5.
Multispectral vegetation indices used in the study.
Multispectral Vegetation Index | Formulation | Reference |
---|
Difference vegetation index | | [32] |
Ratio vegetation index | | [33] |
Green chlorophyll index | | [34] |
Red-edge chlorophyll index | | [34] |
Normalized difference vegetation index | | [35] |
Green normalized difference vegetation index | | [36] |
Green-red vegetation index | | [33] |
Green-blue vegetation index | | [37] |
Normalized difference red-edge | | [38] |
Normalized difference re-edge index | | [39] |
Simplified canopy chlorophyll content index | | [40] |
Enhanced vegetation index | | [41] |
Two-band enhanced vegetation index | | [42] |
Optimized soil adjusted vegetation index | | [43] |
Modified chlorophyll absorption in reflectance index | | [44] |
Transformed chlorophyll absorption in reflectance index | | [45] |
MCARI/OSAVI | MCARI/OSAVI | [44] |
TACRI/OSAVI | TACRI/OSAVI | [45] |
Wide dynamic range vegetation index | | [46] |
2.4. Yield Estimation Model
Two repeats (Repeat 1 and Repeat 2) were employed for all cultivars to construct the winter wheat yield estimation model from each vegetation index. Pearson’s correlation coefficient (
r) was used to demonstrate the yield estimation accuracy by comparing the yield measured in the field and estimated yield from the regression models below using Student’s
t-test at a 95% confidence level. After excluding the outliers described in
Section 2.2.1, 144 samples were used to establish the yield estimation model in the study.
The most commonly used remote-sensing based approaches for crop yield estimation involve the use of empirical statistical models, which demonstrate the relationship between yield and canopy spectrum characteristics in an intuitive way. Therefore, the simple linear regression function (SLR) was used to analyze the relationship between winter wheat yield and vegetation indices in individual growth stages (Equation (1)).
where
represents the winter wheat yield;
represents the vegetation index values at the booting, heading, flowering, filling, or maturation stages; and
and
are parameters calculated from a least-squares fitting method.
In addition, the multiple linear regression function (MLR), which was used to combine the vegetation indices among the multiple growth stages to estimate crop yield, is presented in Equation (2).
In this equation, represents the winter wheat yield; , , , , and represent the vegetation index values at the booting, heading, flowering, filling, and maturation stages, respectively; and , , , , , and are parameters calculated from a least-squares fitting method.
2.5. Validation of the Crop Yield Estimation Model
After eliminating outliers, 70 samples of Repeat 3 for all cultivars were used to validate the regression model using root mean square error (RMSE) and mean absolute percentage error (MAPE). The RMSE and MAPE equations are presented in Equations (3) and (4).
In these equations, is the crop yield measured in the field of sample i, is the crop yield predicted by the estimation models of sample i, and n is the number of valid samples.
4. Discussion
Currently, NDVI is the most widely used vegetation index for crop yield estimation. However, because of the characteristics of NDVI, it has significant saturation under a high vegetation coverage level, thereby affecting estimation accuracy [
47,
48,
49,
50]. According to the spectral reflectance characteristics of the plant, the absorption of chlorophyll on the red-edge waveband is weaker than that of red band, with the red-edge region having stronger transmission ability with the crop canopy [
51]. The use of a red-edge band rather than a red band in the vegetation index of NDVI can reduce the saturation phenomenon, improving crop yield estimation accuracy [
51,
52]. Additionally, the spectrum of each irrigation group for the winter wheat cultivator of ‘C9’ at the flowering stage is shown in
Figure 11. According to the figure, there were significant differences among different irrigation groups at the near-infrared band. Therefore, in this study, vegetation indices composed of the red-edge and near-infrared bands, both for MS imagery (NDRE) and HS imagery (‘
’), demonstrated reasonable robustness in terms of their winter wheat yield estimation.
Remote crop yield estimation methods are commonly based on the high correlation between the crop yield and the vegetation index taken at a specific crop growth stage [
50]. The success of the development of a vegetation index is dependent on the use of bands with different sensitivities to the key parameter that is to be monitored [
51]. The potential of multi-spectrum and hyper-spectrum for winter wheat yield estimation was systematically compared in this study. The lower yield estimation accuracy based on multispectral vegetation indices is mostly due to the obvious shortcoming of the limited fixed bands with a wide resolution. The hyperspectral sensor captured much richer information and is more sensitive to crop canopy characteristics with the continuous acquisition of reflectance at narrow wavelengths [
45,
52]. Additionally, for all vegetation indices calculated from the UAV-hyperspectral imagery used in this study, with the number of bands including in the vegetation index, the yield estimation accuracy of winter wheat assumed a rising tendency. According to the study of Thenkabail et al. [
53], four sensitive band combination-based optimum multiple narrow band reflectance models could explain up to 92% of the crop biophysical parameter variability. Hence, the four-band combination type based on hyperspectral imagery achieved a more significant correlation with winter wheat yield than the vegetation indices derived from multispectral imagery throughout the whole growth period of winter wheat in this study.
According to Qader et al., yield estimation models that use VIs from the crop’s critical growth stage can obtain a higher accuracy across remote sensing data [
54]. Our results regarding winter wheat yield estimation based on singular growth stages confirmed this and strongly indicated that the flowering stage was a critical period for winter wheat yield estimation. Some studies have shown that the accumulative vegetation index can improve the stability of yield estimation and that adding one or more growth stages (except the critical growth stage) could improve estimation accuracy [
28,
55,
56]. In this study, the correlation coefficient between crop yield and the hyperspectral vegetation index of ‘
’ increased from 0.80 to 0.84, and
RMSE decreased from 0.75 t/ha to 0.69 t/ha when the booting, heading, filling, and maturation stages were added with the MLR model in the flowering stage of the simple linear model.
Although the MLR model combining multiple temporal hyperspectral vegetation indices calculated according to hyperspectral imagery acquired good yield estimation accuracy for winter wheat, machine learning algorithms have demonstrated the potential to retrieve crop characteristics using multispectral satellite data, aerial multispectral data, aerial hyperspectral data, and so on [
57,
58,
59]. Therefore, future research should aim to explore machine learning regression models to strengthen crop estimation ability via multiple temporal UAV-derived hyperspectral datasets. Additionally, data fusion approaches which could integrate UAV data and satellite data-based vegetation index time-series curves together and improve temporal resolutions will also be investigated in the context of estimating crop yield in a future study.
5. Conclusions
This study assessed the accuracy of winter wheat yield estimation values based on UAV-derived multispectral and hyperspectral images from single and multiple growth stages. The results suggested that the proposed multiple linear regression model, constructed by the vegetation index of ‘’ with central wavelengths of 782 nm, 874 nm, 762 nm, 890 nm, which were calculated from UAV-based hyperspectral images using the growth stages from booting to maturation, can be used as a fast and reliable method for winter wheat yield estimation to contribute to the breeding of drought-resistant varieties of winter wheat in a field plot scale over a short amount of time. Moreover, the red-edge and near-infrared bands are recommended for use in the context of crop yield estimation.
From a longer-term perspective, more in-depth investigations into crop yield estimation (including rice, maize, soybean, and other crops) based on UAV-mounted hyperspectral datasets (via not only linear regression models but also machine learning algorithms, data assimilation, and so on) are expected.