Utilisation of unmanned aerial vehicle imagery to assess growth parameters in mungbean (Vigna radiata (L.) Wilczek)

Study site and experimental design

Three experiments were conducted at the University of Queensland, Gatton Campus (27.55°S, 152.34°E), to establish mungbean crops with variable growth and biomass/yield dynamics. Weeds, pests and diseases were managed as required. Experiments were sown in January 2018, October and November 2019.

Experiment 1 was conducted from 16 January 2018 (sowing date) to 9 April 2018 on a 2-ha plot comprising five replicates of two mungbean genotypes (Jade-AU and Satin II). Additionally, a zinc (Zn) fertiliser was applied with seeds as a starter, at a rate of 20–30 kg/ha. This experiment was conducted under conditions of high-water availability with 100 mm of irrigation at sowing, and a further 50 mm of irrigation after emergence due to a high-temperature period. No additional water was applied after this as adequate rainfall of 263 mm occurred. Row spacing was 40 cm with six rows per pass.

Experiments 2 and 3 were respectively conducted from 22 October 2019 (sowing date) to 2 February 2020 and 28 November 2019 (sowing date) to March 2020. Two mungbean cultivars (Jade-AU and Crystal) were grown in four replicates under irrigated conditions. A Zn fertiliser was applied with seeds as a starter, at a rate of 20–30 kg/ha. Each plot size was 13 m long and 4.35 m wide. A total of 381.4 mm and 333.1 mm of irrigation were applied in Experiments 2 and 3 across the season. These two experiments were planted on 50-cm row spacing with eight rows per plot.

Data collection

Meteorological data, including maximum and minimum temperature, total rainfall, solar radiation and relative humidity, were obtained from a weather station for Experiment 1. The weather station was located at the University of Queensland in Gatton, which was 1 km north of the field experiment. The data was downloaded through the Bureau of Meteorology website. For Experiments 2 and 3, similar weather data were collected from a HOBO® RX3000 logger (HOBOlink®, Australia) weather station installed within the field trial area.

Four to six biomass harvests were taken in the experiments. In Experiment 1, four inner rows along a 1-m transect (1.2 m²) inside the plot area were sampled from 10 plots (two genotypes, five replicates). In Experiments 2 and 3, a sampling area of 1 m² within three inner rows per plot was used to collect data across the experiments in eight plots (two genotypes and four replicates). Table 1 presents the dates for manual harvesting and UAV data collection at different growth stages for Experiments 1, 2 and 3. Growing degree days (GDD) were calculated based on the difference between the daily mean temperature and the base temperature (Arnold 1959), using a base temperature of 7.5°C (Robertson et al. 2002; Rachaputi et al. 2015; Chauhan and Williams 2018). When mungbean plants reached 50% flowering during all three experiments, the date of flowering was recorded.

The total number of plants and leaf fresh weight of each sampling plot were recorded. The leaf area of each sample was obtained from a LI-3100C (LI-COR) leaf area meter. Subsequently, all leaves and stems were oven-dried for 6 days in paper bags at 65°C and their dry weights were recorded. From flowering onwards, leaves and stems, as well as reproductive organs (flowers, green and black pods), were split into components, weighed and dried separately. At the final harvest, pod counts and seed weight per plot were recorded.

Light interception measurements were collected using an AccuPAR LP-80 ceptometer (Meter Environment®, USA). Measurements prior to every biomass harvest in experiment plots were performed under the full sun between 11:00 hours and 13:00 hours in six different locations under the canopy with four measures of incoming radiation above. The locations were fixed per plot to calculate accumulated light intercepted as the same canopy structure expanded over time, corresponding to the biomass and leaf area measurements.

All the UAV images were collected to match with the ground truth data measurements undertaken in the vegetative and reproductive stages. Therefore, weekly or fortnightly flights were launched during the growing seasons at 15 m altitude in Experiment 1 and 20 m in Experiments 2 and 3 over the field trials. These flights were conducted under low wind speed conditions and clear sky from 11:00 hours to 14:00 hours. The UAV platform used to collect images was a DJI Phantom 4 Pro. This UAV has a 20 Megapixel RGB camera that captures light in the visible spectrum (400–700 nm) in red, green and blue wavebands. To ensure accurate geolocation of the images, at least four ground control points (GCPs) were placed in the field prior to capturing the first set of images. The GPS coordinates of these targets were obtained using a GPS survey kit (Aeropoints https://www.propelleraero.com/aeropoints/), which allowed for precise location determination of the trials. The reference of GCPs ensures accurate geolocation of the images.

Ground truth data pre-processing

Image data pre-processing

RGB image data were processed using the PhenoCopter tool (https://phenocopter.netlify.app/) and commercial software Pix4D (https://www.pix4d.com/) to produce orthomosaic images. We then applied plot segmentation, binary image classification and extraction of phenotype values: FVC and vegetation indices per plot for each UAV flight. Although PhenoCopter is no longer accessible, this pipeline can be replicated using Pix4D and GIS tools such as QGIS to create a matrix of plots to extract the data. Image classification is also available using EasyPCC (https://www.quantitative-plant.org/software/easypcc) directly (Guo et al. 2017).

Production of orthomosaic images

The set of 80% overlapped images for these three experiments collected in each flight was imported directly to the PhenoCopter website to generate an orthomosaic image. The raw images were stitched together to develop the orthomosaic images for each flight in different mungbean growth stages, as shown as an example in Fig. 1a.

Fig. 1.

Orthomosaic image and plot segmentation. (a) Orthomosaic image of mungbean field on 14 February 2018 in Experiment 1 (left) and 17 January 2020 in Experiments 2 and 3 (right); (b) plot segmentation in the field trial for UAV flight during Experiments 1 (left), 2 and 3 (right).

Data analysis

Customised Python code scripts (Python ver. 3.10.9) were used to analyse the UAV and ground truth data. A two-sample t-test at a 5% significance level was calculated to compare the performance of the varieties for each experiment at the observation date. A test for homogeneity of variances and normality of residues was performed to validate these analyses. In addition, to validate the use of UAV imagery from RGB cameras, regression analysis and Pearson correlation tests were used to compare ground data (biomass, leaf area and fractional light interception) and UAV imagery (vegetation indices and FVC) for the three mungbean cultivars and each sowing period. After preliminary analysis showed no cultivar and seasonal differences, the data from three cultivars from three experiments were combined to represent the mungbean features to analyse the relationship between UAV imagery and ground data during pre- and post-flowering stages. The UAV analysis grouped the data into two different subgroups (pre-flowering and post-flowering) based on the flowering dates across the three experiments.

For Experiment 1, only three UAV flights (1, 2 and 3) were used to match with three sets of field sampling from the same period due to technical issues and data losses of the later three flights. However, these three early flights were sufficient to perform the correlation analysis between field data and UAV imagery and predict biomass at the pre-flowering stage. As a result, UAV data collected in Flights 1, 2 and 3 in Experiment 1 and Flights 1 and 2 in Experiment 2 (Table 1) was considered pre-flowering data. UAV data from Flights 5 and 6 in Experiment 2 and Flights 2, 3 and 4 in Experiment 3 were considered post-flowering data. Other flights in Experiments 2 and 3 were excluded due to the discrepancy between the field collection and drone flight dates.

Trial weather data across Experiments 1–3

The relationship between growing degree days (GDD) and the average maximum and minimum temperatures, as well as rainfall in mungbean Experiments 1, 2 and 3 are depicted in Fig. 2a. The average maximum and minimum temperatures for Experiment 1 were 31.8°C and 17.6°C, respectively. In Experiment 1, large temperature fluctuations with higher rainfall were apparent pre-anthesis (690 GDD, 20% flowering) but were more stable post-anthesis. The highest maximum daily temperatures and rainfall occurred towards the end of the vegetative stages.

Fig. 2.

Weather data for Experiments 1, 2 and 3. (a) Average maximum and minimum temperature (°C) and rainfall (mm); (b) average measured relative humidity (%) and solar radiation (MJ/m²).

In Experiment 2, the average minimum and maximum temperatures were 17.7°C and 35°C, respectively. In Experiment 3, the average minimum and maximum temperatures were 20°C and 34.37°C, respectively. In Experiment 2, the highest temperature was observed during the flowering period, whereas in Experiment 3, it occurred during the early growth stages. During Experiment 2, higher temperatures with relatively higher rainfall were seen during the flowering period (766 GDD, 50% flowering). In Experiment 3, higher temperatures and lower rainfall were observed in the period around the flowering (752 GDD, 50% flowering).

In-crop rainfall was relatively low across the three experiments. Total rainfall of 263 mm, 93 mm and 192 mm was observed in Experiments 1, 2 and 3, respectively. Most of the in-season rainfall occurred before flowering in Experiment 1 and after flowering in Experiments 2 and 3 (Fig. 2a). Relative humidity (RH) varied across the seasons and experiments. Experiment 1 had consistently high RH (49.5–91.5%) across the season, whereas Experiments 2 and 3 had increasing RH values as the season progressed (Fig. 2b). The solar radiation (MJ/m²) among the experiments showed a relatively stable trend with an average of 19.7 MJ/m², 24.1 MJ/m² and 21.73 MJ/m² for Experiments 1, 2 and 3, respectively (Fig. 2b). The average of solar radiation in Experiment 1 was lower than in Experiments 2 and 3.

Ground truth data

LAI

There were no significant differences (Experiment 1, P = 0.649; Experiment 2, P = 0.178; Experiment 3, P = 0.317) over time between the varieties used in Experiments 1, 2 and 3. As a result, two varieties were combined to present LAI for each experiment. For Experiment 1, LAI rapidly increased until canopy closure (around 619 GDD), then slowly rose and finally decreased as leaf senescence began. In Experiments 2 and 3, LAI increased to a peak at the pod and grain development stage (around 1319 GDD and 1259 GDD, respectively). During pod development, accelerating leaf senescence increased biomass partitioning to developing pods, thereby reducing canopy development. Maximum LAI for the three experiments were observed at the grain-filling stage as shown in Fig. 3a.

Fig. 3.

Growth parameters: leaf area index, fractional light interception and total biomass for Experiments 1, 2 and 3. (a) Leaf area index (LAI); (b) total biomass (g/m²); (c) fractional light interception (%).

Radiation extinction coefficient (k)

The radiation extinction coefficient for the different mungbean varieties was computed for each experiment, using a linear relationship between the −ln(1−fractional light interception) and LAI and forcing this relationship to pass through 0 (Fig. 4a).

Fig. 4.

Growth parameters: Extinction coefficient values (k) and radiation use efficiency for Experiments 1, 2 and 3. (a) Extinction coefficient values (k) from the relationship between −ln (1−fractional light interception) and LAI for three different mungbean varieties (Jade-AU, Satin II and Crystal); (b) radiation use efficiency (RUE).

Statistically significant differences between varieties for k were only observed in Experiment 3 (P = 0.027). In Experiment 3, k values of 0.66 and 0.48 (R² = 0.88) were calculated for Jade-AU and Crystal, respectively. During Experiment 1, where LAI values were relatively higher, k values for Jade-AU and Satin II were 0.72 and 0.68 (R² = 0.93), respectively. In Experiment 2, Jade-AU and Crystal had k values of 0.45 and 0.41 (R² = 0.73), respectively.

UAV imagery data

All three experiments’ data were combined to analyse the relationship between UAV imagery and ground data, because no significant differences existed in LAI, biomass and fractional light interception between cultivars across the season. However, ground truth data was split into pre- and post-flowering data as the relationships between biomass development and light interception changed significantly in the post-flowering period. As a result, UAV data analysis was also divided into two different subgroups (pre-flowering and post-flowering) across the three experiments.

Relationship between total biomass, FVC and fractional light interception at pre-and post-flowering stages

Pre-flowering there was a strong linear relationship between fractional light interception and FVC (R² = 0.90), but post-flowering this relationship was weak (R² = 0.09) in Fig. 5a. There was also a relatively strong relationship (R² = 0.69) between total biomass and FVC at pre-flowering stages, which disappeared at post-flowering stages (R² < 0.01), seen in Fig. 5b.

Fig. 5.

Relationship between fractional light interception, total biomass, and fractional vegetation cover for pre-flowering and post-flowering stages. (a) Relationship between the fractional light interception (%) and fractional vegetation cover; (b) relationship between total biomass (g/m²) and fractional vegetation cover (%).

Growth parameters under different sowing times

This study found no significant differences in leaf area index (LAI), biomass, fractional light interception, radiation extinction coefficient (k) or radiation use efficiency (RUE) among three genotypes: Jade-AU, Satin II and Crystal. Given that the three varieties have similar morphological characteristics, including a determinate growth pattern and similar plant architecture, with flower buds differentiating at the apex of the main stems or branches (Chauhan and Williams 2018), this may not be surprising. The lack of differences in the measured traits could be due to the similarities in the photosynthetic capacities, environmental responses and resource utilisation of the three varieties (Geetika et al. 2022a). Jade-AU, Satin II and Crystal, which account for over 90% of the country’s mungbean seed production, all have large green seeds (Chauhan and Williams 2018). Geetika et al. (2022a) reported similar findings, showing that there was no significant difference in radiation extinction (k) or RUE between the genotypes Jade-AU, Opal-AU and Satin-II in two field trials. Opal-AU, like the mungbean varieties used in this study, is also a large-seeded shiny green genotype (Australian Mungbean Association 2020).

Although Experiments 2 and 3 received additional irrigation after sowing, the canopy structure of Experiment 1 was substantially more vigorous and larger throughout the growing season in terms of LAI, fractional light interception and biomass, possibly due to the high in-season rainfall during the vegetative stage in Experiment 1. In Experiments 2 and 3, the rainfall mainly occurred during the post-flowering stages when the canopy growth rate had started to decrease. This led to variations in canopy development among the experiments.

Biomass accumulation is determined by LAI, radiation extinction coefficient (k), the duration of fractional light interception, and RUE (Tesfaye et al. 2006; Chauhan and Williams 2018). Biomass in legumes is generally positively correlated with LAI under conditions with no water stress; LAI is reduced dramatically when the plants are stressed with high temperatures and water stress during the vegetative stages (Tesfaye et al. 2006). Extremely high temperatures (above 40°C) and direct effects of heat during flowering (Experiment 2) and pod stages (Experiment 3) may have affected biomass accumulation of Jade-AU and Crystal. Patriyawaty et al. (2018a) found that high temperatures significantly reduced dry matter production and grain yield of mungbean cultivars Jade-AU and Crystal. High temperatures impact floral abortion rates, decrease pollen viability and reduce biomass accumulation (Patriyawaty et al. 2018b; Rachaputi et al. 2019). Temperature fluctuations and unpredictable rainfall have been observed to negatively affect mungbean biomass. This is consistent with previous research in Geetika et al. (2022b), which has shown that extreme temperatures and moisture stress during the growing season can reduce mungbean growth and yield. For example, high temperatures during the flowering stage can reduce pod set and seed filling, and prolonged drought can reduce leaf area and photosynthesis, leading to reduced biomass accumulation.

Sowing time also affects biomass. As observed in this study, the optimal time to achieve maximum mungbean biomass is the summer season (around January) due to the favourability of warmer weather at the vegetative stages (Robertson et al. 2000). The highest biomass accumulation occurred in Experiment 1 planted in February 2018, whereas biomass was reduced in Experiments 2 and 3 sown in October and mid-November 2019, respectively. High levels of in-season rainfall in Experiment 1 also contributed to increased biomass.

Higher LAI results in higher fractional light interception, which enhances crop growth rate and consequently leads to higher biomass (Yousefi et al. 2021). Greater LAI can contribute to higher biomass in mungbean, but it is not the only determining factor. Experiment 1 had much higher LAI and fractional light interception than Experiments 2 and 3, which likely contributed to its higher biomass accumulation. Agronomic management practices such as row spacing and plant density have been recognised as crucial strategies that can significantly impact the canopy radiation environment under favourable growing conditions (Geetika et al. 2022a).

Experiment 1 showed higher estimated k and RUE compared to Experiments 2 and 3, possibly due to temperature–radiation imbalances and higher temperatures during mungbean’s reproductive stages. According to Rachaputi et al. (2015), a decrease in k can result in a reduction in fractional light interception per unit of LAI and dry matter accumulation. The RUE values of the three cultivars in Experiments 1 and 2 were consistent with earlier studies on mungbean, which reported baseline RUE values of 0.8–1 g/MJ in well-watered environments (Muchow et al. 1993; Rachaputi et al. 2015; Chauhan and Williams 2018). Experiment 1, which was conducted during the summer season of 2018, was associated with the highest RUE, of up to 1.3 g/MJ. This value is consistent with the findings of Geetika et al. (2022a), who reported a RUE of 1.3 g/MJ for the same genotypes (Jade-AU and Satin II) in field trials conducted in the summer season of 2019 and 2020. Variations observed among the three experiments could be attributed to fluctuations in temperature–radiation equilibrium, particularly during the reproductive stages of mungbean growth (Geetika et al. 2022b).

Relationship between UAV imagery and ground data on mungbean biomass prediction

FVC has been utilised to monitor vegetation conditions and estimate biomass from UAV-based RGB imagery (Yan et al. 2019; Yue et al. 2019). In this study, FVC from the UAV image data strongly correlated with the fractional light interception (R² = 0.90) during the pre-flowering stage. This high level of correlation means that using FVC derived from UAV data provides a valid alternative to more detailed ground-based, time and labour-intensive ceptometer measurements of light interception (Steven et al. 1986; Gitelson 2013). This study shows that UAV-based measurements can provide an alternative approach to estimate RUE using FVC, which is strongly correlated with biomass but is not equivalent to biomass itself.

Crop canopy structure, field background, separation and classification in the process of orthomosaic images can affect FVC computation in later stages (Bendig et al. 2015; Wan et al. 2021). Factors such as shading, increased canopy density with overlapping leaf layers, cessation of vegetative growth and later leaf senescence are likely to influence the accuracy of fractional light interception and consequently, the quantification of FVC at post-anthesis stages in all three experiments.

At the pre-anthesis stages, we observed that the relationships between FVC and fractional light interception, as well as between FVC and biomass, were comparable to the strong correlation between UAV-derived vegetation indices and biomass or LAI. Specifically, all the vegetation indices showed a strong relationship (all R² > ~0.80, except CIVE) at pre-flowering stages. These relationships all became poor (all R² < 0.41) at post-flowering stages (Table 3). In particular, biomass and LAI had the most robust relationship with NGRDI in the pre-flowering stage. Based on these results, NGRDI can accurately predict growth parameters of mungbean before flowering. NGRDI is considered to be a greenness index representing the combined effect of canopy architecture, leaf area and chlorophyll concentration that is strongly linked with biomass (Torres-Sánchez et al. 2014). Additionally, Jannoura et al. (2015) showed a relatively good relationship (R² = 0.58–0.78) between the above-ground biomass and NGRDI in three different crops/combinations of crops (sole peas, intercropped peas/oats, and sole oats).

In this study, a significant correlation was also found between biomass, LAI, and the vegetation indices GLI, ExG and VARI during the pre-flowering stages. High positive GLI values indicate the presence of healthy green leaves and stems, which are useful for monitoring vegetation growth and assessing vegetation health (Louhaichi et al. 2001). The advantage of ExG is that it can provide a clear contrast between plants and soil and produce near binary images, whilst VARI can reduce the sensitivity to atmospheric factors to estimate biomass and LAI (Woebbecke et al. 1995; Gitelson et al. 2002). Marcial-Pablo et al. (2019) compared vegetation indices from the visible spectrum (ExG, CIVE and NGRDI) and other vegetation indices, such as normalised difference vegetation index (NDVI), green NDVI and normalised green from NIR sensors in maize by separating vegetation and bare soil as accurately as possible. They demonstrated high accuracy of ExG to monitor FVC at earlier growth stages, whereas later growing stages would require using a near-infrared sensor. (Kataoka et al. 2003) claimed that CIVE could separate vegetation from soil background to evaluate crop growing status and have a high correlation with growth parameters in sugar beet and soybean. However, CIVE in this study had the lowest correlation with biomass and LAI compared to other vegetation indices during the pre-flowering period.

Other studies have also observed weak correlations between biomass/LAI and vegetation indices at later growth stages, possibly to due saturation at high LAI and biomass. Hunt et al. (2005) argued that dry biomass from zero to 120 g/m² (early stage: before canopy closure) was linearly correlated with NGRDI obtained from digital cameras in soybeans, but beyond this point, NGRDI saturated and did not show any further increase when soybean biomass increased from 150 to 600 g/m² (later stage). Motohka et al. (2010) noticed significant saturation happened in the middle of the growing period using NDVI to monitor biomass, although it usually had a marginally better cross-validation result than the NGRDI index. According to Yue et al.(2019), the estimates of biomass by using vegetation indices based on UAV data from RGB camera presented high accuracy at an early stage compared to later crop growth stages in wheat, since saturation and no unchanged cover are seen in the later growth stages. Across all crops, including mungbean in this study, these relationships are strong pre-flowering and weak post-flowering as canopy density increases.

Another potential reason for the better biomass prediction models originating from FVC and NGRDI (highest correlation with biomass and LAI) during pre-flowering stages may relate to canopy morphological structure. Mungbean plants grow tightly together and overlap each other after flowering in the field, although row spacing was applied within the range recommended for this crop. This made accurately predicting biomass more difficult as the season progressed because UAV images were captured above the ground, and overlapped biomass in image pixels cannot represent real biomass at later stages in the field.

The study on the relationship between UAV imagery and ground data on mungbean biomass prediction found a strong correlation between the two methods at pre-flowering stages. The research team was able to accurately and efficiently measure growth parameters and biomass estimation using UAV imagery. These findings suggest that UAV imagery can be a reliable and cost-effective method for measuring crop growth parameters, particularly in crops like mungbean where ground-based measurements are challenging. The study’s implications for future research in mungbean and other crops include the potential for wider adoption of UAV technology in crop monitoring and management. Additionally, the study highlights the need for further research to optimise the use of UAV imagery for crop monitoring and to better understand the relationship between UAV and ground-based data. Other modified vegetation indices from high-resolution RGB cameras or combining an RGB with NIR sensor to assess biomass in the mungbean’s later stages may need to be considered. More studies with UAV-based RGB imagery in legumes, particularly mungbean, are encouraged by using FVC through different crop growth stages, locations including glasshouse and fields, genotypes and experimental designs in future years.