Use of Unmanned Aerial Vehicles (UAVs) Imagery in Phenotyping of Bambara Groundnut

In this experiment, proximal measurements and Unmanned Aerial Vehicle (UAV) imagery was used to determine growth stages for bambara groundnut (Vigna subterranea (L.) Verdc.). The crop is a high potential crop due to its ability to yield in marginal environments, but neglected and underutilised due to lack of information on its growth in different environments. This study evaluated the correlation between Normalised Difference Vegetation Index (NDVI) derived from the ground as well as airborne sensors to test the ability of remotely sensed data to identify growth stages. NDVI and chlorophyll content of bambara groundnut leaves were measured at ground level at 18, 32, 46 and 88 days after planting (DAP) comprising vegetative, flowering, pod formation and maturity growth stages. The UAV imagery for the experimental plots was acquired with 0.2m resolution at maturity. The result showed a significant (p < 0.05) linear relationship between proximal NDVI and chlorophylls content at all growth stages of growth. The R varied from 0.57 in the vegetative stage to 0.78 in the flowering stage. Furthermore, NDVI derived from proximal measurements and UAV data showed a significant (p < 0.05) correlation. The observed high correlation between proximal sensors, UAV data and crop parameters suggest that remote sensing technologies can be used for rapid phenotyping to hasten the development of models to assess the performance of underutilised crops for food and nutrition security.


Introduction
Plant phenotypes are dynamic and the result of plant interactions with the environment. Understanding these activities in a constantly changing climate is important for the development of plant science, crop management and breeding of new varieties (Pieruschka & Schurr, 2019). The plant research community need to accurately measure the diverse characteristics of plants in order to understand their adaptation to resource-limiting environments. Bambara groundnut (Vigna subterranea (L.) Verdc.), is a legume crop that is commonly grown in low-input systems across sub-Saharan Africa and Southeast Asia (Mayes et al., 2019). In addition to having good nutritional characteristics, bambara groundnut is highly tolerant to drought and is able to yield on lands that are not fertile enough for the cultivation of many other crops. However, despite its potential, it remains underutilised due to lack of information on its performance in different environments and in particular, its phenotypic development. This limits the ability to assess its suitability for new locations (Suhairi et al., 2018).
Crop phenotype is the result of interaction between the genotypic and environmental factors. It comprises geometric traits such as height, leaf area index, canopy cover and spectral features and physical traits such as chlorophyll content, biomass and photosynthesis; nutrient contents and yield (Yang et al., 2017). Understanding how these traits change over time is one of the crucial steps in monitoring the crop growth. Contrasting these events with crop management events such as irrigation, fertiliser and pesticide application an essential source for understanding the crop conditions (Prasad et al., 2006).
Various technological approaches based on remotely sensed measurements have been proposed to assess these traits in the field condition (Yang et al., 2017). The commonly used trait for high-throughput screening and phenotyping is the Normalized Difference Vegetation Index (NDVI) derived from canopy reflectance. NDVI is measured using wavelengths within the near infrared (NIR) and visible (VIS) regions of the electromagnetic spectrum. The NDVI is associated to chlorophyll content in the leaf molecules that in turn is related to photosynthetic capacity of the plants. NDVI can be used to estimate the relative crop biomass at different crop developmental stages as well as nitrogen deficiency at crop senescence (Tattaris et al., 2016). There is enough scientific evidence to suggest that NDVI can be successfully used to estimate different crop traits (Jewan et al., 2019;Johnson, 2003;Wall et al., 2008). For example, Leaf Area Index (LAI), which is one of the most important indicators of crop growth has been indirectly estimated using NDVI for soybean, maize (Colombo et al., 2003;Johnson, 2003) and bambara groundnut (Jewan et al., 2019). The NDVI has also been used to forecast the yield of barley, canola, field peas and spring wheat (Mkhabela et al., 2011), bambara groundnut (Jewan et al., 2019), wheat (Wall et al., 2008) and maize (Shanahan et al., 2001). Therefore measurements of NDVI or its estimates can be used in yield assessment models (Prasad et al., 2006).
Phenotyping using ground-mounted vehicles can provide information about plant traits on a timescale of many hours for a plot. However, this method is time-consuming and is not practical for large scale and remotely located plots (Han-Ya et al., 2010). Using multiple sensors to take measurements concurrently for many plots may increase the costs (Candiago et al., 2015a;Gevaert et al., 2015). This has recently motivated the use of high-resolution data processing in phenotyping. In addition, field-based phenotyping to monitor the phenology and crop parameters for bambara groundnut landraces has recently been shown to be ineffective (Jewan et al., 2019). Determination of leaf chlorophyll content, which requires sampling from several locations in the leaf to obtain adequate characterisation (Candiago et al., 2015b) usually takes a long time to accomplish. This in turn, hinders the process of calibrating crop models, particularly for less-researched, neglected and underutilised species. Development of rapid NDVI estimation methods for crop parameters using remote sensing approaches can streamline modelling efforts for these crops.
Satellite remote sensing has also proven to be a valuable tool for monitoring crop health, crop modelling, climate change adaptation and mitigation and others (Cobb et al., 2013;Li et al., 2014). However, currently available satellite data are costly, lack sufficient spatial resolution to identify desirable features, cloud cover and have long-term visiting periods (Cobb et al., 2013;Tattaris et al., 2016). Alternatively, unmanned airborne platforms have the ability for monitoring large scale crop parameters using high spatial and spectral resolution images. Remote sensing platforms using low altitude and flexible unmanned airborne platforms provide more affordable tools for crop phenotyping and precision agriculture (Candiago et al., 2015b). Therefore, UAVs can play a crucial role in the high-performance, near real-time phenotyping for large number of plots and field trials to reduce the potential costs. These UAVs provide high spatial and spectral imagery, useful for determining crop vegetation indices (VIs) and plant phenotyping. In this study, proximal measurements and UAV imageries were used to derive NDVI values for bambara groundnut. The objectives of this study were i) to evaluate NDVI obtained from a proximal sensor and its relationship with the selected phenotypic characteristics of bambara groundnut, ii) to compare the data derived from the UAVs with proximal sensors and chlorophyll content, and iii) to evaluate the ability of UAV data in predicting the crop growth stages.

Description of Study Site
The field trial was conducted at the Field Research Centre of Crops for the Future Research Centre, Semenyih, Selangor (2°55′56.96″ N, 101°52′33.59″ E) during July-October 2019. Block 3 which is 0.1 ha in size with a slight slope was designated for this experiment. Randomised sub blocks were created within the field to take measurements of crop phenotypes. The experimental field is shown in Figure 1. jas.ccsenet.

Evalua
The averag planting (F growth sta The NDVI Then the N Figure 6

Evalua
There was VARI did suggesting org ure 4. The corre

Correlation between Proximal Sensors in Different Crop Stages.
Figure 3 shows that SPAD values and NDVI derived from the GreenSeeker were correlated at all growth stages although during vegetative growth (18 DAP) there was a relatively weak correlation. During this stage, the canopy has a low number of leaves and the small leaf area affects the signal received by the sensor. This is due to the reflectance of plant canopy in visible and near infrared regions that is effected by the amount of green tissue present. The higher the absorption, the higher the value of NDVI. Both biomass and chlorophyll content affect the proximal sensing measurements above the crop (Amaral et al., 2015). This result confirms other results obtained for maize and wheat (Eitel et al., 2008;Solari et al., 2008).
The slope of the regression decreased to 45.27 at 88 DAP (maturity stage, Figure 3). This can be associated with the leaf senescence. During the senescence, at least 10% of leaves are senesced without new leaves being formed to replace them, this phenomenon indicates the beginning of canopy decline. Thus reddening of canopy leaves (Mabhaudhi & Modi, 2013). Similarly, other studies show that temporal profiles of NDVI vary every fortnight, especially at the beginning and at the end of each cycle (Junges et al., 2019). These results show that NDVI has the potential to reflect the vegetation change and canopy development corresponding to canopy photosynthetic capacity (CPC) at different crop growth stages for Bambara groundnut.

Applicability of Both Methods in Deriving the Vegetation Indices
Data from both proximal and remote sensing methods were positively correlated. The proximal measurements shown the capability to predict yield and biomass in maize, wheat and for plant breeding in field condition (Eitel et al., 2008;Solari et al., 2008;Tattaris et al., 2016). However, both systems have their own advantages and drawbacks. The present study is based on spatial and spectral relationship between GreenSeeker and UAV imagery. GreenSeeker provides low resolution data, but with higher focus on the crop and therefore little influence of the space between rows that in turn leads to limited covered area. UAV imagery provides high spatial resolution spectrometry that can be used to generate data for large numbers of plots, in a fraction of time that is required to make ground-based measurements (Gnädinger & Schmidhalter, 2017). In addition, the use of high resolution and low altitude UAVs can address other drawbacks of proximal sensing systems, such as non-simultaneous measurement of various plots, trafficability, small row spaces, plot geometries requiring specific sensor configurations, and vibrations resulting from uneven field slope (Tattaris et al., 2016). However, GreenSeeker is an active system, it is less influenced by lighting conditions. Although highly correlated, the NDVI derived from the GreenSeeker did not exhibit the same frequency distributions, which means caution should be exercised when using this data for site specific crop management.
The variance of the bandwith between both methods in deriving the vegetation indices also limits the extraction of vegetation information. The proximal sensor has broad bandwith compare to the multispectral sensor deploy on the UAV which having narrow bandwith. The proximal sensor (Greenseeker and SPAD) is using a group of bands which leads to a lack of sensitivity especially when applying the vegetation indices on the heterogeneous canopies which consists of cover crop, weed and soil in the interows. This hetegeneneous canopies will lead to the difficulties in discriminating area of interest particularly when vegetation indices respond to other vegetation such as weeds rather than area of interest (Xue & Su, 2017). However, multispectral sensor by using the UAV has the advantages of discriminating heterogenous canopies and sensitivity towards detection of spectral properties on green vegetation. Previous studies had been conducted using UAV imagery for different application such as vegetation and soil segmentation (Hassanein et al., 2018) and crop row detection (Hassanein et al., 2019). The segmentation of vegetation and soil fraction can be implemented by vegetation indices (VIs) using different spectral bands and their combinations (Mesas-Carrascosa et al., 2020). Color vegetation indices (CVIs) are used to emphasize plant greenness using common red, green and blue (RGB) sensors on UAV platforms (Torres-Sánchez et al., 2014). Similar in Jiangsu, China, the study was conducted for estimation of nitrogen where it stated that the indices from multispectral sensor images derived from UAV performed better in the most cases compared to other indices from proximal sensor .
Although the other vegetation index, VARI, was not correlated well with the proximal sensing data at all growth stages, it demonstrated that UAV-based RGB imaging with visible wavebands for assessing vegetation indices was consistent with the results shown in Figure 3 (c). Previous studies have demonstrated that UAV-based RGB indices can be used for crop health monitoring and estimating growth traits in oilseed rape (McKinnon & Hoff, 2017;Wan et al., 2018). RGB images only provide limited crop physiological information. The canopy reflectance reacts strongly to the blue and green light (Amaral et al., 2015;Sulik & Long, 2015). Thus, although VARI is not a replacement for NDVI, the VARI algorithm applied to an RGB sensor can provide valuable information and also be a useful tool to assist farmers in identifying crop stress, monitoring field for crop phenotyping it is shown for sugarcane and oilseed rape (Amaral et al., 2015;McKinnon & Hoff, 2017;Wan et al., 2018).

Advantages, Limitations and Future Work
Aerial based imagery for crop phenotyping is an efficient, cost-effective and suitable technique for complex environments. It can help with rapid identification of growth information with high resolution data. Similar studies were conducted for crop phenotyping (Liebisch et al., 2015;Tattaris et al., 2016) and precision agriculture (Candiago et al., 2015b) which they provide a low-cost approach in order to obtain accurate results for phenotyping in the field environment. However, it will be more useful if hyperspectral sensors could also be used with the aerial-based sensors. This is the key limitation in current crop phenotyping, which limits the amount of information that can be derived from these platforms. In fact, by having hyperspectral data which combine properties of imaging and spectroscopy (Kumar et al., 2016) high resolution spectral vegetation indices such as soil adjusted vegetation indices (SAVI), enhance vegetation indices (EVI) etc. along with crop parameters such as leaf area index (LAI), soil fertility, soil moisture, level of crop stress, yield prediction, biomass can be estimated.
Further study is required to evaluate the capability of data fusion between proximal sensors (SPAD, GreenSeeker) with canopy temperature or any other related data with aerial-based sensors (UAV imagery) to improve monitoring of other crop growth-related traits in field observations. Fusion of data from multiple sensors could provide more information for crop phenotyping which may be especially helpful for underutilised crop studies.
Finally, collecting low cost UAV data (Wang et al., 2018) and linking this data to the global satellite remote sensing databases, in a consistent format that can be shared with other stakeholders working on neglected and underutilised crops will help with the inclusion of these crops in the global crop monitoring projects such as GEOGLAM (Becker-Reshef et al., 2018) for yield forecasting and crop monitoring.

Conclusion
Developing rapid crop phenotyping methods for neglected and underutilised crops is an important step towards ensuring food and nutrition security in a warming world. In this experiment, we found positive relationship between different sensors used for phenotyping and determining developmental stages for bambara groundnut, a neglected and underutilised crop. The results show that there is a potential application for UAV based crop phenotyping in the field. However, there is still a need for validation of results in different environments using different genotypic varieties of bambara groundnut before it can successfully be used for predicting growth stages.