Prediction of Biometric Variables Through Multispectral Images Obtained From Uav in Beans (Phaseolus vulgaris L.) During Ripening Stage

Here, we report the prediction of vegetative stages variables of canary bean crop by means of RGB and multispectral images obtained from UAV during the ripening stage, correlating the vegetation indices with biometric variables measured manually in the field. Results indicated a highly significant correlation of plant height with eight RGB image vegetation indices for the canary bean crop, which were used for predictive models, obtaining a maximum correlation of R2 = 0.79. On the other hand, the estimated indices of multispectral images did not show significant correlations.


Introduction
Bean (Phaseolus vulgaris L.) is a legume with a protein content of 20 to 25% and 50 to 60% carbohydrates, and is part of the human diet worldwide, mainly in developing countries [1]. It is widely cultivated for its enormous genetic diversity, it is a nitrogen-fixing plant, highly adaptable and productive in a wide range of environments [2,3]. This crop will likely play a key role in guaranteeing food security for millions of people around the world in the near future [4]. In Peru in 2019, a total of 73,298 ha of beans were cultivated, representing 0.8% of the Gross Value of Agricultural Production (GVaP) [5]. Canary beans is the most outstanding cultivar in the Peruvian coast due to its preference in the national diet, and cultural aspects. Therefore, due to the importance of this crop, it is a great challenge to monitor its development together with an appropriate agronomic management in the field [6] in the context of climate change.
Quantitative evaluations of biometric variables such as plant height, leaf area index, and chlorophyll content influence yield and are becoming a high priority under precision agriculture [7]. Efficient and non-destructive monitoring of crop growth is essential for accurate crop management and is key to digital agriculture [8]. Determining data manually requires a significant amount of time and resources (measuring equipment, reagents, and researchers, among others). To increase agricultural production with limited resources, important technological advances have been implemented such as the use of unmanned aerial vehicles (UAVs) [9].
UAVs are tools that provide new alternatives of monitoring crops, without direct contact on them [10], allowing the prediction of crop development in a spatial-temporal way. The sensors coupled to a UAV allow estimating vegetative development variables with different vegetative indices [11]. Recently, in the north of China, [12] used UAV images to predict yield in corn crop. Similarly, other researchers, predicted aerial biomass in various crops such as sunflower, corn and wheat [13][14][15][16]. Similarly, in Kenya RGB images were used to estimate growth and nutritional yield of cassava [17]. Since evaluations are necessary during the development of crops and as beans are one of the main foods in the basic family basket, it is essential that research is conducted to guarantee food security under the context of climate change. Therefore, the objective of the present work was to estimate the prediction of plant height, chlorophyll content and leaf area index through multispectral images obtained from UAV of canary bean crop during the ripening stage.

Study site
The study site is located in the research field of the National Institute of Agrarian Innovation (INIA for its acronym in Spanish) (12°4ʹ30.10ʹʹS, 76 ° 56ʹ33.86ʹʹW, 240 masl) ( Figure 1). The experiment was developed during the winter-spring seasons (Jun 26-Oct 20, 2020), using commercial canary beans. This research was conducted in an experimental field of 0.30 ha ( Figure 2) with distances between plants and rows of 0.2 m and 0.9 m, respectively. We used drip irrigation with a distance between drippers of 0.2 m with a flow of 1.20 l/h. The study site is arid according to the climatic classification of Warren Thornthwaitees [18], recording for the year 2020 averages of 76.8% RH, wind speed 3.3m/s and temperature 19.2 °C (Figure 1), and a total annual rainfall of 8 mm. The meteorological data was recorded from an automatic station (VANTAGE Pro2 Plus Davis, California, USA), located at the Alexander Von Humbolt meteorological observatory of the La Molina National Agrarian University at a distance of 0.87 km from the research area.

Field data collection
Three biometric variables were recorded on 16 plots of 7m x 5m at 90, 97 and 101 days after sowing (DAS): • Plant height (cm); it was measured manually from the soil surface to the highest stem apex.
• Leaf area index (LAI); it was estimated from digital images taken on the canopy cover of the plant on a 0.2m x 0.9m frame. Image processing was performed with the Green CropTracker software (v. 1.0, Agriculture and Agri-Food Canada) using a histogrambased threshold method according to [19].
• Chlorophyll content (mg / m2); it was measured with the CCM-300 equipment (Opti-Sciences Hudson, NH, USA) following the methodologies of [20] choosing leaves from the upper third free of mechanical and biotic damage.

UAV RGB and multispectral image acquisition and processing
We used a Quadcopter type UAV platform, DJI Phantom 4 Pro (Shenzhen Dajiang Baiwang Technology Co., Ltd., Shenzhen, China) with a built-in 4864 × 3648 pixel resolution RGB camera. In addition, a multispectral sensor Parrot Sequoia (Parrot SA, France) which is a synchronized array of 4 single-band multispectral camera with 1.2 MP global shutter was attached, taking images in green (550 nm), red (660 nm), red edge (735 nm) and near infrared (790 nm).
The images were collected at 90, 97 and 101 DAS during sunny days with wind speeds lower than 12 m/s, from 11:00 to 13:00 hours approximately on the 16 research plots, as detailed in Figure 2. The flight plan was established with the Pix4Dcapture software (v. 4.12.1, Pix4D SA, Prilly, Switzerland), considering a frontal and lateral overlap of 80%, height of 30 m, speed of 2.8 m/s and the camera focused at nadir position (perpendicular to the ground surface), allowing to obtain a resolution of 0.8 cm and 2.5 cm for RGB and multispectral images, respectively. Figure 3 shows a flow chart for data acquisition from UAV platform to image processing. Image processing was performed with Pix4Dmapper Pro software (v. 4.3.33, Pix4D SA, Prilly, Switzerland), according to the following steps: (i) alignment of geolocated images, (ii) generation of point clouds and geometric correction, and (iii) generation of the digital surface and orthomosaic model using the inverse distance weighting method. The geometric correction was performed considering nine ground control points (GCP) (Figure 2) previously installed and registered with differential GNSS (Global Navigation Satellite System) (South Galaxy G1 model, South Surveying & Mapping Instrument Co. Ltd, Guangdong, China).

Calculation of vegetation indices
The calculation of the vegetation indices from RGB images was carried out after to a normalization of the bands (R = red, G = green and B = blue) in the Pix4Dmapper software. In addition, the multispectral bands were obtained (R = red, G = green, RE = red border and NIR = near infrared). In this research, 19 vegetation indices were estimated to assess biometric parameters (Table 1)

Statistical analysis
We recorded the quantitative data (plant height, leaf area index and chlorophyll content) in a field book for the 16 plots and were then analyzed with the R v software. 4.1.0 [21] with the following packages: GGALLI v. 2.0.0 [22], Hmisc v. 4.5-0 [23]. In addition, we employed R base codes to correlate variables evaluated in the field with those estimated through vegetation indices. Statistical indicators such as the Pearson correlation coefficient (r) and determination coefficient (R 2 ) were determined. Moreover, a principal component analysis was performed with libraries factoextra v.1.0.7 [24] and FactoMiner v.2.4 [25].

Principal component analysis
The principal component analysis (PCA) (Figure 4) defined the interactions between the variables evaluated in the field (plant height, leaf area index and chlorine-row content.) with the vegetation indices (19 in total) monitored by UAV with a cumulative variance of 80.6% for the first two dimensions. In addition, PCA generated differentiated clusters between 90 and 97 DAS. However, the last evaluation (101 DAS) did not report a significant difference, indicating that during ripening stage, these variables in the bean crop show minimal rates of increase.
The PCA variables on the axis of dimension 1 that contribute to the variance represent an approximate of 5% in each one. We can also observe that the multispectral indices were greater at 90 DAS as well as the chlorophyll content ( Figure 4).

Correlation between spectral indices and growth variables
The correlation analysis between the vegetation indices and growth variables were carried out for 90, 97 and 101 DAS, finding significant correlations (p-value <0.05) at 97 DAS for plant height. In contrast, the other variables (chlorophyll content and leaf area index) did not register significant correlations, so we decided not to consider them for the multiple linear regression model.

Prediction model for estimating plant height
After correlating the vegetation indices and plant height at 97 DAS, multiple linear regression analyzes were performed for three prediction models, as detailed in Table 3. Figure 5 shows the three prediction models for the measured plant height and its predicted values: (i) for model I, 10 vegetation indices were used (RGBVI, GBDI, VDVI, NGBDI, NGRDI, MGRVI, VARI, ExGR, CRRI, RGRI) with significant correlations 0.59 < r <0.64, (p-value <0.05), generating a predictive model with an R 2 = 0.79 (p-value <0.01), (ii) in model II four vegetation indices were considered (RGBVI, GBDI, VDVI, ExGR) with significant correlations of 0.63 <r <0.64, (p-value <0.01), obtaining a predictive model with R 2 = 0.60, and (iii) to generate the prediction model III, three vegetation indices were considered ( ExR, ExG, ExB) with negative correlations -0.66 <r<-0.32, generating a predictive model with an R 2 = 0.46. Although for model II four vegetation indices with high correlation values were chosen, their prediction decreases due to the fact that it presents an R 2 lower than model I but higher than that of model III.

Discussion and conclusions
To our best knowledge, this is the first research work in Peru using technology such as UAV for an agronomic monitoring of canary beans. A highly significant correlation for plant height was obtained from eight estimated vegetation indices from RGB images in the canary bean crop at ripening stage. It was possible to predict the height of this crop with high performance (R 2 = 0.79) using model I. Recent studies [2,17] reported similar results for beans and cassava crops, respectively, demonstrating the utility of UAVs and vegetation indices for the compilation of biometric variables. The use of these tools becomes more important due to the need to conduct more agronomic evaluations in a timely manner in crops and mainly in those that are cultivated in arid environments such as the Peruvian coast. We are currently conducting experiments with new technologies using remote sensors such as UAV for the prediction of biometric variables estimated from vegetative indices obtained from RGB and multispectral images in bean crops and others of national importance. In addition, images from the PeruSat-1 (Peru) and Kompsat 3 (South Korea) satellites will be used, allowing timely monitoring and prediction of crop development under different agronomic management scenarios (irrigation, fertilization, among others) on a larger scale.