**Rapid and Non-Destructive Methodology for Measuring Canopy Coverage at an Early Stage and Its Correlation with Physiological and Morphological Traits and Yield in Sugarcane**

**Raja Arun Kumar 1,\*, Srinivasavedantham Vasantha 1, Raju Gomathi 1, Govindakurup Hemaprabha 2, Srinivasan Alarmelu 2, Venkatarayappa Srinivasa 2, Krishnapriya Vengavasi 1, Muthalagu Alagupalamuthirsolai 1, Kuppusamy Hari 1, Chinappagounder Palaniswami 1, Krishnasamy Mohanraj 2, Chinnaswamy Appunu 2, Ponnaiyan Geetha 1, Arjun Shaligram Tayade 1,3, Shareef Anusha 1, Vazhakkannadi Vinu 2, Ramanathan Valarmathi 2, Pooja Dhansu <sup>4</sup> and Mintu Ram Meena <sup>4</sup>**


**Abstract:** Screening for elite sugarcane genotypes for canopy cover in a rapid and non-destructive way is important to accelerate varietal/clonal selection, and little information is available regarding canopy cover and leaf production, leaf area, biomass production, and cane yield in sugarcane crop. In the present investigation, the digital images of sugarcane crop by using *Canopeo* software was assessed for their correlation with the physiological and morphological parameters and cane yield production. The results revealed that among the studied parameters, canopy coverage has shown a significantly better correlation with the plant height (0.581 \*\*), leaf length (0.853 \*\*), leaf width (0.587 \*\*), and leaf area (0.770 \*\*) in commercial sugarcane clones. Two-way cluster analysis has led to the identification of Co 0238, Co 86249, Co 10026, Co 99004, Co 94008, and Co 95020 with better physiological traits for higher sugarcane yield under changing climate. Additionally, in another field experiment with pre-breeding, germplasm, and interspecific hybrid sugarcane clones, the canopy coverage showed a significantly better correlation with germination, shoot count, leaf weight, leaf area index, and plant height, and finally with biomass (*r* = 0.612 \*\*) and cane yield (*r* = 0.458 \*\*). It has been found that the plant height, total dry matter (TDM), and leaf area index (LAI) had significant correlation with the cane yield, and the canopy cover data from digital images act as a surrogate for these traits, and further it has been observed that CC had better correlation with cane yield compared to the other physiological traits viz., SPAD, total chlorophyll (TC), and canopy temperature (CT) under ambient conditions. Light interception determined using a line quantum sensor had a significant positive correlation (*r* = 0.764 \*\*) with canopy coverage, signifying the importance of determining the latter in a non-destructive way in a rapid manner and low cost.

**Keywords:** sugarcane clones; canopy cover; light interception; biomass; cane yield

#### **1. Introduction**

Sugarcane is one the most important industrial crops in global agriculture, and it has emerged as a multiproduct crop benefiting producers and consumers [1]. Sugarcane is the

**Citation:** Kumar, R.A.; Vasantha, S.; Gomathi, R.; Hemaprabha, G.; Alarmelu, S.; Srinivasa, V.; Vengavasi, K.; Alagupalamuthirsolai, M.; Hari, K.; Palaniswami, C.; et al. Rapid and Non-Destructive Methodology for Measuring Canopy Coverage at an Early Stage and Its Correlation with Physiological and Morphological Traits and Yield in Sugarcane. *Agriculture* **2023**, *13*, 1481. https:// doi.org/10.3390/agriculture13081481

Academic Editors: Koki Homma, Jibo Yue, Chengquan Zhou, Haikuan Feng, Yanjun Yang and Ning Zhang

Received: 5 June 2023 Revised: 25 June 2023 Accepted: 17 July 2023 Published: 26 July 2023

**Copyright:** © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

second most important industrial crop after cotton in India, occupying about 5 million ha of land with a sugar production of 32.38 metric tons [2]. The sugar industry is the second largest agro-industry in India, and it contributes to 1.1% of the national GDP besides providing for 4% of the population residing in rural areas [3]. Due to the burgeoning population and other constraints (abiotic stress), Ref. [3] the cultivated area of sugarcane will mostly remain static; hence, the only option for the increasing production is to go the vertical way/enhance crop productivity. Sugarcane is a C4 crop that produces four carbon compounds as the primary product in the carbon assimilation cycle, and it is commonly grown from latitude 36.7◦ N to 31◦ S and from sea level to 1000 m of altitude, and generally sugarcane grows slowly during the early part of its growing period compared to other tropical gramineous crops, taking up to 4 months to produce a complete leaf canopy which intercepts nearly all the incoming radiation [4–7], while maize (*Zea mays*) and pearl millet (*Pennisetum glaucum*) normally produce a complete leaf canopy within 2 months of sowing [8–10]. Owing to the slow production of a complete leaf canopy, dry biomass production is slow in sugarcane during the early part of the growth period. A comparison of sugarcane and maize made in Zimbabwe [11] showed that the growth in dry mass was faster at 4 months after sowing. Sorghum (*Sorghum bicolor*) and maize grow at similar rates [12], suggesting that sorghum grows faster than sugarcane. On the other hand, Bull and Glasziou [4] showed that early growth in dry mass in sugarcane is slow in most of the regions, and high yields produced by sugarcane are mainly due to an extended growth period rather than superior photosynthetic efficiency.

Canopy cover is a useful trait for monitoring crop productivity [13], and canopy photosynthesis is greatest when the crop reaches its maximum canopy cover to intercept nearly most of the incident light and absorb the required photosynthetic radiation for photo-biochemical processes and yield formation [14,15]. The most common method for measuring canopy cover is by determining the light interception with a line quantum sensor [16,17]. Shepherd et al. [13] reviewed the notion that this system would be timeconsuming and costly, as the measurements should be collected near solar noon [16,18].

Another method involves using drone-based digital image capturing and processing to predict the canopy coverage. However, such a facility may not be equally accessible to all in the scientific community.

In this context, a recently developed method of Oklahoma State University for measuring canopy coverage, called *Canopeo*, which rapidly determines the canopy coverage (%) using digital images, employs an application for iOS (Apple) and Android (Google) devices and Matlab (Mathworks) [19]. *Canopeo* (Oklahoma State University App Center, Stillwater, OK, USA) is an automatic colour threshold (ACT) image analysis tool that analyses pixels based on the red-to-green (R/G) and blue-to-green (B/G) colour ratios and an excess green index [13]. *Canopeo* was accurate and faster at computing canopy cover than other software and is widely being used in many crops such as alfalfa, cover crops, soybean, sorghum, wheat, potatoes, and turf grass (https://canopeoapp.com/, accessed on 10 July 2022).

Canopy cover (CC), leaf area, and biomass production are reported to be the most important physiological components resulting in better cane yield in sugarcane; hence, quantification of canopy cover, which is the primary factor for biomass production, is highly essential, and the later requires a lot of labour, resources, and time through a leaf area measurement by destructive sampling or by light interception method using line quantum sensors or by the drone-based image capturing. Several reports are available on various crops regarding canopy cover by *Canopeo*, and little information is available regarding canopy cover, leaf production, leaf area, and biomass production in sugarcane crop. The robustness of the *Canopeo* tool needs to be validated and compared with data generated from line quantum sensor and leaf area measurements for light interception measurements in sugarcane crop. The ICAR-Sugarcane Breeding Institute, Coimbatore, India, a century-old historical institute known for the "Nobilization of cane", has evolved more than 3500 sugarcane clones, and to sustain sugarcane production the canopy coverage (CC) trait is highly essential for screening climate-resilient sugarcane clones. Therefore, the

present investigation was carried out to (i) evaluate sugarcane canopy cover measured with *Canopeo* and with the light interception method using a line quantum sensor to find an association between two different methods and (ii) to analyse the canopy cover in sugarcane including commercial, interspecific hybrids and germplasm clones and to establish its correlation with physiological and morphological parameters, biomass, and cane yield traits in field conditions.

#### **2. Materials and Methods**

#### *2.1. Plant Material and Crop Management of Commercial Hybrids of Sugarcane Clones*

Sugarcane clones of commercial hybrids types viz., CoM 0265, Co 86249, Co 99004, Co 10026, Co 86010, CoC 671, Co 1148, Co 95020, Co 2001-13, Co 86032, Co 7717, Co740, Co 62175, Co 8371, Co 0218, CoLK 8102, BO 91, Co 775, Co 0212, Co 91010, ISH 100, Co 94008, Co 0238, Co 86011, Co 8338, Co 85019, Co 8208, Co 419, CoV 92102, Co 13006, and Co 8021 (Table 1) were grown at the ICAR-Sugarcane Breeding Institute, Coimbatore (11◦0 34 N, 76◦55 2 E, 430 m above mean sea level), Tamil Nadu, India. Two budded sets, thirty-eight per row of 6.0 m, were planted, and a full dose of phosphorous (P2O5) was applied in the furrows before planting as basal fertilization, while nitrogen (N) and potassium (K2O) were applied in two equal measures at 45 days after planting (DAP) and at full earthing-up (90 DAP). Detrashing of dried leaves was done at 5, 7, and 10 months after planting for proper sunlight penetration. The crop stand was free from significant disease or insect damage. The morpho-physiological data, viz., germination percent, leaf length, leaf width, leaf number, leaf area, shoot thickness, and plant height, were determined by following standard procedure.


**Table 1.** Sugarcane clones (source: Hemaprabha et al., 2018) [20].


**Table 1.** *Cont.*

[a] Asterisks (\*) indicate clones suitable for drought conditions [1].

#### 2.1.1. Germination%, Plant Height, and Shoot Thickness

The number of germinants/row was recorded at 30 DAP, and germination % was derived. The plant height was measured from the base to the top most visible transverse mark on the 60 DAP using a measuring tape and the shoot thickness with a digital vernier calliper (Mitutoyo, Kawasaki, Japan) [21].

#### 2.1.2. Leaf Traits

Leaf area (*LA*) was determined in a non-destructive manner by linear measurement method as mentioned by Montgomery (1911):

$$LA = LBK \left(\text{cm}^2\right) \tag{1}$$

where *L* = maximum length of length, *B* = maximum breadth, and *K* = constant (0.75 based on regression analysis).

#### 2.1.3. Biomass

During the formative stage, the biomass samples were collected in a one-meter square area, and all the samples were oven-dried (60 ± 5 ◦C) until a constant weight was reached.

2.1.4. Determination of Canopy Cover in Commercial Hybrids of Sugarcane Crop at Early Formative Phase

The non-destructive method of canopy coverage was recorded at 60 DAP using *Canopeo* software installed in Android mobile phone. *Canopeo* is an application for iOS (Apple, Cupertino, CA, USA) and Android (Google, Mountain View, CA, USA) mobile devices and Matlab (Mathworks, Natick, MA, USA) that can rapidly analyse canopy cover (Figure 1a) from pictures [19]. The accuracy of the CC recorded through *Canopeo* software is 91% (correctly classify pixels as green/true positive), and specificity is 89% (nongreen/true negative) as mentioned by the original author of the software from Oklahoma State University [19]. The distance between the mobile and sugarcane plant while recording the measurements is 80 cm. In order to facilitate easy CC recording, a 35-inch length selfie stick was also used for a few taller clones.

The captured image was processed rapidly through the *Canopeo* software immediately after image acquisition on an Android mobile device, and the derived canopy coverage (%) was saved as a separate folder for further analysis. The canopy coverage image was captured by keeping the mobile parallel to the soil [19]. However, in our study, another method (keeping mobile perpendicular to the soil) was also followed along with the standard method (keeping mobile parallel to the soil), and finally, both methods were compared by correlation to identify the best method/position for capturing the image in the sugarcane crop.

The samples are recorded for the canopy coverage from 9.00 to 11.00 AM, and data are recorded just opposite to the sunlight direction in order to avoid the shade of the observer. The represented values are the average of four observations per replication, i.e., a total of 8 observations per treatment.

2.1.5. Determination of Light Interception in Commercial Hybrids of Sugarcane Crop at Early Formative Phase

The light interception (LI%) was determined using line quantum sensor LI-191SA (LICOR Inc., Lincoln, NE, USA) connected with the LI-1400 a multipurpose datalogger that functions both as a data logging device and a multichannel, auto-ranging meter between 11.00 to 12.00 IST (Figure 1b). The intercepted photosynthetically active radiation (IPAR) for a particular day was computed as the difference between incident PAR at the top and the transmitted PAR received at the bottom of the canopy (the radiation reflected from the crop and soil was also taken into account for deriving the LI%), and the correlation between the CC% and LI % was conducted. Also, the radiation reflected by the soil surface was also determined and finally incorporated for the LI% calculation.

#### *2.2. Plant Material and Crop Management of the Breeding Population, Interspecific Hybrids, and Basic Species Clones of Sugarcane Clones*

In order to determine the correlation between the canopy coverage with biomass and cane yield, 38 sugarcane clones including improved breeding population clones (004-73, 04-423, 14-154, 07-520, 04-595, 04-472, 12-127, 97-77, 01-807, 20-158, 20-614, 20-335, 99-45, 98-290, WL 10-40, 14-161, 81 GUK 192, 81 GUK 527, 92 GUK 220, 97 GUK 111, 98 GUK 116, GUK 00-910, GUK 02-91, GUK 06-402, 88 GUK 072, 97 GUK 9, 97 GUK 74, 987 GUK 124, 20-191, 07-776, 99-19, 99-291, 06-013, and 01-803), interspecific hybrids (ISH 107 and ISH 111), and germplasm clones (Kheli and Pathri) were planted in the randomized block design in two replications at the experimental farm of ICAR-Sugarcane Breeding Institute, Coimbatore, India (11◦0 34 N, 76◦55 2 E, 430 m above mean sea level) during the years 2021–2022. The canopy coverage was recorded at 60 DAP using *Canopeo* software and analysed as mentioned in Experiment 1. The germination, shoot thickness, and plant height were determined as mentioned in Experiment 1.

During the formative stage, the fresh biomass samples were collected in a one-meter square area, and all the samples were separated into leaf, sheath, and stem parts and were oven-dried (60 ± 5 ◦C) for determination of constant weight. The constant dry weight was used for computing the overall dry matter production.

#### 2.2.1. SPAD

Non-destructive chlorophyll estimation was recorded using a SPAD meter (Soil Plant Analysis Development) (atLeaf, Wilmington, Delaware, NC, USA) that computes the chlorophyll content of a leaf by recording the transmission of red light and infrared light at 660 nm and 940 nm, respectively, and converts the reading into a digital signal [1].

#### 2.2.2. Canopy Temperature

The canopy temperature was measured with a thermal imaging infrared camera (FLIR E6) between 11:30 a.m. and 12:00 noon on cloudless days. The image captured was processed through FLIR software (FLIR Tools version 5.1.15036.1001), and the final data were used. The thermal imaging camera was held to view the crop at a 30◦ angle from horizontal at a 90◦ angle to the row, with the minimum exposure to the soil, and the emissivity factor of 0.95 was used for the green canopy. Each canopy temperature measurement was the average of three readings at different locations in each clone. Images were registered in the Thermal MSX® mode (FLIR Systems, Wilsonville, OR, USA), and files were saved in standard 14-bit JPG format.

#### 2.2.3. Chlorophyll Fluorescence

Chlorophyll fluorescence (*Fv*/*Fm*) was measured in intact sugarcane leaves using a chlorophyll fluorometer (model OS30p, Opti-Sciences, Hudson, NH, USA). The leaves were dark-adapted for 15 min using leaf clips (Opti Sciences), and the (*Fv*/*Fm*) readings were recorded by passing a saturating light:

$$\frac{F\_v}{F\_m} = \frac{F\_m - F\_o}{F\_m} \tag{2}$$

where *Fv*/*Fm* = ratio of variable fluorescence to maximal fluorescence, *Fm* = maximal fluorescence, *Fo* = minimal fluorescence, and *Fv* = variable fluorescence of photosystem II [1].

#### 2.2.4. Sucrose, Cane Yield, and CCS

Sugarcane juice was extracted in a crusher with 65% extraction capacity, and the juice quality was analysed as total soluble sugars (TSS) (Brix) and sucrose content (Pol%) according to the standard method [22]. Cane yield was estimated at the 12th month of the crop stage, and the middle 4 rows of canes were harvested and weighed for the plot yield, and the yield per hectare was calculated and expressed as t ha−1. Commercial cane sugar (CCS) was determined and expressed in percentage and t ha−<sup>1</sup> according to Equations (3) and (4), respectively [21].

$$\text{CCS\%} = \frac{\text{Success content} \times 1.022}{\text{TSS} \times 0.292} \tag{3}$$

$$\text{CCS (t/ha)} = \frac{\text{CCS\%} \times \text{Yield}}{100} \tag{4}$$

#### *2.3. Statistical Analysis*

Analysis of variance (ANOVA) was performed on the data following the method of Gomez and Gomez (1984) [23], and the least significant difference (LSD) values were calculated at the 5% probability level. Duncan multiple range test (DMRT) was performed to separate significant genotypes, and alphabets were superscripted for easy view. The Pearson-product-moment correlation coefficients (*r*) between leaf length, leaf width, leaf number, leaf area, shoot thickness, plant height, germination percent, and canopy cover were computed using SAS 9.3 (SAS Institute, Cary, NC, USA) [24]. The scatterplot matrix showing the correlation and frequency counts among the studied parameter, i.e., canopy coverage % (CC), cane yield (CY), dry biomass at formative phase (DWFP), shoot diameter (SD), leaf area (3rd top visible dewlap), leaf area (cm2), leaf width (cm), leaf length (cm), leaf number (L.No), and plant height (PH), was created using JMP genomics software version 6.1. Two-way cluster analysis with the wards method showing the grouping of sugarcane clones with the studied parameter was also conducted using JMP genomics software. Regression analysis was carried out between the CC and biomass, cane yield, and their corresponding slope (*β*), and significance was determined following "F" test at 0.05 probability. A correlation diagram displaying the correlation between the studied parameters along with the *p*-value was conducted through R software version 4.1.3.

#### **3. Results**

#### *3.1. Association between Sugarcane Canopy Cover Measured with Canopeo at Different Positions*

The canopy cover (CC) data was determined at the early formative phase (60–150 DAP) through *Canopeo*, and both digital images acquired parallel to the ground and perpendicular to the ground were analysed for their association and relevance in sugarcane crop. Among the four phases of the sugarcane crop, the formative phase which starts at 60 DAP is reported to have high relevance to the cane yield; hence, the CC data were recorded at 60 DAP. The correlation between canopy coverage (%) from an image acquired parallel to the ground and perpendicular to the ground is shown in Figure 2a. A significantly better correlation of r = 0.870 \*\* was observed between the canopy coverage (%) data through images acquired parallel to the ground and perpendicular to the ground of the sugarcane crop. Canopy cover images were taken in properly weeded/weed-free fields to reduce the data error; i.e., the background images of weeds mimic the crop, and this results in an overestimation of CC data. The data revealed a significant linear relationship at 1% probability level between the data captured through two positions.

**Figure 2.** (**a**) Correlation between canopy coverage (%) image acquired parallel to the ground and perpendicular to the ground. \*\* denotes significant at 1%. (**b**) Canopy coverage (%) images, i.e., original image (left) and classified image (right) of sugarcane clone.

#### *3.2. Association between Sugarcane Canopy Cover Measured with Canopeo and with the Light Interception by PAR (Photosynthetically Active Radiation) Line Quantum Sensor*

The light interception (LI) data were recorded simultaneously while capturing the canopy cover images through *Canopeo* using Android mobile. Light interception data were acquired through multi-channelled PAR quantum sensors; i.e., one line quantum sensor was placed diagonally between the rows of sugarcane crop, and another line quantum sensor between and above the crop for measuring the transmitted PAR and reflected PAR simultaneously. Incident PAR measurement was achieved through a point sensor for ease of work.

Further, a significant correlation (r = 0.764 \*\*) was observed between canopy coverage (%) from images acquired in parallel to the ground, and light interception by a line quantum sensor (Figure 3) confirms the accuracy of the CC data of *Canopeo*. A positive coefficient indicates that as the value of the independent variable (canopy cover) increases, the mean of the dependent variable (light interception) also tends to increase. The slope coefficient or *β* value of the regression was 0.695, and the coefficient represents the mean increase of LI% for every additional increment of CC. In the present regression equation, (Figure 3) for every 0.695 increment in CC, a correspondingly one unit increment in LI was observed, and the model was found statistically significant at 1% probability through the "F" test.

**Figure 3.** Correlation between canopy coverage (%) image acquired in parallel to the ground and light interception. \*\* denotes significance at 1%.

#### *3.3. Canopy Cover in Sugarcane Crop, and Its Correlation with Morphological Parameters, Biomass, and Cane Yield Traits in Field Conditions*

The results for CC%, germination %, leaf area (cm2), leaf length (cm), leaf width (cm), leaf number, plant height, and shoot thickness are shown along with LSD at 5% in Table 2.

**Table 2.** Variation in canopy coverage (CC), germination % (G), and leaf and shoot morphology in sugarcane clones under field condition.


CC%: Canopy coverage %, G%: Germination %, LL: Leaf length, L.No: Leaf number, LW: leaf width, PH: Plant height, SD: Shoot thickness. NS: Non-significant. n=3Values carrying the same letters as superscripts in each column are not significantly different from each other treatment.

#### 3.3.1. Canopy Coverage

The mean canopy coverage (CC%) of the sugarcane crop was 25.7%, and the minimum and maximum CC% were 15.9 and 32.8, respectively (Table 2). Among the studied clones, Co 0212, Co 0238, Co 10026, Co 62175, Co 85019, Co 86010, Co 86249, Co 94008, Co 95020, Co 99004, and CoM 0265 were recorded with better canopy coverage of more than 25%, while Co 13006, BO 91, and Co 1148 indicated a poor CC% of less than 17%.

#### 3.3.2. Plant Height

The mean plant height of the sugarcane crop was 22.5 cm, and the minimum and maximum plant height were 18.5 and 27.33, respectively (Table 2). Among the studied clones, Co 0212, Co 0238, Co 10026, Co 1148, Co 13006, Co 2001-13, Co 62175, Co 740, Co 8021, Co 85019, Co 86010, Co 86032, Co 86249, Co 94008, Co 95020, and Co 99004 were recorded with better plant height of more than the mean plant height (22.5 cm). The clones, viz., Co 95020, Co 10026, Co 62175, and Co 99004, were observed with significantly better plant height compared to other studied clones.

#### 3.3.3. Germination Percentage

The mean data of germination % were 41.25, and the clones, viz., Co 2001-13, Co 86249, Co 94008, Co 10026, CoV 92102, and BO 91, recorded significantly better germination %, while the clones Co 13006, CoLk 8102, and CoM 0265 showed relatively less germination % (<40%) (Table 2).

#### 3.3.4. Leaf Area, Leaf Number, Leaf Length, and Leaf Width

The mean leaf length (3rd top visible dewlap) of the sugarcane clones was 80.58 cm, and the clones, viz., Co 95020, Co 0212, Co 0238, Co 10026, Co 94008, Co 86249, Co 85019, Co 8021, and Co 62175, were recorded with significantly better leaf length (>85 cm) than other clones, while Co 86032, BO 91, Co 740, and Co 1148 observed with poor leaf length (Table 2). The mean leaf no. per shoot was 5.7, and non-significant differences were observed among the studied clones. The mean leaf width (3rd top visible dewlap) of the sugarcane clones was 2.4 cm, and the clones, viz., Co 94008, Co 95020, Co 86249, Co 86010, Co 85019, Co 10026, Co 99004, Co 2001-13, and Co 0238, exhibited better leaf width (>2.4 cm), while BO 91 and CoLk 8102 recorded poor leaf width compared to other clones.

#### 3.3.5. Shoot Thickness

The mean shoot diameter of the sugarcane clones was 11.9 mm, and the clones, viz., Co 85019, Co 740, Co 10026, Co 0238, Co 62175, and Co 95020, were observed with significantly better shoot diameter, while BO 91, Co 0212, Co 1148, and CoLk 8102 recorded less shoot diameter.

#### 3.3.6. Scatterplot Matrix

The scatterplot matrix showing the correlation and frequency counts among the studied parameter, i.e., canopy coverage % (CC), cane yield (CY), dry biomass at formative phase (DWFP), shoot diameter (SD), leaf area (3rd top visible dewlap), leaf area (cm2), leaf width (cm), leaf length (cm), leaf number (L.No), and plant height (PH), is shown in Figure 4. The canopy coverage % data acquired through the image have shown a significantly better correlation with plant height (0.581 \*\*), leaf length (0.853 \*\*), leaf width (0.587 \*\*), and leaf area (0.770 \*\*).

**Figure 4.** Scatterplot matrix showing the correlation, frequency counts among the studied parameter, i.e., canopy coverage % (CC), dry biomass at formative phase (DWFP), shoot diameter (SD), leaf area (3rd top visible dewlap) (LAI), leaf area (cm2) (LA2), leaf width (cm) (LW), leaf length (cm) (LL), leaf number (L.No), and plant height (PH).

#### 3.3.7. Two-Way Cluster Analysis

Two-way cluster analysis showing the grouping of sugarcane clones with the studied parameter is shown in Figure 5. The results revealed three distinct clusters: **Cluster I**: BO 91, CoLk 8102, Co 1148, Co 13006, Co 86032, and CoV 92102; **Cluster II**: Co 0212, CoM 0265, Co 0238, Co 86249, Co 10026, Co 99004, Co 94008, and Co 95020; and **Cluster III**: Co 2001-13, Co 85019, Co 62175, Co 86010, Co 8021, and Co 740. Among the three clusters, Cluster I was recorded as relatively lesser in plant height (20.33 cm), leaf number (5.49), leaf length (78.9 cm), leaf width (2.1 cm), total leaf area (695 cm2), shoot diameter (11.2 mm), dry biomass (510 g dry-weight m−2), and canopy coverage (18%), while Cluster II was observed with better plant height (23.64 cm), leaf length (94 cm), leaf width (2.65 cm), total leaf area (1085 cm2), and canopy coverage (31%). Cluster III was recorded with better leaf number, dry biomass, and shoot thickness among the studied parameters.

**Figure 5.** Two-way cluster analysis displaying the ward method grouping of sugarcane clones based on the studied parameter.

*3.4. Canopy Cover in Sugarcane Crop (Improved Breeding Population, Interspecific Hybrid, and Basic Germplasm Clone) and Its Correlation with Physiological, Morphological, and Cane Yield Traits*

#### Canopy coverage:

The mean canopy coverage (CC%) of the sugarcane crop was 32.5%, and the minimum and maximum CC% were 17.2 and 49.0, respectively (Table 2). Among the studied clones, the 004-73, 04-423, 14-161, GUK 06-402, and 01-803 were recorded with better canopy coverage of more than 40% (Figure 2b), while 04-595, 97 GUK 111, and 98 GUK 116 indicated a poor CC% of less than 20% (Figure 6).

**Figure 6.** The mean cane canopy (CC%) coverage, cane yield (t/ha), and total dry matter (TDM) of various pre-breeding, germplasm, and interspecific hybrid sugarcane clones.

Cane yield:

The mean, minimum, and maximum cane yield in sugarcane clones (improved breeding population, interspecific hybrid, and basic germplasm) were 72.4, 37.3, and 123.8 (t/ha). Among the studied clones, 07-520, 12-127, GUK 06-402, 987GUK 124, 99-291, Pathri, and ISH 111 recorded better cane yield compared to other clones (Figure 6).

Distribution of the dry matter partitioning and physiological and morphological traits:

The distribution of the dry matter partitioning (LWT: leaf weight, SHWT: sheath weight, STWT: stem weight, and TDM: total dry matter (g .dwt.m−2) and S.Hgt (cm) in the studied sugarcane clones are shown in Figure 7a. The mean LWT, S.Hgt, SHWT, STWT, and TDM were 349, 191, 251,579, and 1179 g .dwt.m−2. The distribution of the SUC%: Juice sucrose, CC: Canopy cover, CCSY: commercial cane sucrose, CT: Canopy temperature (◦C) (Figure 7b), CY: cane yield, GC: germination count, LAI: leaf area index, NOC: number of canes, NOL: Number of leaves, SHC: shoot count, and SPAD: Soil Plant Analysis Development ratios are shown (Figure 7b). The mean SUC%, CC, CCSY, CT, CY, GC, LAI, NOC, NOL, SHC, and SPAD were 16, 32, 8.2, 33, 72.4, 12, 2.09, 8.2,77, 61, 23, and 26, respectively. The distribution of the chlorophyll fluorescence (*Fv/Fm*) and total chlorophyll (mg.cm−2) are shown in Figure 7c. The mean chlorophyll fluorescence (CFL) and total chlorophyll (TC) were 0.609, and 0.0204, respectively.

**Figure 7.** *Cont*.

**Figure 7.** (**a**) Box plot displaying (upper quartile, lower quartile, median, upper extreme, lower extreme, whisker, outlier, and mean (horizontal green line) and the distribution of the dry matter partitioning LWT: leaf weight, SHWT: stem weight: STWT, and TDM: total dry matter (g .dwt.m<sup>−</sup>2) ratios) and S.Hgt: shoot height (cm). (**b**) Box plot displaying the distribution of the SUC%: Juice sucrose, CC: Canopy cover (%), CCSY: commercial cane sucrose (ton/ha), CT: Canopy temperature ( ◦C) CY: cane yield (ton/ha), GC: germination count/row, LAI: leaf area index, NOC: number of canes/row, NOL: Number of leaves, SHC: shoot count/row, and SPAD: Soil Plant Analysis Development ratios. (**c**) Box plot displaying the distribution of the physiological (CFL: Chlorophyll fluorescence and TC: total chlorophyll content mg·cm−<sup>2</sup> ratios). (**d**) Thermal image displaying the canopy temperature (◦C) sugarcane crop.

#### *3.5. Correlation between Physiological and Morphological with Canopy Coverage*

The correlation between physiological and morphological traits and cane yield is shown in Figure 8. Cane yield, SUC%, CCSY, and TDM had significant correlations with canopy coverage (r = 0.46 \*\*, 0.42 \*\*, 0.51 \*\*, 0.62 \*\*, respectively). The germination count, shoot count, initial plant height, and final plant height also had significant correlation with CC (r = 0.46 \*\*, 0.56 \*\*, 0.60 \*\*, and 0.35 \*, respectively), while chlorophyll fluorescence (*Fv/Fm*), canopy temperature (CT), SPAD, and total chlorophyll (TC) showed a non-significant association with CC (r = −0.19, 0.00, 0.00, and −0.01, respectively). The

leaf weight and stem weight also revealed a positive correlation (0.36 \* and 0.65 \*\*) with CC. Also, the correlation between physiological traits, viz., chlorophyll fluorescence, SPAD, total chlorophyll (TC), and canopy temperature (CT), with cane yield (CY) (r = −0.10 ns, −0.23 ns, −0.23 ns, 0.01 ns, respectively). The leaf area index (LAI), plant height (PH), and total dry matter (TDM) had significant correlations with CC (r = 0.44 \*\*, 0.60 \*\*, and 0.62 \*\*, respectively).

**Figure 8.** Correlation between various physiological and morphological traits with yield, viz., SPAD, TC:total chlorophyll (mg/cm2), CT: Canopy temperature (◦C), NOC: number of canes/row, NOL: number of leaves, S.Hgt: shoot height (cm) at final stage, SHWT: Sheath weight (g.m<sup>−</sup>2), LWT: leaf dry weight (g.m−2), LAI: Leaf area index, GC: germination count, SHC: shoot count-early stage, PH: plant height at early stage (cm), CC: canopy coverage (%), STWT: stalk weight (g.m−2), TDM: total above-ground dry matter (g.m−2), SUC%: Juice sucrose%, CCSY: Commercial cane sugar (t ha−1), and CY: Cane yield (t ha−1), \*\*\* denotes *p* < 0.001, \*\* denotes *p* < 0.01; \* denotes *p* < 0.05; and ns denotes non-significant *p* ≥ 0.05. The intensity of the colour indicates the strength of the correlation.

Based on the canopy coverage data, the simple regression analysis has revealed the prediction of cane yield and total dry matter (TDM) as per Equations (5) and (6) mentioned below:

$$\text{Cane yield} = (33.00 + 1.213 \,\text{C} \,\text{°C} \,\text{°C}) \tag{5}$$

$$\text{Total dry matter} = (198.18 + \text{30.204 CC\%})\tag{6}$$

These simple Equations (5) and (6) suggest the usefulness of the canopy coverage trait in forecasting cane and total dry matter in sugarcane in a rapid and accurate way.

#### **4. Discussion**

#### *4.1. Position of the Camera and Comparison of Light Interception (LI) with Canopy Cover (CC)*

This paper describes the methodology of canopy cover (CC) determination in sugarcane and its association with dry matter and cane yield. A significant correlation (r = 0.870 \*\*) was observed between canopy coverage (%) from images acquired parallel to the ground and perpendicular to the ground (Figure 2). Also, the canopy coverage (%) from images acquired parallel to the ground had a significantly better correlation with the leaf area thus revealing that the parallel position of the camera for capturing the image for CC% is suited for the sugarcane crop. A significant positive correlation coefficient (0.764 \*\*) between light interception (by line quantum sensor) and canopy cover coverage (by *Canopeo*) indicates (Figure 3) the similarity between both data. There is a strong correlation between the canopy coverage (%) image acquired in parallel to the ground and light interception (Figure 3) by a line quantum sensor which we have observed in sugarcane crops also in the present investigation [13]. Similar to our findings, others have also reported that the ground coverage values estimated from digital images taken above the canopy have been correlated to light interception measurements which are limited by the time of measurements and the presence of clouds [25]. The limitation of this light interception method is that the measurements should be taken in unobstructed sunlight and close to solar noon [26]. The canopy cover methodology for estimating light interception in soybeans has been reviewed to have advantages over the above limitations [27]. In this technique, ground area coverage was determined by digital images taken above the canopy. The canopy coverage values were similar throughout the day and were correlated in a one-to-one relationship with light interception measurements made with a line quantum sensor at solar noon. Shepherd et al. (2018) [13] have also reported a linear relationship between canopy cover measured with pictures (R2 = 0.94) and videos (R2 = 0.92) in *Canopeo* and light interception.

#### *4.2. Germination*

Better germination of sugarcane sets in the field is often reported to be linked with the early vigour. In our study, the mean germination % was 44.21, and the clones, viz., BO 91, Co 10026, Co 740, Co 8021, Co 86032, Co 86249, Co 94008, CoV 92102, Co 95020, and Co 99004, recorded a significantly better germination (>44) percentage. Several reports [28,29] suggest that, due to the genetic nature and environment, there exists high variability in sett germination percent in sugarcane varieties, and these reports corroborate our findings.

#### *4.3. Leaf Length, Width, Leaf Number, and Leaf Area*

The rate of leaf appearance is cultivar-dependent and determined mainly by temperature [30], but it can also be altered by water stress that decelerates expansive growth [31]. Our experiment also confirms the previous report [30] having greater variability in leaf number which suggests that the variation is mainly due to clonal dependence at ambient conditions. Differential thermal requirements for nine sugarcane cultivars to produce the first leaves and the association of the rate of leaf appearance which has the potential for increasing yield [30] are determined by the extent of genetic variation apart from environmental influence.

Leaf arrangement was associated with higher sugar/metric ton, and selection by breeders for higher leaf area indices and for optimum leaf arrangement is suggested [32]. A significant positive correlation between leaf area index and ground cover in potatoes (*Solanum tuberosum*) under different management conditions has been reported [33], and this shows that the canopy coverage (%) image acquired non-destructively through *Canopeo* software using simple android mobile will be useful in determining the leaf area of the sugarcane crop at an early stage rapidly compared to the conventional destructive methods which consume a lot of labour and other resources. *Canopeo* is faster at calculating a canopy cover percentage and can be easily done while in the field. It took less than 1 min to take

three pictures or one video per plot, and with the line quantum sensor, data collection time per plot was variable due to cloud cover.

#### *4.4. Dry Matter Production or Biomass*

Most of the better-performing sugarcane clones (Co 86010, Co 85019, and Co 10026) identified in this study had a drought-tolerant parent [1], and, in addition to that, Co 62175, Co 85019, and Co 10026 were high-biomass clones. The poor performance of the clones, viz., BO 91, Co 1148, and CoLK 8102, might be plausibly due to their best suitability to subtropical Indian areas rather than a tropical condition in India, while the clones Co 10026, Co 86249, Co 99004, Co 94008, and Co 95020 are of high biomass type with better leaf area production resulting in better canopy coverage.

#### *4.5. Tiller Number and Plant Height*

The variability (high tillering and shy tillering) in sugarcane tillering and its relation to sugarcane productivity [34,35] have been widely discussed [36]. It was reported that the number of tillers and plant height at six months after planting are highly correlated with canopy cover (*rg* = 0.72) and canopy height (*rg* = 0.69), respectively [37]. Our results are in line with the previous study of [38] which reported that early biomass had a high genetic correlation with unmanned aerial vehicle (UAV)-derived canopy height (0.810) and canopy cover (0.710). Capturing spectral reflectance by means of UAV at the whole canopy level rather than at the individual leaf level has been an important contributing factor for the high trait-yield correlation compared to individual leaf spot measurements which do not represent whole-canopy dynamics [38].

Canopy cover is a useful trait related to crop growth, water use, and stalk number, and cane yield is considered an important parameter in crop monitoring [37].

#### *4.6. Canopy Temperature and Cane Yield*

Canopy temperature, a surrogate trait for canopy conductance, has been previously monitored in sugarcane, and it showed a significant genotypic variation and a strong negative genetic correlation with biomass [39,40]. Our study (Figures 7b,d and 8) observed similar findings (r = 0.04 ns, r = 0.01 ns between CT and TDM, CY, respectively) and also corroborates the report [41] where canopy temperature has been reported as highly negatively correlated with stalk productivity (r = −0.53 \*\*) under drought stress, while there is a non-significant correlation (r = −0.18 ns)). Under ambient conditions, the canopy temperature is generally observed with less variability (poor r value with cane yield) among the sugarcane clones, and the better expression of canopy temperature is observed only under abiotic stress conditions where the deeper roots function in tapping of water at deeper zones and support transpiration with subsequent higher canopy conductance, canopy cooling, and better correlation with crop yield.

#### *4.7. Chlorophyll Fluorescence vs. TDM and Cane Yield*

Chlorophyll fluorescence is being reported to be one of the best traits for screening the healthy crop under abiotic stress and in the present investigation (Figure 3) where the crop responses under ambient conditions did not translate in the form of TDM (r = −0.24 ns) and cane yield (r = −0.10 ns). The chlorophyll fluorescence exhibits a non-significant correlation (r = 0.02 ns) with cane yield under ambient conditions, while a positive correlation of *Fv/Fm* with stalk productivity (r = 0.56 \*\*) under drought stress [41].

#### *4.8. SPAD Index vs. TDM and Cane Yield*

The SPAD index is a widely discussed trait for the rapid determination of chlorophyll content, and it is also reported to have a significant correlation with crop yield. Chlorophyll is the basic molecule that helps in the absorption of solar radiation and aids in the synthesis of carbohydrates through photosynthesis and finally crop yield. In our experiment, a non-significant correlation of *<sup>r</sup>* <sup>=</sup> −0.01 ns and *<sup>r</sup>*<sup>=</sup> −0.23 ns was observed between the TDM, cane yield, and SPAD (Figure 3). These findings corroborate the findings of conclusions of Silva (2007) where the SPAD index has been reported to have a non-significant correlation with stalk productivity (r = 0.19 ns) under ambient condition, while there is a significant correlation (r = 0.36 \*\*). Thus, it reveals that the SPAD index is a useful trait preferably under abiotic conditions, where the stress leads to loss of chlorophyll and declined photosynthesis and reduced synthesis of carbohydrates and finally crop yield.

#### *4.9. Canopy Coverage vs. TDM and Cane Yield*

Canopy cover is a valuable trait for monitoring crop productivity [13], and canopy photosynthesis is greatest when the crop reaches its maximum canopy cover to intercept virtually most of the incident light and absorbs the required photosynthetic radiation for photo-biochemical processes and yield formation [14,15]. From the present study, it is clear that the GUK clones had significantly better CC and cane yield compared to other clones. The GUK clones have the parental genes of *Erianthus sps* which is fast growing, with more leaf area, CC, biomass, and cane yield. It has been reported that *Erianthus sps* exhibits vigorous growth, high biomass production, and high tillering ability and is suitable for abiotic stress conditions [42]. Our experiment results (Figure 8) also confirm the previous reports by displaying significant correlations of r = 0.46 \*\*, 0.42 \*\*, 0.51 \*\*, and 0.62 \*\*, respectively, of CY, SUC%, CCSY, and TDM with canopy coverage. The germination count, shoot count, initial plant height, and final plant height also had a significant correlation with CC (r = 0.46 \*\*, 0.56 \*\*, 0.60 \*\*, and 0.35 \*, respectively). The leaf weight and stem weight also revealed a positive correlation (0.36 \* and 0.65 \*\*) with CC (Figure 8). From the overall discussion, it has been found that the plant height, total dry matter (TDM), and leaf area index (LAI) had significant correlation with the cane yield, and the canopy cover data from digital images act as a surrogate for these traits, and further it has been observed that CC had better correlation (Figure 8) with cane yield compared to the other physiological traits, viz., SPAD, total chlorophyll (TC), and canopy temperature (CT).

### *4.10. Summary of Key Findings, Advantages, and Limitations*

#### 4.10.1. Key Findings

In the present investigation, the canopy covering digital images of sugarcane crop by using *Canopeo* software was evaluated for its correlation with the physiological and morphological parameters and cane yield production. The results show that among the studied parameters, canopy coverage had a significantly better correlation with the plant height (0.581 \*\*), leaf length (0.853 \*\*), leaf width (0.587 \*\*), and leaf area (0.770 \*\*) in commercial-type sugarcane clones (Figure 4).

Canopy cover data of sugarcane clones (improved breeding population, interspecific hybrid, and basic germplasm) also revealed a significant correlation of r = 0.46 \*\*, 0.42 \*\*, 0.51 \*\*, and 0.62 \*\*, respectively, of cane yield (CY), juice sucrose (SUC%), commercial cane sugar yield (CCSY), and total dry matter (TDM) with canopy coverage (CC). The germination count, shoot count, initial plant height, and final plant height also had a significant correlation with CC (r = 0.46 \*\*, 0.56 \*\*, 0.60 \*\*, and 0.35 \*), respectively, while the chlorophyll fluorescence, canopy temperature, and SPAD index revealed a poor correlation with TDM and cane yield. The leaf weight and stem weight also revealed a positive correlation (0.36 \* and 0.65 \*\*) with CC (Figure 8). From the overall discussion, it has been found that the plant height, total dry matter (TDM), and leaf area index (LAI) had a significant correlation with the cane yield.

#### 4.10.2. Advantages

The traditional light interception method for determining canopy coverage using a line quantum sensor also had a significant positive correlation (r = 0.764 \*\*) with canopy coverage captured through *Canopeo*; thus, our results signify the importance of canopy coverage determination by *Canopeo* in a rapid, non-destructive way and low-cost way.

#### 4.10.3. Limitations

The presence of weeds in the crop field background poses difficulty to classifying or differentiating the crop and weed, and for measuring the canopy coverage, the crop should be in a completely weed-free as well as also detrashed field (removal of senescence leaf) which is more suitable to avoid overestimation of the canopy coverage. If the camera lens were nearer to the crop, then the canopy few crop portions may be excluded in the analysis, and on the other hand, extra sugarcane rows would have been included in the image if the camera lens were positioned at a greater height above the top of the canopy. The vegetation taller than about 2.5 m requires the use of aerial images or special equipment [19].

#### *4.11. Future Research Direction*

The canopy coverage data measurement through the drone/unmanned aerial vehiclebased image and the utilization of pix4d software version 4.8.4 and other software are an emerging trend for the determination of canopy coverage which is valuable for yield forecasting in sugarcane and other crops.

#### **5. Conclusions**

The present investigation revealed that in commercial sugarcane clones the mean data of canopy cover were 25.77%, and the clones, viz., Co 95020, Co 0212, CoM 0265, and Co 86249, showed significantly better canopy cover % (>30%) compared to other clones, while the clones Co 13006, BO 91, and Co 1148 were observed with poor canopy coverage (<20%). Also, among the observed traits, canopy coverage % data acquired through image have shown a significantly better correlation with the plant height (0.581 \*\*), leaf length (0.853 \*\*), leaf width (0.587 \*\*), and leaf area (0.770 \*\*). Further, there is a significant correlation (r = 0.585 \*\*) between the canopy coverage (%) image acquired in parallel to the ground and the light interception through line quantum sensors which consume more labour and costly instruments/sensors. The canopy coverage (%) image acquired non-destructively through using simple Android mobile will be useful in determining the leaf area of the sugarcane crop at an early stage rapidly compared to the conventional destructive methods which consume a lot of labour and other resources. Two-way cluster analysis revealed that Cluster II comprising Co 0212, CoM 0265, Co 0238, Co 86249, Co 10026, Co 99004, Co 94008, Co 95020 Co 0238, Co 86249, Co 10026, Co 99004, Co 94008, and Co 95020 was observed with better plant height (23.64 cm), leaf length (94 cm), leaf width (2.65 cm), total leaf area (1085 cm2), and canopy coverage (31%). In a second field experiment with diverse sugarcane clones (improved breeding population, interspecific hybrid, and basic germplasm), the canopy coverage showed a significantly better correlation with biomass (*r* = 0.612 \*\*) and cane yield (*r* = 0.458 \*\*), while the chlorophyll fluorescence, canopy temperature, and SPAD index revealed a poor correlation with TDM and cane yield. Light interception determined using a line quantum sensor had a significant positive correlation with canopy coverage signifying the importance of canopy coverage determination in a non-destructive way.

**Author Contributions:** Project formulation and execution of the experiment: R.A.K., S.V., G.H., K.V., S.A. (Shareef Anusha), S.A. (Srinivasan Alarmelu), K.M. and V.S.; Canopy cover data analysis: R.A.K., P.G. and C.P.; Data analysis: R.A.K. and M.A.; Physiological analysis: R.A.K., K.H., V.V. and R.V., drafting of the manuscript: R.A.K., V.S., K.V., C.A., A.S.T., R.G., P.D. and M.R.M. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding.

**Institutional Review Board Statement:** Not applicable.

**Data Availability Statement:** The data presented in this study are available on request from the corresponding author.

**Acknowledgments:** The authors are thankful to the Director, ICAR-Sugarcane Breeding Institute, Coimbatore, for the constant encouragement and support for carrying out the research work.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


**Disclaimer/Publisher's Note:** The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

### *Article* **Technology of Automatic Evaluation of Dairy Herd Fatness**

**Sergey S. Yurochka 1, Igor M. Dovlatov 1,\*, Dmitriy Y. Pavkin 1, Vladimir A. Panchenko 2, Aleksandr A. Smirnov 1, Yuri A. Proshkin <sup>1</sup> and Igor Yudaev <sup>3</sup>**


**Abstract:** The global recent development trend in dairy farming emphasizes the automation and robotization of milk production. The rapid development rate of dairy farming requires new technologies to increase the economic efficiency and improve production. The research goal was to increase the milk production efficiency by introducing the technology to automatically assess the fatness of a dairy herd in 0.25-point step on a 5-point scale. Experimental data were collected on the 3D ToF camera O3D 303 installed in a walk-through machine on robotic free-stall farms in the period from August 2020 to November 2022. The authors collected data on 182 animals and processed 546 images. All animals were between 450 and 700 kg in weight. Based on the regression analysis, they developed software to find and identify the main five regions of interest: the spinous processes of the lumbar spine and back; the transverse processes of the lumbar spine and the gluteal fossa area; the malar and sciatic tuberosities; the tail base; and the vulva and anus region. The adequacy of the proposed technology was verified by means of a parallel expert survey. The developed technology was tested on 3 farms with a total of 1810 cows and is helpful for the non-contact evaluation of the fatness of a dairy herd within the herd's life cycle. The developed method can be used to evaluate the tail base area with 100% accuracy. The hungry hole can be determined with a 98.9% probability; the vulva and anus area—with a 95.10% probability. Protruding vertebrae—namely, spinous processes and transverse processes—were evaluated with a 52.20% and 51.10% probability. The system's overall accuracy was assessed as 93.4%, which was a positive result. Animals in the condition of 2.5 to 3.5 at 5–6 months were considered healthy. The developed system makes it possible to divide the animals into three groups, confirming their physiological status: normal range body condition, exhaustion, and obesity. By means of a correlation dependence equal to R = 0.849 (Pearson method), the authors revealed that animals of the same breed and in the same lactation range have a linear dependence of weight-to-fatness score. They have developed an algorithm for automated assessment of the fatness of animals with further staging of their physiological state. The economic effect of implementing the proposed system has been demonstrated. The effect of increasing the production efficiency of a dairy farm by introducing the technology of automatic evaluation of the fatness of a dairy herd with a 0.25-point step on a 5-point scale had been achieved. The overall accuracy of the system was estimated at 93.4%.

**Keywords:** dairy cows; body condition score; 3D TOF sensor; non-contact evaluation; recognize area of interest

#### **1. Introduction**

Over the last few decades, the global trend in dairy farming has been to automatize and robotize milking processes on commercial farms [1,2]. The common average production period of dairy animals is 3.5 lactations [3]. Due to the rapid development of dairy farming,

63

**Citation:** Yurochka, S.S.; Dovlatov, I.M.; Pavkin, D.Y.; Panchenko, V.A.; Smirnov, A.A.; Proshkin, Y.A.; Yudaev, I. Technology of Automatic Evaluation of Dairy Herd Fatness. *Agriculture* **2023**, *13*, 1363. https:// doi.org/10.3390/agriculture13071363

Academic Editors: Jibo Yue, Chengquan Zhou, Haikuan Feng, Yanjun Yang and Ning Zhang

Received: 31 May 2023 Revised: 3 July 2023 Accepted: 4 July 2023 Published: 8 July 2023

**Copyright:** © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

new technologies are increasingly required to achieve a higher economic efficiency and achieve an improved production [4–6]. On the one hand, intensive production results in an increased milk yield of a cow; on the other hand, intensive production leads to the rapid deterioration of dairy cows—i.e., a reduction in the number of lactations [7]. The reduction of the production life of dairy animals also depends on the premature culling of animals that have a high or low body condition score. Lack of a normal body condition score during lactation is primarily due to dietary deficiencies [7]. Another negative consequence is culling of animals due to poor body condition because an increased body condition score reduces fertility and thereby extends the service period.

A body condition score (BCS) evaluation is important in technological milk production. First and foremost, the BCS score is used to place animals within productivity groups and determine their status. In Russian dairy farm conditions, veterinarians and livestock breeding technicians rotate animals into production groups once a month, provided the milk production technology is well established. The body condition score helps make a decision individually for each cow, based on her current physiological condition, rather than simply on accepted technological norms. In intensive milk production, dairy cows are divided into 5 main groups: group 1, the step-ladder milk yield increasing group, includes new cows from 6 to 100 days after calving, and also cows with a daily milk yield of more than 24 kg per head per day. The total productivity of this group of animals should not be lower than 6000 kg per head per year. The main objectives of the group are: quality feeding with full-fat mixes and good care to achieve the peak milk production by day 40–50; elimination of post-calving complications to inseminate the animals on day 65. During this period, the animals give up to 65–70% of their milk volume during the lactation period. High-yielding cows are transferred to group 1 and should be in group 2, but they need increased nutrition according to milk yield and body condition score. The typical fatness score for group 1 is 3.5 to 3.25 from day 6 to day 30, and 3 to 2.75 from day 31 to day 100. The normal decrease of the body condition score of cows in group 1 is due to an intensive milk production, which requires a large amount of energy. The energy expended cannot be fully compensated by the energy gained from feeding. Therefore, it results in a natural decrease of the body condition score. Maria Ledinek et al., in a study [8], showed that during the calving period, the body condition score decreases, and body fat reserves provide for an increased milk production. By 40–65 days after lactation, animals should be milked as often as possible, and the body condition score should not be reduced by more than 0.5. During this period, the cow consumes up to 12 kg of high energy feed. Insemination takes place when the animal is at peak production and consumes the highest amount of feed and the fatness score is within 3 points. At the same time, the animal's body condition score may not deviate by more than ±0.25 points.

Group 2—milking cows from 101 to 305 days after calving with 24 to 16 kg of milk per head per day. The main objective for the animals in this group is to ensure that the milk yield does not fall by more than 9% per month, and to increase the body condition to 3–3.5 fatness points.

Group 3—milking cows from 101 to 305 days after calving with a milk yield below 16 kg/head/day. The main task for this group is to prevent diseases, correct body weight to a fatness of 3.5–3.75 points and prepare the animals for drying off.

When animals are in the second and third physiological group, from 101 to 305 days after calving, it is necessary to monitor their condition. A cow should have 3.5–3.75 by the start of the dry period. If it is under-conditioned, it should be kept in the first or second group and its milk production should be ignored. Otherwise, under-conditioning can lead to complications during parturition or at the beginning of the next lactation [9]. Overconditioning of a cow above 3.75 will result in an increase in fetal weight. As the cow's weight increases, so does the calf's weight. An increased calf weight at calving causes birth complications and injuries that are equally detrimental to the cow and calf.

Group 4—the first 45 days from day 306 after drying off. During this period, no adjustment is made to the animal's body condition score. It is assumed that the animal already has a body condition score of 3.5–3.75.

Group 5 is the maternity group. The animals are kept 15 days before calving and 5 days after calving. During this period, no adjustment is made to the body condition score of the animals.

The fatness assessment of dairy cows is not only a valuable indicator for evaluating the quality of feeding and the response of the animal to feed, but is also an indirect indicator of its reproductive function. Dairy performance correlates with feed intake. An increase in milk production is associated with a decrease in fertility. During peak lactation, cows require 3.5 times more protein and energy for milk synthesis than protein and energy for life support, as lactation and calf feeding have a higher biological priority than body weight gain and fecundity. At peak milk production, quit estrus and overcalving are the most significant problems. A negative energy balance, which is also affected by decreased body condition dynamics, results in a delayed onset of first heat and ovulation after calving in underfed cows, reduced probability of conception after first insemination, negative effects on follicle growth, corpus luteum function, oocyte quality, impaired intrauterine development, and embryo survival and growth [10].

Cows with a body condition at day 60 of 3.25–2.75 have a 67% chance of conception, and those with a body condition below 2.75 have a 44% chance [11].

Mohamed A.B. Mandour, in a study [12], found that high fatness in first-year heifers increases the risk of ketosis to 3.71%, which is twice as high as in adult cows. The study mentions that cows with a high body condition score consume less feed than cows with a normal body condition score and have a high negative energy balance due to a higher concentration of fatty acids in the plasma, which is associated with an increased risk of ketosis.

Thinawanga Joseph Mugwabana et al., in a study [13], found no relationship between the fatness of cows and the calving rate.

Wynnton C. Meteer [14] found in their study that animals given 70% of the required feed energy had more embryos at the next insemination and a higher probability of insemination than animals that received an energy excess of 130% of the norm. Changing the level of feeding in animals in groups 4 and 5 (middle and late stage before calving) did not significantly affect the amount of pregnancy hormones excreted in the blood.

These studies confirm the above information that the main management of feeding, control, and changing the body condition score of cows should be done during lactation, in animals in groups 1 and 2, to increase the probability of reproductive success in the next insemination of animals.

Poczta W. et al., in their study [15], established a relationship between cow fatness and the likelihood of subclinical ketosis, where cows with a fatness score ≥ 3.25 were more susceptible to the disease than lean cows with a fatness score ≤ 3.

Vanholder T. et al., in a study [16], found a relationship between the body condition score of dairy cows and weight loss within 30–40 days after calving. Of the 47 cows studied, 37 cows lost ≥ 0.75 BCS points at 14 days post calving, and 10 cows lost ≤ 0.75 BCS points. Weight loss is associated with a negative energy balance in the cow after calving and subsequent mobilization of body reserves for recovery.

In [17], the authors found a correlation between the propensity for metritis and the BCS of cows ≤ 3. In [18], the authors evaluated the relationship between BCS points during the transition period and the development of disease and changes in milk yield. A total of 232 cows were assessed and the fatness was scored from 1 to 5 in a 0.25 step. After a blood test, a conclusion on the health of the animals was made. Changes in the body condition of dairy cows using the BCS scale were measured at 21 days before calving and 21 days after calving. The percentage of cows that increased BSC (fatness) during this period was 28%, lost BCS—22%, and retained BCS—50%. Additionally, 18% of the cows that lost BCS during this period had health problems compared to the cows that retained the BCS points. Furthermore, 28% of the cows that had an increased BCS were less likely to have subclinical ketosis.

The results confirmed that developing ketosis can be detected in an automatic, noncontact method. An alternative way of detecting ketosis is presented in studies [19–21], where blood tests were required to detect disease. On large farms with more than 200 milking herds, the continuous active assessment of animal health by blood testing is not possible, due to the lack of specialists, the time-consuming process, and the need for laboratory equipment. The BCS can be evaluated both manually and automatically.

Studies [18–22] describe the manual method of BCS evaluation. Study [23] gives a detailed review of automatic systems for automatic BCS evaluation. Study [24] describes the development of an automatic BCS evaluation system using a deep learning neural network algorithm using a convolutional neural network. The researchers achieved a recognition accuracy of 94% at a step of 0.5, and 78% at a step of 0.25. In [25], the authors used a convolutional neural network (CNN) to evaluate BCS. The accuracy of the system was assessed using the Kappa index and was within a moderate range (values between 0.41 and 0.60). In [26], the authors also used a convolutional neural network (CNN) to evaluate BCS. The accuracy of the results obtained in the study was 78%, indicating a successful real-time classification. In [27], the authors used the point cloud method to evaluate BCS. Experiments show that the proposed BCS evaluation model achieved an accuracy of 49, 80, and 96% within a deviation of 0, 0.25, and 0.50 points, respectively.

In [28], a dynamic background model (Gaussian Mixture Model, GMM) was used to distinguish the cow from the background. Subsequent Image Processing Algorithms have made it possible to automatically obtain reliable images, to find areas of interest, and to extract image elements without any manual intervention. With 5-fold cross checking, the model has achieved an average accuracy of 56% with a 0.125-point variance, 76% with a 0.25-point variance, and 94% with a 0.5-point variance.

Having studied the modern experience of the world community on the automation of BCS evaluation, our team had set a goal and fully fulfilled the tasks on the development of technology of an automatic system of BCS evaluation. The aim of the research was to improve the production efficiency of dairy farms by implementing the technology of an automatic BCS evaluation of a dairy herd with a 0.25-point step on a 5-point scale.

The main approach we used in developing the technology was to minimize the use of neural network algorithms to find areas of interest. This decision was based on the experience of the team [29,30]. Training neural networks was a labor-intensive and timeconsuming process. For example, a trained neural network for standardized breeds of EU countries—Holstein, Brown Latvian, Swiss, etc.—will give a big error during a BCS evaluation of Black-Motley Holstein, Kalmyk breeds, etc. To minimize the error, it was necessary to retrain the neural network. The method we proposed is based on the study of standardized breeds of EU countries to adjust the model and to carry out a further BCS evaluation, avoiding the training of the neural network algorithm on each farm.

Thus, the research resulted in the development of a universal automatic system capable of estimating the BCS of an animal with high accuracy (more than 90%) at a step of 0.25. The developed system was intended for the implementation in automated and robotic free-stall farms. The system was designed to evaluate the BCS (fatness) of animals and to provide recommendations for a wider range of functions to be carried out by a specialist.

#### **2. Materials and Methods**

#### *2.1. Farm, Field Data Collection*

In earlier studies, we had already achieved a result of algorithmic evaluation, where the system evaluated a fatness score between 2 and 4 with a 10% error, while scores 1 and 5 were evaluated with a 25% error (the results of the automatic evaluation were checked against the results of an expert panel) [29,30]. In the study we conducted, the unsatisfactory result that required further research was on the evaluation of the boundary body condition scores 1 and 5. The difficulty lies in the fact that for the algorithm, cows with a body condition score of 1 and 2 and a score of 4 and 5 are similar. Therefore, in this study, we focused our field data collection on animals with a body condition score of 1 to 3 and 4 to 5.

We selected 3 commercial farms with a total of 182 animals. On the selected farms, all cows have similar traits, the animals are emaciated and of poor performance, and part of the herd features are overweight. Data on animals were collected in the Moscow and Yaroslavl regions.

Field data were collected between August 2020 and November 2022:


All animals had two milkings per day. Expert panels were formed to assess the body condition score by data collection site. The panel consisted of at least two independent veterinarians and two trained specialists. The average body condition score obtained from all experts was the benchmark value.

In addition, a weighing platform was used as an implement to further increase the accuracy of the body condition score by comparing the values obtained.

The live weight of animals was collected, in particular, by the Klüver–Strauch method [31], and the remaining animals were weighed on the platform. A disadvantage of the Klüver– Strauch method is an error in measuring the live weight of up to 10%. In the experiment, a discount of 1% of the actual weight was taken into account for bulk (mud adhered to the animal), and a discount of 3% for the contents of the gastrointestinal tract, when animals were weighed on the platform. Calculation of the live weight, including the discounts made, was automatic.

Based on the Pearson's correlation coefficient, we obtained proof of the representativeness of the animal samples and the relationship between the body condition score and live weight of the animals.

#### *2.2. Equipment and System*

Images of the cows' backs were collected on a 3D commercial camera, the O3D 303 3D ToF camera. The system is powered by 220 V, the power supply converts to 24 V to ensure the 3D sensor is operational. The sensor is mounted at a height of +2.2 m above the floor level (Figure 1).

**Figure 1.** Demonstration of how the distance of the scanned surface varies with the inclination angle of the sensor. (**a**) Position 1 of the optical module 03D "looks" vertically down relative to the back of the cow; (**b**) position 2 of the optical module 03D "looks" at an angle of 5◦ from the vertical axis; (**c**) position 3 of the optical module 03D "looks" at a 10◦ angle from the vertical axis.

The height and inclination angle of the three-dimensional sensor are based on four parameters: cow height, cow length, minimum working distance between the camera and the object, and the camera's allowable error. The height of the animals under study ranged between 1300 mm and 1500 mm, the minimum working distance of the camera between the surface and the object under study was 300 mm. The signal from the identification antenna of the animal's RFID tag triggered the three-dimensional image production.

The inclination angle of the sensor taking into account the given 4 parameters is chosen to be 5 degrees, as it can cover a sufficient area of the animal's back under analysis, while keeping the pixel spacing to 0.006 m as the point of interest moves away from the 3D camera. The distance of 0.006 m between pixels is the set distance on which the least squares method is based when forming clusters of points related to areas of interest.

For the correct calculation of the required parameters between the camera and the object under study (coordinates of the received *Z*-axis pixels), we performed angle normalization (because the tilt angle of the 3D camera relative to the cow's back was introduced), presented in the expression using the R matrix:

where X, Y, Z—the areas of interest point coordinates, and J—the required distance between the interest points areas.

The total dataset contained 546 images from 182 animals with body condition scores from 1 to 5 with a step of 0.25 points (17 classes) (Table 1). Based on the earlier studies, it is sufficient for this camera to take three pictures of each cow, then the images are combined and the system starts determining the fatness.


**Table 1.** Number of images and proportion of cows for each body condition score.

\*—the number of animals; \*\*—the number of images.

From the data collected, we can see that the predominant body condition scores are 4.75; 4.25; 3.25; 3; 2.5; 2.25; 2; 1.5. The distribution of animals by body condition score was made by the expert panel, whose opinion is considered to be the benchmark (Figure 2).

The 3D camera is able to calculate and output Point Cloud as a multidimensional array I × J × K, where I and J are camera resolution, e.g., 352 × 264, K is X, Y, Z coordinates. Output of received data is in "dat" and ".h5" formats. The recording speed of the video images is 5 fps. Due to this feature, we obtained 3 to 5 images of each cow in the initial image. The images were collected according to the scheme shown in Figure 1. The camera

error stated in the manufacturer's specifications is ±0.01 m for each meter between the lens and the object. Therefore, assuming that the working distance between the cow's rump (1.5 m) and the 3D camera lens (2.2 m) is 0.4 m, the error amounted to ≤0.01 m. The optical module was installed at an angle of 5◦.

**Figure 2.** Developed installation used for field data collecting. (**a**) Scheme of the developed installation to determine the body condition score, height, and weight of dairy cows up to 1200 kg: 1—automatic gates; 2—weighing module; 3—03D 303 three-dimensional camera; 4—a single control unit; (**b**) three-dimensional camera for the body condition score evaluation—03D 303 and software.

#### *2.3. Assessment of the Body Condition Score and Analysis of the Results*

We used our previously developed software [32] to process the obtained threedimensional maps and determine the body condition score and standard tools; excel for primary data processing and formatting was used to process the study results.

The results were obtained automatically and those of an expert evaluation were compared manually. The expert evaluation of the body condition score was a benchmark value.

In terms of searching and determining the main areas of interest, the developed software was based on the application of the least squares method (regression analysis) to find the areas of interest.

As the camera was mounted on top of the animal and the points of the cow's back were presented to the data analysis, the points of greatest interest were those near the contour and describing its perimeter. Using the spine of a cow as an example, we can consider the basic expressions to identify it. The algorithm developed is based on the ordinary method of least squares (LS).

The entire surface of a cow's back is represented by an array of points without regard to depth, after which the regression tool is applied. We represent the whole surface as a set of points:

$$((y\_1, x\_1), (y\_2, x\_2), \dots (y\_{n\_\prime}, x\_n))\tag{1}$$

We can apply the method of least squares to minimize the sum of squares of RSS RRS deviations:

$$\text{RSS} = \sum\_{i} \left( y\_i - (a + bx\_i) \right)^2 \tag{2}$$

To find fixed points for RSS, the following expressions are used:

$$\begin{cases} \frac{\partial RSS}{\partial a} = \sum\_{i} 2(y\_i - a - bx\_i) = 0\\ \frac{\partial RSS}{\partial b} = \sum\_{i} 2(y\_i - a - bx\_i) = 0 \end{cases} \tag{3}$$

$$\begin{cases} \sum\_{i} y\_i - na - b \sum\_{i} \mathbf{x}\_i = 0\\ \sum\_{i} \mathbf{x}\_i y\_i - a \sum\_{i} \mathbf{x}\_i - b \sum\_{i} \mathbf{x}\_i^2 = 0 \end{cases} \tag{4}$$

$$\begin{cases} \overline{y} - a - b\overline{x} = 0\\ \overline{x}\overline{y} - a\overline{x} - b\overline{x^2} = 0 \end{cases} \tag{5}$$

$$\begin{cases} \begin{aligned} a &= \overline{y} - b\overline{x} \\ \overline{x}\overline{y} - (\overline{y} - b\overline{x})\overline{x} - b\overline{x^2} &= 0 \end{aligned} \tag{6} \end{aligned} \tag{6}$$

$$\begin{cases} \begin{aligned} a &= \overline{y} - b\overline{x} \\ \overline{x}\overline{y} - \overline{x}\overline{y} + b\left[\left(\overline{x}\right)^2 - \overline{x^2}\right] &= 0 \end{aligned} \tag{7} \end{cases} \tag{7}$$

Thus, the regression and refinement of the ridge line to the point cloud produces the result shown in Figure 3.

**Figure 3.** Defining the spinal column axis with extracting the area of interest. 1—Filtered area; 2 cow's contour; 3—unspecified animal's spinal column; 4—specified animal's spinal column; 5—tail head; 6—hips. (**A**) initial ridge line plotting by linear regression. (**B**) the ridge construction as a set of points on each longitudinal axis.

Figure 3 showed the results of the regression method. Figure 3B showed the ridge construction as a set of points on each longitudinal axis constructed. The lighter silhouette shows the silhouette of the cow, represented as a cloud of points, disregarding the *Z*-axis. Figure 3A is an initial ridge line plotting by linear regression.

To determine the cow's height, it is necessary to estimate the coordinates (xyz) of each point along the ridge line and find the extremum along the *Z*-axis. The point that is the extremum is the withers from which the cow's height is determined.

Table 2 shows the two approaches to BCS evaluation, the upper part was used for BCS evaluation by the expert panel, the lower part of the table was used for automatic evaluation. The numerical values were determined manually by analyzing the resulting field database of animals. The numerical characteristics are the average values for each body condition score and are relevant for the black-motley and the Holstein black-motley breeds raised in Central Russia. For other breeds, the numerical characteristics described in Table 2 may differ [32–34].

*Agriculture* **2023**, *13*, 1363


When the system evaluates the spinous processes of the lumbar and back, and the transverse processes of the lumbar and dorsum, the system first draws straight lines along the ridge and parallel to the ridge lines in the area of the transverse processes of the lumbar then measures the pixel height along the lines (Table 2, side view, node B, parameter h1).

In the hip's area, the system assesses the angle: two lines are drawn along the protruding parts of the back and then the angle is assessed (Figure 4, fatness score 3). The angle at 136◦ is an indication of a body condition score of 3, and the angle 125◦> is a fatness score of 1.75 to 1.

**Figure 4.** Desired areas of interest.

The points for estimating the angle are plotted on the boundary of the protruding parts of the body: the rump bone is A1, the hip protrusion is A2, and the first point at the junction of the transverse processes is A3. To find the point A3, we applied neural network tools with the preliminary training on 80 animals in the 5–6 months of lactation with a body condition score of 1–3 points.

To determine the depth of the "hunger hallow", the following procedure was used: step 1—point XN1/3, which is 1/3 of the length of the segment XN; step 2—at 1/2 the length of the segment BXN1/3, set point h2. Then, we compare the difference in height between points h2. The depth of the "hunger hallow " for a 5-point animal is 0.06 m and for a 1-point animal the depth of the hunger hole is 0.12 m.

The h3 points are determined by the lowest point in the tail base and the highest tail base.

As the last step before determining the fatness, the system checks all criteria and determines the body condition score on a 5-point scale in 0.25-point steps.

Figure 5 shows three-dimensional images converted into the black-and-white format. The pictures show animals with a body condition score from 1 to 5 on a 5-point scale and an explanation of which area of the animal's back is manipulated by the algorithm.

**Figure 5.** Animals' body condition scores and areas of interest. 1—Spinous processes of lumbar and back/dorsum; 2—transverse processes of lumbar and hunger hollow area; 3—hips and pin bone; 4—head of tail; 5—vulva and anus area.

#### **3. Results and Discussion**

#### *3.1. Results*

To understand the developed system's overall effectiveness, it was necessary to analyze and evaluate the effectiveness of each area of interest. All the resulting field data were evaluated using the developed method. The results were compared with the experts' evaluation. To understand the overall effectiveness of the developed system, it was necessary to analyze the evaluation effectiveness of each area of interest. To this end, a graph was plotted (Figure 6). The graph shows in the percentage terms the areas of interest and their detection probability, where 0% was not detected in all animals and 100% was detected in all animals.

**Figure 6.** Efficiency of the system when detecting the areas of interest in the studied animals. (**a**) measurement efficiency of detecting the areas of interest; (**b**) difference in the BCS evaluation between the automatic measurement of the developed system and the evaluation made by the expert group and the difference between the obtained values.

The graph analysis results of Figure 7 show that the developed method can estimate the tail base area with the 100% accuracy. The hunger hollow is determined with a 98.9% accuracy and the vulva and anus area with a 95.10% probability. Protruding vertebrae namely, spinous processes and transverse processes—are evaluated with a 52.20% and a 51.10% accuracy. The accuracy of 50% was explained by the fact that according to Figure 2, these areas were not determined or were determined incorrectly in animals with a body condition score ranging from 3.25 to 5. The overall accuracy of the system was estimated by the experts at 93.4%, which was a positive result.

**Figure 7.** Distribution graph of the fatness of 5–6 month old animals obtained in an automatic evaluation.

Additionally, Figure 6 showed that the evaluation of the system and that of the experts have more discrepancies when the body condition score is 4–5, with the largest error of 1.25 and the smallest error of 0.25.

Figure 7 shows that animals with a condition score of 2.5 to 3.5 at 5–6 months are healthy. The developed system gives reasons to divide the animals into three groups, confirming their physiological status: normal range body condition, exhaustion, and obesity. In this case, it is worth bearing in mind that the system has an accuracy of 93.4%. Then, in this study, 4 animals with a body condition score of 3.75, and 15 animals with a body condition score of 2.5 had a 6.6% probability of belonging to another physiological status group. This is due to the fact that the system was wrong by 0.25. Errors caused by other nutritional scores are not critical, as technologically, an animal is either healthy and does not require any manipulation even though the system gave a nutritional score of 3 ± 0.25, or it has exhaustion/obesity, which requires manipulation of the animal to improve its physiological status.

In our observations, most of the animals with a fatness score of 3.75–5 were on the second farm (percentage of the total herd when ranked by score) with an average annual milk yield per cow per day of 15 kg/milk and a fat content of 5.1–6%. This farm was financially sound, and the main activity was getting milk from the animals for cheese production. When analyzing the cause of overweight animals, it was found that the farm staff were disrupting the feeding ration and the animals were receiving more energy than they needed; the animals were kept in loose housing.

On the second farm, a correlation was established between live weight and body condition score for 32 animals (Figure 8). The correlation determined by Pearson's method is R = 0.849.

**Figure 8.** Distribution graph of the fatness of 5–6-month-old animals obtained during an automatic evaluation in 32 animals.

The Pearson correlation tool was chosen to determine if there was a relationship between the live weight and body condition score of the cows under study, as the data obtained have a normal distribution. Discussion of the data shows that part of the values on the scale from 4 BCS points to 5 BCS points and a live weight above 525 kg have a chaotic distribution. The Pearson correlation is R = 0.849, which does not guarantee 100% correlation. This is explained by the following: the live weight of animals consists of basic parameters—the amount of dirt accumulated on the animal, the amount of feed eaten, and the month of pregnancy. Additionally, live weight was obtained using the Klüver–Strauch method [33], where the method itself has a margin of error. This correlation did not allow the evaluation of live weight by the fatness score, but is only an additional signaling indicator that draws attention to live weight gain. Therefore, by analyzing the data by the Pearson correlation, our main aim was to understand if there is a relationship between obesity and weight gain. This was important for the purpose of additional animal monitoring, where the developed software will signal if an animal is overweight, which in turn negatively affects the probability of successful insemination. If it was detected that an animal is gaining excessive live weight, it was therefore necessary to move the animal to another housing group to change the feeding ration.

Studies [33] found that an optimal range of body weight for an increased performance does exist due to the non-linear relationship between milk yield and body weight. Dairy breeds respond more strongly to bodyweight range than dual-purpose breeds. Cows with an average weight are the most productive in the population. Heavy cows (>750 kg) produce much less milk. Special attention should, therefore, be paid to the daily ration, and further increases in body weight of dairy cows should be avoided. Animals with a body condition score of 1 to 2.5, in most cases, were found on a farm with an average annual milk yield per cow per day of 16.8 kg/milk, and a fat content of 3.6 to 3.8%. After examining the keeping conditions of the animals, several criteria influencing the emaciation of the animals were observed. The main criterion was the feed ration. The animals under study received mainly legume–grass hay with the addition of micro and macro nutrients in their diets. The animals were continuously fed a complete daily ration consisting of 4 kg of legume–grass hay, 15 kg of mixed grass silage, 6 kg of root crops, 5 kg of high energy mixed fodder, 1 kg of barley powder, and 50 g of table salt. In addition, it was recorded that animals were kept in concrete buildings, typical for buildings constructed in the 1980s, with a disturbed microclimate and tethered housing without regular walks.

Based on the research results, the algorithm for the automatic evaluation of the animal body condition (fatness), followed by the staging of their physiological condition, was supplemented and modified. The algorithm was divided into two parts and is shown in Figures 9 and 10. The second part of the algorithm is an integral part of the first one. The algorithm was included in the software code of the developed software.

**Figure 9.** First part of the BCS Algorithm.

**Figure 10.** Second part of the BCS algorithm.

The explanation of Figure 9 starts with a three-dimensional map containing X, Y and Z coordinates of each pixel, and then searches for points at height (in the range of 1–1.6 m from the floor). When a cluster of points is detected, the algorithm determines the location of the main features: the spine, the hips, the tail head, and the vulva region. The topography of the spinous and transverse processes relief was then determined and the initial BCS value was determined.

As for Figure 10, for the initial BCS of 2.25–5 with no scalloping of the spinous and transverse processors and a low depression, the difference in tail head height was measured. For the primary BCS values in the range 1–3 with spinous and transverse processes in relief and a large hollow volume, the spinous and transverse processes of the lumbar and dorsum were measured. Depending on the results, the algorithm outputs had three evaluation options—'normal BCS', 'obesity', 'exhaustion'.

The developed software gave information about the animal by scanning RFID tags: date and time of the last measurement, sex, status, date of birth, actual weight, weight excluding animal contamination (bulk), weight excluding GIT contents, animal height, and the BCS value. The software developed for each animal provided more detailed information, traced the dynamics of changes in physical parameters, kept the herd log, and had service settings.

#### *3.2. Research limitation*

We refer to several factors as limitations of the research.

Factor 1 is the capability of the machinery and equipment and the environmental conditions in which they were operated. Dairy cows were evaluated both indoors and in the open field. Based on our experience with the equipment, we have found that 3D TOF cameras with a 840 nm wavelength, when shooting animals outdoors in bright sunlight, had noise that prevented effective fatness scoring. As such, 3D TOF cameras at 940 nm may be considered for further research. According to the manufacturers (the study does not specify a specific manufacturer), the 940 nm 3D cameras solve the problem of not being able to produce 3D maps in bright sunlight. In this study, three-dimensional cameras based on 940 nm were not tested.

Factor 2—while evaluating the body condition score of an animal, we could not estimate the animal's weight to an accuracy of 1 kg. We considered it possible to install an additional 3D camera to measure the torso depth of the animal—automatically using the Klüver–Strauch method [32] based on the digital data obtained for the torso depth, height and body condition score. However, this method would also not give accurate information about the animal's live weight, as there was no information about the cow's pregnancy, degree of contamination, and gastrointestinal contents. Additional discounts and coefficients relative to live weight would have to be introduced, but this may result in a high margin of error.

Factor 3 is the use of artificial intelligence to find the areas of interest. The matter is that if all fields of interest are calculated by means of artificial intelligence, then the exploitation of the system with each new breed or farm will demand resources for additional training of the system, which is not practical. Using the proposed research method for code development is a more labor-intensive process than collecting a data array and training artificial intellect. However, this method would definitely prove to be more practical, because it covered dairy cattle breeds, which are bred in Russia.

Factor 4 is the use of specific equipment. For these studies, it was not the specific manufacturer of the 3D cameras that was important, but their characteristics. It was important to choose a 3D camera that has a wavelength of 840 nm and a resolution of 352 × 264, and the factory error rate of used cameras is not higher than 1 cm per 1 m distance at a distance from the object in question.

#### *3.3. Economic Efficiency*

The proposed technology will improve production efficiency on large dairy farms by reducing animal stress, controlling animal nutrition when necessary, and early detecting physical deviations (Table 3).


**Table 3.** Cost of implementing the technology using the example of the farms under study.

The percentage of culling and mortality was planned to be reduced by adjusting the ration and improving the general maintenance condition of the animals on the farms. We also proposed to increase the actual milk yield per day.

Often, farms have in-house veterinarians, but with the introduction of the biometric system, costs can be reduced, and external specialists can be called in only when necessary. Feed costs would also be reduced, as feed rations for the animals can be monitored and adjusted.

The main profit increase was expected to come from the improved life quality of the cows, and as a consequence, the birth rate of calves will also increase.

Sales are planned by age groups. The distribution will be as follows: 80% of all calves born on the farm during the year will be sold. Of these, 50% will be sold at the age of 1 month, 35%—6 months, and 15%—12 months.

As far as in the first year, Farm 1 and Farm 3 would make 38.4% and 43.8% more profit, respectively, but this technology did not look profitable on the second farm. We recommend that this biometric system should only be installed on large farms with 560 heads or more.

#### *3.4. Technology Applicability*

Having confirmed the cost-effectiveness of the developed BCS estimation technology, we can now describe how we implement automatic BCS evaluation for milk production.

The automatic livestock monitoring system operated in two ways. The first way was stationary, and the second way was mobile.

The stationary method consisted in the fact that on the farm, in the places of the daily pass of the animals, for example, the system of BCS evaluation was mounted behind the

milking parlor in the "gallery". The system consisted of a three-dimensional camera and data collection and processing unit, as well as an identification antenna, which read the ID number of the cow. The data were sent to a server.

The mobile method implied bringing the system once a month to the box where a group of animals was kept to scan the BCS score (Figure 11).

**Figure 11.** Schematic diagram of the installation of a mobile BCS evaluation system in a cubicle housing a group of animals.

The system was brought in by a forklift or an ATV to the cubicle where the animals were kept in a loose housing. Two staff members then turn the system around; they set the fence in the desired position, pointing to one side or the other. Then, one employee drove the animals in and out of the system, a second employee took care of reading the cow number, assessed the body condition, and recorded the data. When the group was finished, the system was assembled in the transport position and transported to the next group of animals. The data were transferred on a flash drive to a server. This procedure is done on a monthly basis. The advantage of the mobile system is that the fatness estimation can be done while the animals are grazing in the fields.

On the basis of the data obtained, the developed software plots a graph—a diagram of the change in the BCS score—and compares it with the set-required values for the current physiological status of each cow. There were several applications of the technology. The first situation was when we had an animal with an increased body condition score. The system recorded that the BSC score was increased, then queried the following data from the herd management software: current physiological status, which group the animal is in, current milk production, day of lactation, insemination status, and specific ration. For example, an animal was on day 75 of lactation, no conception had occurred, the BCS score was 3.75, milk yield was 17 kg/day, and fatness was 3.7%. Then, an automatic decision was made that the ration should be adjusted by reducing the amount of energy the animal receives without changing the animal's maintenance group, as the animal was at the peak of lactation and its milking requirements should be met. When moving to the next group, a gradual decrease in milking should be observed, accordingly. At the same time, we have to monitor the animal's condition so that by the end of lactation, the animal has a corrected condition. For example, an animal on day 190 of lactation and the conception on day 110, the BCS condition score was 2.5, milk yield 14 kg/day, and fat content 3.5%. In this case, the animal should be moved from group 3 to group 1 or 2 in order to adjust the feeding level to ensure an energy surplus.

The BCS evaluation system was needed as an additional tool to monitor feeding and assist in decision making for each cow when moving them to different housing groups. The development of an automatic fatness estimation system will make it possible to collect data sets and statistics for each animal. This will make it possible, when collecting data on feeding, animal genetics, breeding material, and diseases, to form animal groups on farms more effectively, revealing their genetic potential in terms of productivity.

#### **4. Conclusions**

The effect of increasing the production of dairy farms had been achieved by implementing the technology of an automatic evaluation of the fatness of dairy herds (BCS) in a 0.25 step on a 5-point scale. The developed technology had been tested on 3 farms, with a total herd of 1810 animals, and provided for a non-contact BCS evaluation of a dairy herd required throughout the life of the herd within the farm. The overall accuracy of the system was estimated at 93.4%. The study has demonstrated the economic effect of implementing the proposed system.

**Author Contributions:** S.S.Y.—project management, methodology, natural data collection, writing editing. D.Y.P.—conceptualization, software development. I.M.D.—system development, writing analysis and editing. V.A.P.—visualization, rude preparation. A.A.S.—software testing. Y.A.P. natural data collecting, editing. I.Y.—natural data collecting, writing—editing. All authors have read and agreed to the published version of the manuscript.

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Conflicts of Interest:** The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.

#### **References**


**Disclaimer/Publisher's Note:** The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
