Smart Agricultural Technology 3 (2023) 100117 Contents lists available at ScienceDirect Smart Agricultural Technology journal homepage: www.journals.elsevier.com/smart-agricultural-technology Estimating body weight in conventional growing pigs using a depth camera Guilherme A. Franchi a, *, Jacinta D. Bus b, Iris J.M.M. Boumans b, Eddie A.M. Bokkers b, Margit Bak Jensen a, Lene Juul Pedersen a a Department of Animal and Veterinary Sciences, Aarhus University, Blichers Allé 20, 8830 Tjele, Denmark b Animal Production Systems Group, Wageningen University and Research, P.O. Box 338, 6700AH, Wageningen, the Netherlands A R T I C L E I N F O A B S T R A C T Keywords: Automated body weight (BW) estimation can be a useful tool for continuous monitoring of growth in commercial Sus scrofa pigs, whereas deviations could indicate welfare problems. We validated a depth camera for BW estimation in 251 Precision livestock farming conventional growing pigs on two farms. Scale-based BW of individual pigs was used as gold standard (Farm 1: Sensors 107 pigs, BW range: 16–130 kg, recorded on three days; Farm 2: 144 pigs BW range: 20–114 kg, recorded on nine 3D image days). The camera was placed above the individual feeding station (Farm 1) or multi-partitioned feeder (Farm 2) Weight gain Animal welfare and combined with a radio frequency identification system. Whenever a pig visited the feeding site, three- dimensional images were taken, and all individual daily images were used to calculate the median individual estimated BW. The pen estimated BW was calculated by taking the median of all daily picture estimates. A very high agreement (Concordance Correlation Coefficient >0.96) between scale-based BW and estimated BW was found on both farms at individual and pen level. Additionally, the individual-level and pen-level BW estimation errors of the median weight over the fattening period were low on both farms (≤3.6%). Yet, the camera’s BW estimation performance decreased in pigs weighing >110 kg on Farm 1. Whereas, on Farm 2, the performance decreased when pigs weighed approximately 60 kg and were subjected to a typical dietary change, which potentially increased the competition for access to the multi-partitioned feeder and, consequently, limited body boundary detection. 1. Introduction not done properly, leading to negative effects on the growth of young animals [12]. Furthermore, photogrammetric techniques typically In pig production, the body weight (BW) of pigs is a crucial indicator require several cameras positioned at different locations taking simul­ of growth and readiness for market [1,2]. Additionally, knowledge on taneous images of individual animals, which can be challenging to pig BW development is helpful to support animal welfare, including synchronise and expensive to implement on-farm. Whereas the animal health, and optimise production efficiency [3]. Typically, on image-analysis-based BW estimation can be an efficient, non-invasive commercial farms, pig weighing is performed manually [4], which is method, this approach relies on visible light and the outcome can considered the most accurate method to determine BW of animals [5]. consequently be affected by the room lighting and the status of the pig However, in the European Union, pig farms have become larger in herd skin (e.g., dark, stained, or dirty) [4]. To deal with these potential size over the last decades [6] and, hence, manual pig weighing involving sources of errors, recent studies explored the use of depth cameras based handling constitutes a time-consuming and relatively stressful proced­ on a structured infrared-light system, which provide three-dimensional ure to both pigs and stockmen [7]. (3D) images and eliminate errors arising from variations in skin colour Alternatively, BW of pigs can be indirectly determined from manual and ambient light [13]. As 3D images can account for animal height, measurements (e.g., girth size, withers height; [7,8]), photogrammetry they may lead to more accurate BW estimates than two-dimensional [9] and digital images (e.g., [10,11,2]) of pigs’ body dimensions, due to (2D) images (e.g., [4,14,13,15]). Yet, the methods proposed by the a high, positive correlation between body dimensions and body mass cited studies required profound physical changes in the farm, such as [7]. Nevertheless, manual body dimension measuring is laborious and modifications in the feeding site or pen corridor, to accommodate the can be dangerous for humans and stressful for the animals, if handling is equipment and avoid incomplete image viewpoint of the pigs and image * Corresponding author. E-mail address: amorimfranchi@anivet.au.dk (G.A. Franchi). https://doi.org/10.1016/j.atech.2022.100117 Received 12 July 2022; Received in revised form 6 September 2022; Accepted 9 September 2022 Available online 12 September 2022 2772-3755/© 2022 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). G.A. Franchi et al. Smart Agricultural Technology 3 (2023) 100117 quality issues arising from ambient light changes or unclear back­ 60 kg: 14.8% crude protein, Svin Struktur E; above 60 kg: 15.5% crude grounds. These physical interventions may be costly to farmers and protein, Svin Ideal Vox; both manufactured by DLG, Denmark), and the require some habituation period for the animals prior to the beginning of feeder containing three partitions were filled four times daily at 07:00 h, the weight estimation. 11:00 h, 16:00 h and 20:00 h. Water was accessible ad libitum in two Accordingly, we aimed to validate a commercially available depth drinking cups. All pens were equipped with dry and fresh wooden sticks camera providing 3D images for continuous BW monitoring of conven­ and a rubber ball as minimum pen enrichment in compliance with Eu­ tional growing pigs throughout the fattening period requiring minimal ropean Union and Danish animal welfare legislation. Artificial light was physical changes on-farm, hence requiring no habituation period for the on from 06:00 h to 21:00 h (182 lx). pigs. 2.2. Body weight recording 2. Materials and methods Two types of BW data were recorded: scale-based BW (gold standard) This study was performed on two locations. The first location was a and BW estimated by a depth camera (iDOL65, dol-sensors a/s, Aarhus, commercial farm in Gronau, Germany (Farm 1) between October and Denmark; technical information available on https://www.dol-sensors. November 2021, and the second location was the experimental farm of com/wp-content/uploads/2019/11/EN-iDOL-65-Camera-TI-20210720. the Department of Animal and Veterinary Sciences, Aarhus University, pdf). On Farm 1, the scale-based BW was recorded on three days Viborg, Denmark (Farm 2) between September and November 2021. throughout the fattening phase (BW range: 16–130 kg), first day in the Ethical approval was not required for the procedures involving pigs on morning and two last days in the afternoon (Table 1). On Farm 2, the any of the two farms, as data was collected during standard management scale-based BW was recorded on nine days throughout the fattening procedures and no additional measures were obtained for the study. phase (BW range: 20–114 kg), all in the morning (Table 1). On each farm, all pigs per pen were moved into the corridor of the farm and 2.1. Animals, housing, and management individually walked into a calibrated digital weighing scale (Farm 1: W- 2000, Welvaarts Weegsystemen, The Netherlands, accuracy: ±0.5 kg; On both farms, pigs were raised under conventional husbandry Farm 2: MTW2-STACON, Schauer Agrotronic GmbH, Germany, accu­ conditions. On Farm 1, 107 Landrace × Large-White pigs were housed in racy: ±0.3 kg). On Farm 1, barn staff scanned pigs’ radio frequency 10 identical pens (9–11 pigs/pen; ≥1.03 m2/pig; Table 1) spread across identification (RFID) ear tag and manually entered the individual BW on five rooms, on fully slatted floors (Fig. 1). Pigs were fed ad libitum with a a computer, whilst, on Farm 2, pigs’ RFID ear tag were read by the scale commercial dry feed (until farm average of 57 kg: 16.2% crude protein, and automatically recorded the individual BW. Pen-level BW was Select Delta 2; between 57-81 kg: 15.3% crude protein, Select Delta 4; calculated by taking the median of the individual BW of pigs per pen. above 81 kg: 13.8% crude protein, Select Delta 5; all produced by ABN, For the estimated BW, each pen was equipped with one depth camera United Kingdom) using an IVOG® electronic feeding station (Hokofarm placed above the individual feeding station (Farm 1; Fig. 1) or three- Group, The Netherlands). From a separate feeder, pigs could obtain partitioned feeder (Farm 2; Fig. 2) at an approximate height of 2.2 m fibre-rich feed (chopped straw mixed with straw pellets, in compliance and worked in combination with an RFID system installed in the feeding with European Union and German animal welfare legislation) ad libitum sites (Farm 1: one RFID antenna per feeding station; Farm 2: one RFID by manipulating a chain. Water was available ad libitum from two antenna per feeder partition). Camera calibration was automatically drinking nipples, and enrichment consisted of a dry and fresh wooden performed upon camera installation and consisted of the camera storing block, a chain with plastic rings and in some pens one or more hosepipes. its position and orientation in relation to the pen floor, as specified in its Pens were naturally illuminated through windows. technical user guide (dol-sensors a/s, Aarhus, Denmark; https://www. On Farm 2, 144 Yorkshire × Landrace pigs were housed in a finishing dol-sensors.com/wp-content/uploads/2019/11/EN-604438-iD pig unit including 8 identical pens (9–18 pigs/pen; ≥0.73 m2/pig; OL-65-Camera-TUG-20211203.pdf). On Farm 1, throughout the Table 1) with approximately 33% slatted, 33% drained and 33% solid fattening phase, cameras were cleaned when they gave an error due to floor (Fig. 2). Pigs were fed ad libitum with a commercial dry feed (until dirt on the lens. If a camera was moved during the cleaning process, it Table 1 Breakdown on the number of pens, number of pigs, number of pigs per pen, scale-based body weight (BW), estimated BW, and number of images per pig used to obtain weight estimates at individual level and scale-based and estimated BW at pen level on each location (Farm 1, Germany; Farm 2, Denmark) and day. Farm Day No. pens No. pigs No. pigs/pen Individual level Pen level Scale-based BW (kg/ Estimated BW (kg/ No. images/pig Scale-based BW (kg/ Estimated BW (kg/ pig) pig) pen) pen) Median IQRa Median IQRa Median IQRa Median IQRa Median IQRa b c d d d 1 1 9 98 10–11 24 21–27 — — — —d 24 23–25 24 23–25 (DE) 2 10 107 9–11c 73 68–80 74 67–81 256 201–327 73 71–74 73 71–74 3 9e 97 9–11c 105 100–112 104 98–109 205 154–252 105 102–108 105 102–108 2 1 8 144 18 28 26–32 29 26–33 206 71–379 28 27–29 26 25–27 (DK) 2 144 36 33–39 36 33–40 268 140–507 36 35–37 35 33–36 3 143 17–18f 42 38–45 42 39–46 582 375–783 42 41–43 40 38–40 4 143 49 44–52 48 45–53 286 186–439 48 47–49 48 46–49 5 143 56 52–60 55 51–61 272 168–426 55 54–57 54 52–55 6 139 16–18f 63 58–69 62 58–70 216 106–348 64 62–66 63 60–66 7 132 9–18f 72 65–77 69 64–76 116 70–223 72 69–74 71 69–72 8 131 9–18f 80 73–87 79 72–86 107 60–198 81 77–83 80 76–83 9 128 9–18f 89 82–96 84 78–92 112 71–206 89 88–92 87 85–90 a IQR=interquartile range. b One camera was not yet calibrated, and the data was unavailable. c All pens should contain 11 pigs. If only 10 or 9 were present, pigs had been removed due to sickness. d Due to data storage malfunctioning, individual images and BW estimates were not available. e One camera malfunctioned and the data was unavailable. f All pens started with 18 pigs. If pens had fewer than 18 pigs, pigs had been removed due to sickness. 2 G.A. Franchi et al. Smart Agricultural Technology 3 (2023) 100117 Fig. 1. (A) Illustration of the pen for fattening pigs on Farm 1. The single-space electronic feeding station is represented by the white rectangle, the iDOL65 3D camera by the white pentagon, the crude fibre station by the white circle, drinking cups by the black hollow circles, and the two solid black squares give the location of enrichment, which differed between pens. Half the pens were in the opposite configuration (feeder always located next to the door/corridor). (B) Illustration of the pen with the depth camera installed above the individual feeding station and indicated by the red arrow (credit: Jacinta D. Bus). was recalibrated. On Farm 2, no camera cleaning or recalibration was estimate, in kg). If a pig had <30 daily images, no individual BW esti­ performed. On Farm 1, the feeding station’s construction was developed mate was calculated (Farm 1: all individual BW estimates were avail­ for only one pig to eat, and consequently be registered by the RFID able; Farm 2: 1056 individual BW estimates (81%) were available). The system, at a time. On Farm 2, the RFID was configured to randomly pen estimated BW was calculated by taking the median of the weight switch between the three partitions and read once every second, hence estimates of all pictures of a particular pen. Image acquisition, seg­ each position could only be read once every 2 s. The camera took a 3D mentation and selection, weight estimation, and removal of outliers, if image at a 10-s interval on Farm 1 at individual and pen level, and on any, were conducted by dol-sensors using their confidential algorithm. Farm 2 at pen level. On Farm 2, at individual level, the camera took a 3D Due to data storage malfunctioning, individual images and BW estimates image every 1 s. from Farm 1 on the first weighing moment were not available and, A YOLO-based algorithm developed and trained by dol-sensors hence, the Farm 1 analyses at individual level included the second and detected whether a pig was present in the picture and, if so, third weighing moments only. segmented out the individual pig. Segmented images were retained only if the pig was in standing position, close to the RFID reader and in full 2.3. Statistical analyses view (determined with the YOLO-based algorithm). As part of dol-sen­ sors’ protocol, unclear images due to dirt on the camera lens were dis­ Statistical analyses were performed for each farm separately in R carded. If multiple pigs were detected in a frame, only the BW estimated v.4.1.1 [16]. The agreement between the estimated and scale-based BW for the pig closest to the RFID antenna (Farm 1: the pig with the head in was assessed at individual and pen level with repeated-measures the feeder; Farm 2: the pig with the head in the feeder partition) was concordance correlation coefficient (CCC) and repeated-measures maintained. If the closest pig to the antenna could not be determined, a Bland-Altman plot (library SimplyAgree v.0.0.2; [17]). Analyses at in­ dummy pig was recorded, and the respective sample was manually dividual level were controlled for pen and pig, whereas analyses at pen discarded. For selected segmented images, a regression neural network, level were controlled for pen. To assess the relationship between the developed and trained by dol-sensors, used available information in the number of daily images per pig on the BW estimation accuracy, we image (e.g., head-to-tail distance, ribs’ width, and ribs’ curve) to esti­ conducted a partial Pearson correlation (rpartial; library ppcor v.1.1; mate pig BW. Using the time stamps of the depth camera and RFID [18]) controlled for pig between the number of individual daily images system, each measurement could be prescribed to an individual pig. If and the BW estimation variation (scale-based BW minus estimated BW, only an RFID reading or only a BW estimate was obtained, or if two pigs in kg). Based on the criteria proposed by Hinkle et al. [19], CCC (range: were registered to be in the feeding station (Farm 1) or in one feeder 0–1) and rpartial (range: -1–1) values were interpreted as: negligible partition (Farm 2) simultaneously, the sample was discarded. The (0.0–0.3 and 0.0–-0.3), low (0.3–0.5 and -0.3–-0.5), moderate (0.5–0.7 weight estimates of all the images (Table 1) of an individual pig taken on and -0.5–-0.7), high (0.7–0.9 and -0.7–-0.9) or very high (0.9–1.0 and each day were used to calculate a median weight (i.e., the daily BW -0.9–-1.0). Based on the Bland-Altman plot reporting guidelines 3 G.A. Franchi et al. Smart Agricultural Technology 3 (2023) 100117 Fig. 2. (A) Illustration of the pen for fattening pigs on Farm 2. Feeders with three partitions are represented by the white rectangles, the iDOL65 3D camera by the white pentagon, drinking cups by the black hollow circles, and the two solid black squares represent two hard wooden sticks in separate vertical racks provided as general enrichment for all pens (retrieved from Larsen et al., 2018). (B) Illustration of the pen with the depth camera installed above the multi-partitioned feeder and indicated by the red arrow (credit: Guilherme A. Franchi). proposed by Flegal et al. [20], a 95% Limits of Agreement (95%LoA) of individual (95%LoA=-9.5–6.5 kg) and pen level (95%LoA=-8.5–5.5 kg) 10% of the median scale-based BW [21] over the fattening period of (Fig 3). The relationship between number of images and BW estimation each farm population (Farm 1: ±8.8 kg; Farm 2: ±5.5 kg) was pre­ accuracy was negligible (rpartial=-0.20; N=204 observations; P-val­ specified as an acceptable degree of agreement between the two mea­ ue=0.004). The RMSE were 1.7% (1.5 kg; R2=0.98) and 2.4% (1.8 kg; surements. The normal distribution of the number of individual daily R2=0.99) of the median weight over weighing moments at individual images and BW estimation variation were graphically confirmed with a and pen level, respectively. histogram and QQ-plot. The camera’s BW estimation error (Root Mean On Farm 2, the agreement between the individual estimated BW and Square Error; RMSE) at individual and pen level was assessed with the individual scale-based BW was also very high at individual (CCC: mixed-effects linear regression (library glmmTMB v.1.1.2; [22]). The 0.98; 95%CI: 0.978–0.982) and pen (CCC: 0.98; 95%CI: 0.974–0.984) individual-level model included estimated BW as outcome variable, the level. At individual level, the BW was overestimated by the 3D camera in scale-based BW, day and their 2-way interaction as fixed effects, number comparison with the scale with a mean difference of 0.2 kg (95%LoA=- of daily images per pig as covariate, and pig nested in pen as random 6.9–7.4 kg) (Fig 3). At pen level, the BW was underestimated by the 3D effect. The pen-level model included estimated BW as outcome variable, camera in comparison with the scale with a mean difference of 1.5 kg the scale-based BW, day and their 2-way interaction as fixed effects, and (95%LoA=-8.9–6.0 kg) (Fig 3). The relationship between the number of pen as random effect. We prespecified an RMSE ≤5% as acceptable [11]. images and BW estimation accuracy was negligible (rpartial=-0.08; Model assumptions of normality and heteroscedasticity were confirmed N=1056 observations; P-value=0.014). The RMSE were 3.6% (1.9 kg; through graphical inspection of the residuals. R2=0.98) and 3.1% (1.7 kg; R2=0.98) of the median weight over weighing moments at individual and pen level, respectively. 3. Results and discussion On both farms, based on the CCC and RMSE outcomes, the perfor­ mance of the depth camera and its underlying algorithm was satisfactory On Farm 1, the agreement between the individual estimated BW and at both individual and pen level, and better than the BW estimation the individual scale-based BW was very high at individual (CCC: 0.97; reported in earlier studies using other types of depth cameras. For 95%Confidence Interval (CI): 0.963–0.971) and pen (CCC: 0.99; 95%CI: instance, Kongsro [13] and Condotta et al. [4], using a Microsoft Kinect 0.989–0.995) level. The BW was overestimated by the 3D camera in camera, reported similar R2 (0.99) compared with our results. However, comparison with the scale with a mean difference of 1.5 kg at both they found higher RMSE ([13]: 3.4 kg, 4.8%; [4]: 3.0 kg, 4.9%) than our 4 G.A. Franchi et al. Smart Agricultural Technology 3 (2023) 100117 Fig. 3. Bland-Altman plots of paired differences between camera and scale BW values against the average of the pairs of methods at individual and pen level on Farms 1 and 2. In each plot, the solid black line indicates the mean difference, the black dashed lines indicate the mean difference’s 95% confidence interval, the green solid lines indicate the upper and lower limits of agreement (LoA), and the green dashed lines indicate the LoA’s 95% confidence intervals. The purple dots indicate individual pigs on the individual-level plots and individual pens on the pen-level plots. study. In Kongrso [13], 71 finisher pigs (37 Duroc, 34 Landrace) with a frequency [23]. Moreover, the camera performed well across two BW range of 29–139 kg were used, whereas Condotta et al. [4] used 234 different feeding site set-ups with minimum physical disturbance of the finisher pigs (78 Landrace, 78 Duroc, and 78 Yorkshire) sampled at four pen layout, except for minor adaptations such as installation of a rack different moments (BW range: 15–110 kg). Additionally, using a second above the feeder to mount the camera and removal of feeding station’ version of the Microsoft Kinetic camera and finisher pigs (unknown top fence to ensure pigs’ body were visible on Farm 1 (Fig. 1) and a metal breed) with an average BW of 120 kg, Fernandes et al. [14] reported an rod to hold the camera at the required position and height on Farm 2 R2 of 0.88 and an RMSE of 4.4 kg (3.6%), both values poorer than the (Fig. 2). ones reported herein. Overall, the iDOL65 camera revealed a satisfac­ Yet, the overall performance of the iDOL65 on Farm 1 was better tory BW estimation performance over farms and development stages, than on Farm 2. We propose two explanations for such a difference. overcoming breed differences that could potentially reflect in differ­ First, the more frequent lens cleaning and camera calibration may have ences in body conformity and, consequently, influence BW estimation benefited the iDOL65’s BW estimation performance on Farm 1. Had the [1]. Furthermore, the negligible correlation between the number of in­ cameras on Farm 2 been inspected during the experimental period, dividual daily images and individual BW estimation accuracy on both potential sources of error compromising high-quality image acquisition farms indicates the camera satisfactorily estimated pigs’ BW regardless might have been avoided and a greater BW estimation performance of the amount of individual daily images available. This is particularly might have been achieved. Second, on Farm 2, several animals were relevant for subordinate pigs who may show a reduced daily feeder visit removed from pens due to sickness from the seventh recording day. Such 5 G.A. Franchi et al. Smart Agricultural Technology 3 (2023) 100117 a disturbance on the social environment likely disturbed the hierarchy under grant agreement no. 862919. We thank dol-sensors a/s for making and the individual access to resources such as the feeding site. Conse­ the weight estimates from both farms available. Thanks to Topigs quently, image sampling and BW estimation, particularly at individual Norsvin for providing the location and animals in Germany and Stefanie level, may have been compromised [24]. Nuphaus (Topigs Norsvin, Germany) for assisting in this study. Thanks Except for the pen-level outcomes on Farm 1, the 95%LoA outcomes to Judith Zuijdwegt (Wageningen University & Research, the were outside the respective prespecified acceptable ranges. Moreover, Netherlands), and Dines Thøger Bolt and Birthe Houbak (Aarhus Uni­ our Bland-Altman plots revealed few points of disagreement (i.e., purple versity, Denmark) for assisting with pig weighing. points outside the LoA in Fig. 3) between the individual estimated BW and the individual scale-based BW on Farm 1 when pigs weighed >110 References kg and several points of disagreement on Farm 2 when pigs weighed between 60–90 kg. According to the manufacturer, the dataset used to [1] C.P. Schofield, J.A. Marchant, R.P. White, N. Brandl, M Wilson, Monitoring pig growth using a prototype imaging system, J. Agric. Eng. Res. 72 (1999) 205–210. train and test the iDOL65’s algorithm included mostly pigs weighing [2] Y. Wang, W. Yang, P. Winter, L. Walker, Walk-through weighing of pigs using between 15–110 kg, and few pigs with BW outside this range. Hence, machine vision and an artificial neural network, Biosyst. Eng. 100 (2008) 117–125. this can explain the BW estimation disagreement cases observed on [3] M. Nilsson, A.H. Herlin, H. Ardö, O. Guzhva, K. Åström, C. Bergsten, Development of automatic surveillance of animal behaviour and welfare using image analysis Farm 1. For the disagreement cases observed on Farm 2, we propose two and machine learned segmentation technique, Animal 9 (2015) 1859–1865. connected justifications. When pigs reach approximately 60 kg, they are [4] I.C. Condotta, T.M. Brown-Brandl, K.O. Silva-Miranda, J.P. Stinn, Evaluation of a typically subjected to a dietary change to meet their increasing nutrient depth sensor for mass estimation of growing and finishing pigs, Biosystems Eng. 173 (2018) 11–18. requirements [25], which, in combination with differences in pig size [5] R. Dohmen, C. Catal, Q. Liu, Computer vision-based weight estimation of livestock: and stocking density, can lead to increased competition for access to the a systematic literature review, N.Z. J. Agric. Res. 65 (2022) 227–247. feeder [23]. Consequently, as multiple pigs could have tried to access the [6] S. Bellini, The pig sector in the European Union. In Understanding and combatting African Swine Fever: A European perspective, Wageningen Academic Publishers, same feeder partition, the camera may have had issues to take images of 2021, pp. 639–648. the actual eating pig (closest to the RFID antenna) and detect the [7] N. Brandl, E Jørgensen, Determination of live weight of pigs from dimensions respective pig body boundary without the presence of another pig’s measured using image analysis, Comput. Electron. Agric. 15 (1996) 57–72. body part. This may also explain the absence of BW estimates of some [8] J.C. Petherick, A note on allometric relationships in Large White × Landrace pigs, Anim. Sci. 36 (1983) 497–500. pigs on Farm 2. Such a challenge was reported by Buayai et al. [26] [9] J. Wu, R. Tillett, N. McFarlane, X. Ju, J.P. Siebert, P. Schofield, Extracting the using a semi-automatic machine vision approach in commercial finisher three-dimensional shape of live pigs using stereo photogrammetry, Comput. pigs whose 2D images were also taken from the top view of the feeder. Electron. Agric. 44 (2004) 203–222. [10] M. Kashiha, C. Bahr, S. Ott, C.P. Moons, T.A. Niewold, F.O. Ödberg, D Berckmans, Due to the algorithm confidentiality, we were unable to examine in Automatic weight estimation of individual pigs using image analysis, Comput. detail the aspects resulting in reduced BW estimation accuracy at indi­ Electron. Agric. 107 (2014) 38–44. vidual level. [11] C.P. Schofield, Evaluation of image analysis as a means of estimating the weight of pigs, J. Agric. Eng. Res. 47 (1990) 287–296. In conclusion, the iDOL65 showed a satisfactory BW estimation [12] A.J. Heinrichs, G.W. Rogers, J.B. Cooper, Predicting body weight and wither height performance at both individual and, especially, pen level over different in Holstein heifers using body measurements, J. Dairy Sci. 75 (1992) 3576–3581. development stages, feeding site set-ups and breeds. The minimal [13] J. Kongsro, Estimation of pig weight using a Microsoft Kinect prototype imaging system, Comput. Electron. Agric. 109 (2014) 32–35. physical interventions needed to install the camera system above the [14] A.F. Fernandes, J.R. Dórea, R. Fitzgerald, W. Herring, G.J. Rosa, A novel automated feeding site and no need for habituation period for the animals or system to acquire biometric and morphological measurements and predict body handling by the barn staff represent an advancement in 3D camera weight of pigs via 3D computer vision, J. Anim. Sci. 97 (2019) 496–508. [15] A. Pezzuolo, M. Guarino, L. Sartori, L.A. González, F. Marinello, On-barn pig technology and a potential solution for BW estimation in commercial weight estimation based on body measurements by a Kinect v1 depth camera, farms. We encourage further camera performance optimisation, partic­ Comput. Electron. Agric. 148 (2018) 29–36. ularly for individual-level BW estimation, by inclusion of BW observa­ [16] R Core Team, R: A Language and Environment for Statistical Computing, R tions over 110 kg in the algorithm’s training and testing and Foundation for Statistical Computing, Vienna, Austria, 2021. https://www.R-pro ject.org/. experimentation of the camera under other pen and feeding site layouts. [17] A. Caldwell, SimplyAgree: Flexible And Robust Agreement and Reliability Additionally, to encourage the iDOL65 adoption by farmers, we suggest Analyses. R package version 0.0.2, 2021. https://CRAN.R-project.org/package=Si an integration of the camera system with equipment (e.g., robots, mplyAgree. Accessed on Feb 02, 2022. [18] K. Seongho, ppcor: Partial and Semi-Partial (part) Correlation. R Package Version drones) allowing for moving the camera across the pen, or even across 1.1, 2015. https://CRAN.R-project.org/package=ppcor. Accessed on Jun 14, 2022. the pig unit section, and taking images of animals at any time and [19] D.E. Hinkle, W. Wiersma, S.G. Jurs, Applied Statistics for the Behavioural Sciences, location. This improvement can potentially permit BW estimation at 3rd ed., Houghton Mifflin, Boston, MA, USA, 2003. [20] K.M. Flegal, B. Graubard, J. Ioannidis, Use and reporting of Bland–Altman analyses individual level independent of a combination with an RFID system in studies of self-reported versus measured weight and height, Int. J. Obes. 44 installed at the feeding site, which is not commercially available yet. (2020) 1311–1318. [21] P. Barrios, J. Martin-Biggers, V. Quick, C. Byrd-Bredbenner, Reliability and criterion validity of self-measured waist, hip, and neck circumferences, BMC Med. Declaration of Competing Interest Res. Method. 16 (2016) 1–12. [22] M.E. Brooks, K. Kristensen, K.J. van Benthem, A. Magnusson, C.W. Berg, The authors declare that they have no known competing financial A. Nielsen, H.J. Skaug, M. Maechler, B.M. Bolker, glmmTMB balances speed and flexibility among packages for zero-inflated generalized linear mixed modeling, R J interests or personal relationships that could have appeared to influence 9 (2017) 378–400. the work reported in this paper. [23] J.D. Bus, I.J.M.M. Boumans, L.E. Webb, E.A.M. Bokkers, The potential of feeding patterns to assess generic welfare in growing-finishing pigs, Applied Animal Data availability Behaviour Science 241 (2021), 105383. [24] T. Banhazi, M. Dunn, A. Banhazi, Are image analysis based weight prediction systems precise enough for on-farm applications? in: D. Berckmans, M. Oczak, Data will be made available on request. M. Iwersen, K. Wagener (Eds.), Proceedings of the 10th European Conference on Precision Livestock Farming Vienna, 2022, pp. 544–550. [25] I. Kyriazakis, G.C. Emmans, C.T. Whittemore, Diet selection in pigs: choices made by growing pigs given foods of different protein concentrations, Anim. Sci. 51 Acknowledgements (1990) 189–199. [26] P. Buayai, K. Piewthongngam, C.K. Leung, K.R. Saikaew, Semi-automatic pig weight estimation using digital image analysis, Appl. Eng. Agric. 35 (2019) This study was part of the EU project ClearFarm and funded by the 521–534. European Union’s Horizon 2020 research and innovation program 6