Effects of Smart Farming on the Productivity of Korean Dairy Farms: A Case Study of Robotic Milking Systems

Abstract

The Korean agricultural sector faces increasing challenges such as an aging population, labor shortages, and the liberalization of agricultural markets. To overcome these challenges, the Korean government has striven to enhance the competitiveness of agriculture by introducing AI-based technologies to the agricultural sector, labeling this as smart farming. This study estimates farm-level benefits of adopting smart farming technologies, robotic milking systems, in Korean dairy farms. The benefits are estimated by comparing the productivity (i.e., the savings of labor input, increased calf production, and increased milk production) of adopting and non-adopting farms. Our study uses the propensity score matching method to address potential problems from confounding factors, sample selection bias, and the small number of adopters. Our results show that farms that adopted robotic milking systems produced 0.10 to 0.11 more calves per year than farms that did not adopt the system. The adopters also increased milk production by 2.44 kg to 2.88 kg per head/day, while reducing labor input by 0.15 to 0.30 per head/week. However, the reduced labor input was not statistically significant. When the analysis was extended to regard the farm characteristics, the labor input became significant from small and family-run farms. We also found that the increase in the number of calves produced per head was statically significant from small farms, family-run farms, and farms with successors. The increased milk production per head was statistically significant from large farms, farms employing hired workers, and farms with successors. Our findings suggest that the Korean government continue promoting smart farming technologies such as the robotic milking system to increase the adoption rate. The findings can also provide useful information about target markets of this technology, which can be used to increase the adoption rate and ultimately enhance the sustainability and competitiveness of the Korean dairy industry.

1. Introduction

The Korean agricultural sector faces increasing challenges such as an aging population, labor shortages, and the liberalization of agricultural markets [1]. To overcome these challenges, the Korean government has striven to enhance the competitiveness of agriculture by introducing advanced technologies such as big data, artificial intelligence (AI), and the internet of things (IoT) to the agricultural sector, labeling this as smart farming. The new data-driven and automated technologies are also called Agriculture 4.0. The new technologies aim to increase economic, environmental, and social sustainability by implementing precision farming, while increasing productivity and reducing inputs and production costs. The adoption of the new farming technologies is considered to be crucial for sustainable agriculture, which ensures food security, environmental conservation, and economic prosperity for future generations [2,3].

However, farmers face various challenges in adopting smart farming or Agriculture 4.0 technologies. These challenges include high initial facility investment costs, a lack of infrastructure required for new technologies, a lack of standards due to the diversity of technologies and products, a lack of information on new technologies and skilled labor, issues surrounding the ownership and sharing of the generated data, compatibility issues with traditional agriculture, and resistance to the adoption of new technologies [4]. The Korean government has made various efforts to overcome these challenges associated with smart farms and to expand the adoption of smart farming in agriculture. Since 2004, the Korean government has promoted smart farming through various research and development (R&D) policies and aims to establish up to 9000 smart livestock farms by 2025 [5].

Several studies examined the effect of smart farming in Korea, particularly focusing on increasing farm productivity, improving the quality of farm products, and reducing labor cost [6,7,8], while some studies examined factors influencing the adoption and diffusion of these technologies [9,10,11,12]. Most recently, a few studies have evaluated smart dairy farming and the robotic milking systems in the Korean dairy industry [3,7,8,13]. Yang et al. [3] investigated the benefits of smart farming technologies from two adopting farms and showed that adopting smart farm equipment significantly increases milk production and values of milk component analysis, while decreasing “days open (DO)” [3]. The DO is the period between when a dairy cow calves and conceives again, which represents the reproductive performance of dairy cows. Therefore, the reproductive performance is high with a low DO. Heo and Seo [13] examined the change in productivity before and after ICT adoption in four types of livestock farms (cattle, dairy cow, hog, and poultry). Overall, the study found that productivity increases in most farms after the adoption of the ICT. Recent studies by the Korean Rural Development Administration [7] and the Korean Institute for Animal Products Quality Evaluation [8] compare productivity and farm income changes before and after the adoption of robotic milking systems on dairy farms. Their results show that the adoption of robotic milking systems improved both farm productivity and income [7,8].

Other studies have also evaluated the benefits of adopting robotic (in other words, automatic) milking systems outside of Korea [14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33]. Some studies have shown that the adoption of robotic milking systems has enhanced dairy productivity by reducing labor requirements [14,15,16,17]. Additionally, studies have indicated that allowing cows to be milked at any time has led to increased milk production [18,19,20,21,22,23,24,25]. However, despite these benefits—such as reduced labor input, increased productivity, and higher milk yield—many studies also suggest that the high initial investment costs have ultimately led to a decrease in profitability for dairy farmers [15,26,27,28,29,30,31,32,33].

For example, Hansen [14] reports advantages and disadvantages of adopting automatic milking systems using interview data collected from 14 dairy farms in Southern Norway. The survey data show that advantages include saved milking time, more engaging in other farming activities, a more stable management of cows, and reduced labor, while the main disadvantage is information overload. Shortall et al. [15] compare the economic viability of automatic milking systems and conventional milking systems over a 10-year period for two farm sizes: one with 70 cows and the other with 140 cows. Results of a stochastic budgetary simulation model find that the automatic milking systems reduce labor input by 36%. However, the results also point out that automatic systems are not always more profitable than conventional systems when initial investment is sufficiently large. The study indicates that although the high investment cost of the new technology does not always guarantee the improvement of profitability, the reduction in labor can still make the technology attractive to some farmers due to their lifestyle choices. Jacobs and Siegford [16] suggest that automatic milking systems increase milk production by up to 12%, reduce labor by 18%, and improve dairy cow welfare by allowing cows to choose their own milking times. However, the study also notes that some farms may not experience labor reduction if some cows do not milk voluntarily or if the automatic milking system is not fully integrated into the farm management system. Wagner-Storch and Palmer [17] compared the feeding and milking behavior, as well as the milk yield, of cows housed in the same barn and fed the same diet, divided into a conventional milking parlor group and a robotic milking group in Wisconsin, USA. Feeding behavior patterns were more varied in the parlor group compared to the robotic milking group, but milk yield was significantly higher in the robotic milking group. Duplessis et al. [20] surveyed 97 dairy farms in Canada in 2014 and 2015 to investigate changes in average milk yield and somatic cell count (SCC) following the adoption of automatic milking systems. On average, after transitioning to an automatic milking system, the herd size increased by 11.3 cows, the milk yield increased by 441 kg per cow per year, the culling rate increased by 1.3%, and the calving interval decreased by 7 days.

The findings of earlier studies provide useful insights about benefits of smart dairy farming technologies including robot milking systems, and facilitate directions for future research. However, most earlier findings are not based on statistical inference as they rely on simple average comparisons of productivity before and after the adoption of the robotic milking systems. When comparing the productivity between adopting and non-adopting farms, it should be necessary to match individual farms in each group based on similar farm characteristics. Without matching the groups by farm characteristics, the analysis is likely to face confounding factors and sample selection bias problems, and therefore, estimates of the causal relationship between the adoption of smart farming and productivity measures may be biased [34]. A proper statistical analysis is required to address potential confounding and sample selection problems.

Unlike many earlier studies, our study conducts a statistical analysis using the propensity score matching (PSM) method to evaluate effects of smart farming technologies in Korean dairy farms, while controlling for potential biases from confounding and sample selection problems. An advantage of applying the PSM method for this type of analysis is that it helps equalize the covariate distribution between the treatment and control groups, thereby enabling proper comparison and addressing the potential endogeneity problem that may arise from confounding and selection bias problems [34]. The confounding and sample selection problem should be particularly problematic when the sample size is small with a low prevalence of treatment. Our dataset includes a small number of adopters because smart farming is still in the early stage of its diffusion curve in the Korean dairy industry. Earlier studies indicate that the PSM method addresses confounding and selection bias problems and provides more accurate results than other methods even when the sample size is small with a low prevalence of adopters [35,36,37]. For example, Pirracchio, Resche-Rigon, and Chevret [36] and Cenzer, Boscardin, and Berger [37] conduct a series of Monte Carlo simulations with various sample sizes to compare the performance of the PSM method when the number of treatment subjects is small. The studies find that the PSM method provides better results than other methods even in cases of small samples or low prevalences of treatment. Moreover, since the PSM analysis is conducted on a refined dataset created by matching the data using propensity scores, the interpretation of results should be more intuitive and clearer than outcomes from other comparison methods [38]. Another contribution of our study is the extension of benefit analysis by farm characteristics. The benefits of smart dairy farming are further examined by breaking down the data based on different farm characteristics such as farm size, family-run farms vs. farms with hired workers, and farms run by government-certified successors or not. Results of this analysis should provide in-depth information for developing marketing strategies and government policies regarding robotic milking systems.

The objective of this study is to assess the benefits of adopting smart farming, robotic milking systems, for Korean dairy farmers by measuring productivity improvement, specifically focusing on labor input, calf production, and milk production. The benefits of the robotic milking systems are evaluated by estimating differences in farm performance between adopting and non-adopting Korean dairy farms. As cows can milk themselves in the robotic milking systems automatically at any time, the new technology is expected to reduce labor input while increasing productivity.

2. Materials and Methods

2.1. Data

The data for our analysis was sourced from a facility and equipment survey for the Korean livestock industry conducted by the Korea Rural Economic Institute (KREI) in 2021 [39]. The full survey included 1017 beef, pork, and dairy farms, and the number of dairy farms included in the survey was 177. In 2021, there were a total of 6105 dairy farms in Korea, raising 400,798 dairy cows. The average number of cows per farm was 68, and each farm produced an average of 913 kg of milk per day. The average milk production per cow was 27.4 kg [40]. In the same year, the total milk production was 2035 million tons at the industry level. The regional distribution of dairy farms and dairy cows is shown in Figure 1.

The surveyed livestock farms were stratified by breeding type, region, and farm size within the population, and then randomly sampled. The survey was conducted in two phases: preliminary and main surveys. The preliminary survey was carried out through on-site visits by a research team and a professional survey agency from 30 August to 6 September 2021. Then, the main survey was conducted from 1 October to 30 December 2021. In the first phase of the main survey, an online survey was carried out by a professional research agency. Then, the follow-up survey was conducted by visiting non-participating farms in the first phase.

2.2. Methodology

Since individual dairy farms experience only one condition either with or without the adoption of smart farming technologies, it is not possible to directly compare farm productivity before and after adopting the technologies using within-group tests. In this case, between-group tests can be used, but it is typically expected that the tests would face problems with variances in factors between groups, which should cause the confounding factor (as a result, the selection bias) problem. Therefore, measuring the effects of smart farming technology on farms’ productivity needs to account for the confounding effects from observable and unobservable factors (i.e., needs to balance the adopter group with the non-adopter group to make an “apples to apples” comparison).

Many studies have used the PSM procedure for statistical analysis to reduce confounding effects when subjects are compared with the between-group tests to assess, for example, effects of new government policies or adopting new technologies and products, e.g., [34,38,41,42,43]. Our study uses the PSM procedure to adjust for confounding effects, particularly for the situation where adopters are systematically different from non-adopters, which should improve the comparability between the two groups. As suggested by earlier studies [35,36,37], the PSM method should provide more accurate results than other methods even when the sample size is relatively small with a low adoption rate.

The propensity score represents the probability that a research subject will be included in the treatment group (i.e., the adopters) rather than the control group based on specific covariates [44], which is mostly estimated using logit or probit models. By dividing the sample into treatment and control groups and calculating propensity scores within each group, similar members can be matched based on these scores, allowing for the analysis of the average treatment effect for the entire population or specific groups. The PSM method is therefore a statistical process that creates a balanced distribution of all the confounders included in the estimation of the propensity scores [45].

The propensity score can be expressed as:

$$PS\left(X_i\right)=P\left(T_i=1|X_i\right),$$

(1)

where $PS\left(X_i\right)$ is the propensity score ranging from 0 to 1, $P$ is the probability, $T_i$ is the treatment status (1 for smart farming adoption, 0 for non-adoption), and $X_i$ is the covariate vector.

To match farms between these two groups based on propensity scores, two key assumptions are necessary: the Conditional Independence Assumption (CIA) and Common Support Assumption (CSA) [38]. The CIA indicates that the potential outcomes $\left(Y_i\left(0\right)\mathrm{or}{Y}_i\left(1\right)\right)$ and the treatment $\left(T_i\right)$ are independent given the propensity score ($PS(X_i)$), which can be expressed as:

$$\left[Y_i\left(0\right),Y_i\left(1\right)\right]\perp T_i|PS\left(X_i\right).$$

(2)

If the CIA holds, as treatment status is probabilistically determined by covariates, the covariates influence both the treatment status and the treatment effect simultaneously, while allowing for the estimation of an unbiased treatment effect [38,45]. The CSA requires that there should be comparable farms in the treatment and control groups, and the distribution of covariates should overlap. To accurately assess the treatment effects using PSM, observations in both groups should be widely distributed. This assumption is expressed as:

$$0<P\left(T_i=1|X_i\right)<1.$$

(3)

Through PSM, the Average Treatment Effect on that Treated (ATT) can be estimated for the actual treated farms as [38]:

$$\mathrm{ATT}=E\left(Y_i\left(1\right)-Y_i\left(0\right)|T_i=1\right)=E\left(Y_i\left(1\right)|X_i,T_i=1\right)-E(Y_i\left(0\right)|X_i,T_i=1).$$

(4)

Equation (4) can be used to measure the productivity improvement due to the adoption of smart farming using outcome variables representing the productivity of dairy farms. Outcome variables used for our analysis include the weekly labor hours per head, the number of calves produced per cow, and the average daily milk production per cow.

2.3. Analytical Procedure

A logit model is used to estimate propensity scores of individual dairy farms as:

$$T_i=X_i\beta +{\epsilon }_i,$$

(5)

where $T_i$ represents the adoption of smart farming (1 for adoption, 0 for non-adoption), $X_i$ is the vector of explanatory variables representing farm characteristics of individual farms, $\beta$ is the vector of estimated coefficients, and ${\epsilon }_i$ represents the stochastic error term.

Using the estimated probabilities from the logit model, we can match treatment (smart farming adopters) and control (non-adopters) groups. There are various methods for matching, but this study uses the commonly employed methods of nearest-neighbor matching and kernel matching [34].

First, the nearest-neighbor matching matches farms of the treatment group with farms of the control group that have the most similar propensity scores, which is expressed as:

$$\mathrm{C}(P_i)=\frac{min}{j}\left|P_i-P_j\right|, j\in I_0,$$

(6)

where $\mathrm{C}\left(P_i\right)$ is the matching distance, $P_i$ and $P_j$ are the propensity scores of each member of the treatment and control groups, respectively, and $I_0$ is the sample set of the overlap.

Second, the kernel matching is a non-parametric method that assigns weights to all members of the control group to derive potential outcomes of group members. To determine the weights $\omega \left(i,j\right)$, a non-parametric kernel function is used. The non-parametric kernel function is defined as:

$$\omega \left(i,j\right)={\displaystyle \frac{K\left(\frac{P_j-P_i}{b}\right)}{{{\displaystyle \sum }}_{k\in c}K\left(\frac{P_k-P_i}{b}\right)}}$$

(7)

where $K\left(·\right)$ is the kernel function, b is the number of samples in the bandwidth, $P_i$ is the propensity score of treatment farm i, and $P_j$ and $P_k$ are propensity scores of the jth and kth control groups in each bandwidth.

3. Results

3.1. Descriptive Statistics

Descriptive statistics of the data obtained from the survey are presented in

Table 1. Descriptive statistics (N = 177).

Variable		Freq.	Mean	S.D.	Min	Max
Region	Gyonggi/Gangwon	70	0.395	0.490	0	1
	Chungcheong	44	0.249	0.433	0	1
	Jeolla/Jeju	40	0.226	0.419	0	1
	Gyeongsang	23	0.130	0.337	0	1
Age	Year	-	59.203	11.542	28	83
Successor of family farm	Yes = 1, No = 0	59	0.333	0.473	0	1
Milk collectors	Coops	93	0.525	0.501	0	1
	Private companies	44	0.249	0.433	0	1
	Committee	40	0.226	0.419	0	1
Full time	Yes = 1, No = 0	154	0.870	0.337	0	1
Experience		-	20.905	6.150	5	25
Recording	Computer = 1	54	0.305	0.462	0	1
No. of heifers	Head	-	48.203	43.615	2	350
No. of multiparous cow	Head	-	51.480	36.855	0	250
No. of cows total	Head	-	99.684	63.511	8	430
Hired workers	Yes = 1, No = 0	45	0.254	0.437	0	1
Farm size	m2	-	3154	2111	200	11,286
Debt	KRW 1,000,000 *	-	422	718	0	4700
Disease prevention and treatment cost	KRW 10,000/head *	-	10.9	10.9	0	53.0
Manure disposal cost	KRW 10,000/head *	-	8.1	10.8	0	20.5
Financial assistance	Yes = 1, No = 0	73	0.412	0.494	0	1
ICT program	Yes = 1, No = 0	16	0.090	0.288	0	1
Eco-friendly farm	Yes = 1, No = 0	27	0.153	0.361	0	1
Clean farm	Yes = 1, No = 0	39	0.220	0.416	0	1
Animal welfare farm	Yes = 1, No = 0	16	0.090	0.288	0	1
Neighbor complaint	Yes = 1, No = 0	38	0.215	0.412	0	1
Labor hour	Hour/week/head	-	1.547	0.806	0.233	7
Calf production	Head/cow	-	0.833	0.138	0.556	1.481
Milk production	Kg/day/head	-	32.081	3.214	25.000	45
Smart farm	Yes = 1, No = 0	13	0.073	0.262	0	1

* USD 1 = KRW 1331 on 27 August 2024.

. Our data include four regions: Gyeonggi/Gangwon, Chungcheong, Jeolla/Jeju, and Gyeongsang. Out of the four regions, the Gyeonggi/Gangwon region has the largest number of farms, accounting for 39.5%, while the Gyeongsang region has 13% of the farms in our dataset. The average age of farm owners is 59.2 years, and roughly a third of farm owners are government-certified successors of family farms. The government-certified successor program has been implemented to attract young farmers as fewer young people choose farming as a profession. Central and local governments provide various forms of financial and education assistance to the family farm successors. There are three major milk collectors: coops, private milk processing companies, and the Korea Dairy Committee. The Committee is a government-funded public company, which comprises dairy farmers, milk processors, government officials, consumers, and academics, and decides the producer price of milk each year. Among the three raw milk collecting bodies, most farmers in our sample sell their milk to dairy coops (52.5%), while 24.9% and 22.6% of farmers send their milk to private processors and the Korea Dairy Committee, respectively.

Most farm owners (87.0%) are full-time farmers with an average of 20.9 years of dairy farming experience, and fifty-four farms (30.5%) record management data on computers. The average number of dairy cows per farm is roughly 100, with 48 heifers and 52 multiparous cows. In total, 25.4% of farms have hired workers, while 74.6% of farms are family-run businesses. The average farm size is 3154 square meters with an average debt of KRW 420 million (USD 315,552). Annual costs for disease prevention and treatment, and manure disposal per head are KRW 108,000 (USD 81) and KRW 81,000 (USD 61), respectively. Over 40% of farms receive financial assistance from the government. About 9% of farms participate in ICT programs, while 15.3%, 22.0%, and 9% of farms have government certifications for eco-friendly livestock farming, clean farming, and animal welfare-friendly farming, respectively. Over 20% of farms had complaints about foul orders and/or spillage of animal waste from neighbors.

Dairy productivity indicators include the weekly labor input per head, the number of calves produced per multiparous cow, and the average milk production per head. The average weekly labor input per head is 1.55 h, the number of calves produced per multiparous cow is 0.833, and the average milk production per head is 32.08 kg per day. Smart farming automates and enhances various forms of equipment for dairy farming, which includes milking machines, environment controllers, feeding machines, and farm management programs. However, the level of automating equipment varies by individual farm, and there is no clear definition of which equipment a farm should have to be considered a smart farm. Therefore, in this study, the criterion for the adoption of smart farming is set based on the installation of the robotic milking system, which is the most-costly among the various automated equipment options, and we believe it is the most effective method of cost-reduction (particularly via input reduction and output expansion) in typical Korean dairy farms. The robotic milking system, also known as the automatic milking system, is a voluntary milking technology because it allows cows to choose when to be milked by a robotic arm without requiring any human labor to be present. As a result, it is expected that farmers would have more flexibility in their time and reduce labor input. It is also expected that the new technology would increase milk and calf productions because the robotic system allows cows to be milked on their own schedule as frequently as possible under a less stressful environment than that of a traditional milking parlor system. In our sample, about 7% of farms (13 farms) installed the robotic milking system. A recent press release from the Korean Rural Development Administration disclosed that out of 6123 dairy farms, approximately 180 farms had installed robotic milking systems as of February 2023, which indicates that only 3% of dairy farms have robotic milking systems [46]. It looks like our sample over-represents adopters, although we may still want to have more adopting farms for the purpose of statistical analysis.

3.2. Probability of Adopting Smart Farming

This study conducted a PSM analysis using STATA 15.1. A logit regression analysis with the stepwise method was employed to select the most relevant explanatory variables. Initially, a model including all available explanatory variables was used, and then variables were sequentially removed to find the model with the highest log-likelihood and Pseudo R². Logit regression results are reported in

Table 2. Estimates of logit regression for smart farming adoption.

Variable	Coefficient	Std. Error	z-Value	P > |z|
Constant	−0.952	1.881	−0.510	0.613
Age	−0.103 ***	0.036	−2.840	0.005
No. of cows total	0.030 ***	0.008	3.720	0.001
Debt	−0.001	0.001	−1.530	0.126
Manure disposal cost	0.088 ***	0.026	3.380	0.001
Financial assistance	−2.906 **	1.221	−2.380	0.017
ICT program	2.368 *	1.349	1.760	0.079
Eco-friendly farm	−1.800	1.187	−1.520	0.129
Neighbor complaint	2.148 **	0.960	2.240	0.025
Log Likelihood	−25.020
Pseudo R2	0.461

*, **, and *** represent statistical significance at 10%, 5%, and 1% levels, respectively.

. The results show that coefficients of Age, the No. of cows total, the Manure disposal cost, Financial assistance, ICT program, and Neighbor complaint are statistically significant at least at the 10% level. Negative coefficients of Age and Financial assistance indicate that younger farmers and farms not receiving financial assistance from the government are more likely to adopt these smart farming technologies. Positive coefficients of the No. of cows total, Manure disposal cost, ICT program, and Neighbor complaint suggest that farms with more cows, high manure disposal costs, and neighbor complaints about odors and animal waste have a higher probability of adopting the technologies.

3.3. Matching Results

The propensity scores are calculated from Equations (6) and (7) and matched between adopter and non-adopter groups, and results are shown in

Table 3. Range of common support between adoption and non-adoption farms.

	N	Nearest-Neighbor (N-N) Matching		Kernel Matching
		Off Support	On Support	Off Support	On Support
Adoption	13	4	9	6	7
Non-adoption	164	-	164	-	164
Total	177	4	173	6	171

. For the smart farming adoption group, the number of samples within the common support (matching) region was nine from nearest-neighbor matching and seven from kernel matching. For the non-adoption group, there were no samples outside the common support region. The control group used to estimate the average treatment effect had 164 samples. Thus, a total of 174 sample data were used for nearest-neighbor matching and 171 for kernel matching. Table 3 shows that overall, although the smart farming adoption group within the common support region was small due to the small number of adopters, the adoption and non-adoption groups were well matched, ensuring the robust estimation of the average treatment effect.

3.4. Average Treatment Effect on Labor Input

To assess the effectiveness of smart farming adoption in dairy farms, we measured changes in labor input, calf production, and milk production using Equation (4). The estimated ATT on labor input is reported in

Table 4. Average treatment effect on labor input by farm characteristics.

Variable	Matching Method	Mean		Difference	t-Value
		Adopters	Non-Adopters
Labor input	N-N	0.966	1.115	−0.149	−0.760
	Kernel	0.994	1.296	−0.302	−1.350
Large farms (≥150)	N-N	0.927	0.834	0.093	0.560
	Kernel	1.180	0.767	0.413	-
Small farms (<150)	N-N	0.996	1.615	−0.619 **	−2.600
	Kernel	1.098	1.558	−0.460 **	−2.010
Hired worker (Yes)	N-N	1.045	1.085	−0.040	−0.150
	Kernel	1.290	0.932	0.357	1.520
Hired worker (No)	N-N	0.895	1.555	−0.660 ***	−2.880
	Kernel	0.997	1.639	−0.643 ***	−2.770
Successor (Yes)	N-N	1.090	1.565	−0.475	−1.340
	Kernel	1.090	1.346	−0.256	−1.060
Successor (No)	N-N	0.930	1.230	−0.298	−1.340
	Kernel	0.995	1.225	−0.230	−0.700

**, and *** represent statistical significance at 5%, and 1% levels, respectively. N-N: nearest-neighbor matching method.

. For the labor input per head/week, smart farming adopters averaged between 0.966 and 0.994 h, while non-adopters averaged between 1.115 and 1.296 h per head/week. The ATT is −0.149 h per head/week for nearest-neighbor matching and −0.302 for kernel matching. Despite the reduction in labor from adopters, differences between adopters and non-adopters are not statistically significant, indicating no strong evidence that smart farming reduces labor input. Then, the treatment effect, i.e., the impact of smart farming on labor input per head, was further examined by farm characteristics. In this case, small farms (farms with less than 150 cows) show a substantial reduction in labor input, ranging from 0.460 to 0.619 h per head, which is statistically significant at the 5% level. Family-run farms that do not employ hired workers also reduce labor input significantly, ranging from 0.643 to 0.660 h per head, by adopting smart farming, which is statistically significant at the 1% level.

3.5. Average Treatment Effect on Calf Production

The estimated ATT effect of smart farming on calf production is shown in

Table 5. Average treatment effect on calf production by farm characteristic.

Variable	Matching Method	Mean		Difference	t-Value
		Adopters	Non-Adopters
Calf production	N-N	0.860	0.755	0.105 ***	3.120
	Kernel	0.870	0.771	0.100 **	2.170
Large farms (≥150)	N-N	0.803	0.771	0.033	0.700
	Kernel	0.750	0.702	0.048	-
Small farms (<150)	N-N	0.894	0.790	0.104 ***	2.570
	Kernel	0.890	0.818	0.072	1.370
Hired worker (Yes)	N-N	0.810	0.775	0.035	0.850
	Kernel	0.790	0.676	0.092	1.250
Hired worker (No)	N-N	0.910	0.790	0.120 ***	2.560
	Kernel	0.910	0.836	0.074	1.230
Successor (Yes)	N-N	0.785	0.815	−0.030	−0.490
	Kernel	0.785	0.788	−0.003	−0.060
Successor (No)	N-N	0.881	0.798	0.084 **	2.220
	Kernel	0.912	0.793	0.120 **	2.100

**, and *** represent statistical significance at 5%, and 1% levels, respectively. N-N: nearest-neighbor matching method.

. Average numbers of calves produced per head for multiparous cows for adopters and non-adopters are 0.860 and 0.755 from the nearest-neighbor matching, and 0.871 and 0.771 from the kernel matching. Then, the corresponding ATT effects are 0.105 and 0.100 and both differences are statistically significant at least at the 5% level.

Table 5 also reports the ATT of smart farming adoption on calf production by farm characteristics. The ATT effect on calf production is also statistically significant from small farms and family farms. It is also noted that the ATT effect is statistically significant in no-successor farms. Small farms show an increase of 0.104 calves per head with nearest-neighbor matching. Additionally, family-run farms have an increase of 0.120 calves per head, and farms without successors show an increase of 0.084 calves per head with the nearest-neighbor matching and 0.120 calves per head with the kernel matching.

3.6. Average Treatment Effect on Milk Production

Table 6. Average treatment effect on milk production by farm characteristic.

Variable	Matching Method	Mean		Difference	t-Value
		Adopters	Non-Adopters
Milk production	N-N	34.222	31.778	2.444 **	2.060
	Kernel	33.857	30.982	2.875 **	2.060
Large farms (≥150)	N-N	34.667	32.533	2.133 ***	2.370
	Kernel	35.000	30.915	4.085 ***	2.440
Small farms (<150)	N-N	33.600	30.560	3.040	1.630
	Kernel	33.500	31.413	2.087	0.930
Hired worker (Yes)	N-N	35.000	30.900	4.100 ***	4.590
	Kernel	35.500	30.815	4.685 ***	2.950
Hired worker (No)	N-N	33.000	31.000	2.000	0.890
	Kernel	32.667	31.984	0.682	0.240
Successor (Yes)	N-N	35.000	30.500	4.500 ***	5.010
	Kernel	35.000	32.077	2.923 ***	2.720
Successor (No)	N-N	34.000	31.571	2.429 *	1.680
	Kernel	33.250	32.132	1.118	0.490

*, **, and *** represent statistical significance at 10%, 5%, and 1% levels, respectively. N-N: nearest-neighbor matching method.

reports the estimated ATT effect on milk production. Smart farming adopters produced more milk than non-adopters by 2.444 kg (nearest-neighbor matching) and 2.875 kg (kernel matching) of milk per head per day. The differences are statistically significant at the 5% level. When the ATT effect on milk production was examined using farm characteristics, we found that large farms (farms with more than or equal to 150 cows) produced 2.133 kg (nearest-neighbor matching) and 4.085 kg/day (kernel matching) more milk per head due to smart farming adoption, whereas no statistically significant effect was observed in small farms. Farms employing hired workers also show an increase in milk production by 4.100 kg and 4.685 kg/day per head, while increased milk production from family-run farms is not statistically significant. Additionally, farms with successors increased milk production by 2.923 kg (nearest-neighbor matching) and 4.500 kg/day per head (kernel matching). Farms without a successor also increased milk production as much as 2.429 kg (nearest-neighbor matching) and 1.118 kg (kernel matching) after adopting smart farming technologies but show a significantly smaller increase than the successor farms.

Our results showed that Korean dairy farms that installed the robotic milking systems improved their farm productivity in terms of labor input, calf production, and milk production, but the improvements in calf production and milk production were only statistically significant. However, when the analysis was extended by farm characteristics, the labor input became significant from small and family-run farms.

4. Discussion

The result of ATT on labor input is somewhat unexpected because the robotic or automatic system is expected to reduce labor hours. However, farmers were not able to fully eliminate their manpower from their milking parlors because they did not feel comfortable doing so. This might have happened because the new system was still in the early stage of adoption in Korea without being fully experienced and trained. In fact, one farm survey conducted with 19 Norway farmers pointed out that farmers never felt off-duty and therefore had to be not too distant from their milking parlor equipped with the robot milking systems [18]. Another reason might be that, as Jacobs and Siegford [16] suggested, some farms were not able to reduce labor because some cows did not milk voluntarily, or the robotic milking system was not fully effective as these farms were still learning about how to operate the new system adequately. These issues could be worked out in the future as more farmers adopt this technology and understand the benefits of the technology more clearly by sharing their experiences each other.

The result of ATT on labor input also indicates that the smart farming technologies are highly more effective in small farms than in large farms. This could be the case because small farms can be relatively more flexible than large farms in integrating new technologies into their existing routines. The labor input reduction in no-hired-workers farms could be attributed to the fact that family-run farms could more easily adjust and reduce labor input when adopting new technologies. In other words, the flexibility in labor allocation in family-run farms might result in greater efficiency gains when they adopt smart farming practices. The results of ATT on labor input show that the presence of a farm successor is not an important factor in determining the effect of smart farming on labor input. The results suggest that whether a farm has a successor or not does not influence the labor efficiency gains achieved through smart farming adoption.

The significant differences in ATT on calf production suggest that adopting smart farming, specifically robotic milking systems, has led to labor input savings, allowing increased effort to be invested in the management of individual dairy cows, which could have resulted in improved calf productivity in the Korean dairy industry. The calf production results by the farms’ characteristics suggest that the effect of smart farming on calf production is more pronounced in small farms, family-run farms, and farms without successors.

The results of increases in milk production are not surprising because when dairy farmers use robotic milking systems, they can have more frequent milking sessions than before without increasing workers’ time. Additionally, the increase in milk production due to smart farming adoption was greater from large farms, non-family farms employing hired workers, and farms with successors than from small farms, family farms, and farms without successors, respectively. These results indicate that the increase in milk production from adopting robot milking systems is more pronounced on larger farms than on smaller ones.

Based on these results, the profitability of adopting a robotic milking system was analyzed, and the findings are presented in

Table 7. Benefit–cost analysis of robotic milking system adoption (case of farm with 110 heads of cattle).

Benefit–Cost	Total Amount	Annual Amount
Robotic milking system cost
Installation (USD/farm)	262,960	26,296
Operation (USD/farm)	112,697	11,270
Total cost (USD/farm)	375,657	37,566 (C)
Robotic milking systembenefit
Labor input saving (hour/head)		7.75–15.7
Labor input saving (hour/farm)		853–1727
Labor cost savings (USD/farm)		13,250–26,826 (A)
Milk production increase (%/head)		7.69–9.28
Milk production increase (USD/farm)		38,045–45,911 (B)
Total benefit (USD/farm)		51,295–72,737 (A + B)
Benefit–cost ratio ((A + B)/C)		1.37–1.94

All cost and benefit amounts are calculated based on a farm with a herd size of 110 cows. USD 1 = KRW 1331 on 27 August 2024. A: annual benefits from labor cost savings. B: benefits from annual milk production increase. C: annual total cost.

. In Korea, the installation cost of a single robotic milking system is approximately USD 262,960, and its operating expenses over a useful life of 10 years are reported to total USD 112,697 [47]. Thus, the annual cost of the robotic milking system can be estimated at around USD 37,566 (depreciation plus operation costs). One robotic milking system can generally serve 50–60 cows, meaning a single unit can operate effectively in a farm with a total herd size of 100–120 cows (including milking cows and heifers). The average production cost per cow on Korean dairy farms in 2023 was USD 7129, including a labor cost per cow of USD 1039 per year and an annual labor input of 66.87 h per cow, resulting in an hourly wage of USD 15.5 [48].

Our results indicate that the introduction of a robotic milking system can on save labor hours per cow by 0.149–0.302 per week, which translates to 7.75–15.7 h annually. For a herd size of 110 cows, this amounts to a labor savings of 853–1727 h, which, at the hourly wage rate of USD 15.5, equates to an annual savings of USD 13,250–USD 26,826. This labor cost savings represents 35.3–71.4% of the annual depreciation cost of the robotic milking system (USD 37,566). In addition to labor savings, dairy farms benefit from an increase in milk production following the introduction of a robotic milking system. The analysis shows that milk production per cow increases by 7.69–9.28%. Assuming each cow produces 9977 kg of milk annually and that the farmer sells it at USD 0.90 per kg [49], a farm with 110 cows could generate an additional annual income of USD 38,045–USD 45,911.

In summary, for a farm with 110 cows, the annual cost of introducing a robotic milking system is USD 37,566. The benefits from adopting the robotic milking system include labor savings of USD 13,250–USD 26,826 and an increase in milk production valued at USD 38,045–USD 45,911. The total benefit thus ranges from USD 51,295 to USD 72,737. When comparing the cost of implementing the robotic milking system with benefits from labor savings and increased milk production, the benefit–cost ratio ranges between 1.37 and 1.94, indicating that the adoption of a robotic milking system is profitable. Our results are similar to findings from Heo and Seo [13], who found the benefit–cost ratio to be 1.2–1.8. Other studies, for example those of Zanin et al. [50] and Salfer et al. [51], also report that farms achieve profitability from the adoption of robotic milking systems.

Table 8. Summary of data, methodology, and key findings of previous studies.

Study	Data	Methodology	Findings
Yang et al. [3]	2 dairy farms adopting ICT in Korea	Comparison of farm productivity before and after adopting ICT	Daily milk production increased by 5.13% and 1.33% from farms A and B, respectively. Days open (DO) was reduced by 17.5% for farm A and 13.3% for farm B.
KRDA [7]	3 dairy farms adopting automatic milking systems (AMS) * in Korea	Comparison of farm productivity before and after adopting AMS	Daily milk production increased by 12.5%, the number of milking times per day increased by 33.3%, and milking labor hours per day decreased by 46.7%.
KIAPQE [8]	6 dairy farms adopting ICT in Korea	Comparison of farm productivity before and after adopting ICT	Daily milk production increased by 12.5%, hired labor decreased by 9.09%, while self-labor hours decreased by 22.41%, and profit increased by 16.33%.
Heo and Seo [13]	25 dairy farms adopting ICT in Korea	Comparison of farm productivity before and after adopting ICT	Daily milk production increased by 0.8%, culling rate decreased by 3.8%, DO decreased by 11.7%.
Hansen [14]	19 dairy farmers in Norway	Interview	To succeed with AMS, farmers must be motivated, behave proactively and adapt the new technology to their specific needs. Reduced time on milking, more stable treatment of cows and less need for relief are advantages, but disadvantages include farmers’ experience of constant on-call and information overload.
Shortall et al. [15]	Collected data from published studies	Simulations to compare AMS and conventional milking system (CMS)	AMS showed 36% reduction in labor demand, medium-specification CMS achieved greater profitability than AMS, irrespective of farm size, and AMS system did not achieve the highest profitability but achieved intermediate profitability at medium farm size.
Jacobs and Siegford [16]	Collected data from published studies	Literature review on benefits of AMS	AMS increased milk production by 12%, while decreasing labor input by 18%; it is also noted that AMS improved animal welfare by allowing cows to choose when to be milked.
Wagner-Storch and Palmer [17]	Behavioral data video-taped hourly for one day per month for 9 months in a dairy barn in Wisconsin, USA	Comparison of feeding and milking behaviors of cows: CMS vs. AMS	Lower concentration of cows in each pen and more consistently eating cows were observed from AMS than from CMS, and milk production was slightly higher in AMS than CMS (26.4 vs. 25.8 ± 0.2 kg/day).
Wade et al. [18]	2,071,662 test-day milkings from 306 herds in the Netherlands	Estimation of milk production regression model with an AMS dummy variable	Milk production increased by 10 to 12% after introducing AMS but increased by 2% after correcting for the year effect.
Sitkowska et al. [19]	2 dairy farms adopting AMS in Poland	Comparison of farm productivity before and after adopting AMS	The share of primiparous cows decreased by 6% and both lactation period and milk production increased after adopting AMS.
Duplessis et al. [20]	A survey of 97 Canadian dairy herds adopting AMS	Comparison of farm productivity before and after adopting AMS	Herd size, milk yield, and culling rate increased by 11.3 cows, 441 kg/cow per year, and 1.3%, respectively, and calving interval decreased by 7 days.
Bijl et al. [21]	62 dairy farms in the Netherlands	Comparison of farm inputs used and revenue between AMS and CMS	Labor input decreased by 29% due to AMS, but CMS farms had larger revenues than AMS farms.
Rotz et al. [27]	Collected variables and parameters from previous studies	Farm-simulation models	The potential benefit of USD 100/cow per year was estimated if AMS increased production an additional 5%, up to USD 100/cow of benefits was expected from AMS with a 20% reduction in initial equipment cost or doubling milking labor cost, and a reduction in USD 110/cow of annual net return was estimated if the economic life of AMS was reduced by 3 years.

*AMS is another way of describing robotic milking systems.

summarizes the data, methodology, and key findings of previous studies. Overall, our findings are consistent with those from most previous studies [3,7,8,13,14,15,16,17,18,19,20,21,22,23,24,25,27]. Most earlier studies found that robotic milking systems reduced labor input but increased milk production as well as calf production. Our study also found that the new technology reduced labor input, although our results from the full sample show no statistical significance. The reduction in labor input was statistically significant when a similar analysis was conducted with small farms. Increased milk production and calf production were statistically significant in our study. As noted earlier, one major contribution of our study is that it statistically tests findings from earlier studies. Another strength of our study might be that it provides a benefit analysis using farm characteristics, which could be helpful for target marketing and policy development.

5. Conclusions

This study estimates farm-level benefits of adopting smart farming technologies, specifically robotic milking systems, in Korean dairy farms. The benefits are estimated by comparing the productivity (i.e., savings of labor input, increased calf production, and increased milk production) of adopting and non-adopting farms. To analyze the effects of smart farming adoption on dairy farms, we use the PSM method to control for potential biases from confounding and sample selection problems. Propensity scores are calculated using predicted probabilities of a logit model. Results of the logit model indicate that the probability of adopting smart farming is higher for young farm owners, large farm owners, farms receiving financial assistance from the government, farms with high manure treatment costs, farms participating in ICT support projects, and farms experiencing complaints (about odor and/or waste) from neighbors.

Results of our PSM analysis found that while the overall effect of smart farming on labor input reduction was not statistically significant, it resulted in statistically significant increases in both the number of calves produced and the daily milk production per head. When our analysis was extended by using farm characteristics, the effect of labor input reduction became statistically significant from small and family farms. We also found that the increase in the number of calves produced per head was statically significant from small farms, family farms, and farms with successors. Similarly, the increased milk production per head was statistically significant from large farms, farms employing hired workers, and farms with successors.

Our results found that the robotic milking systems, introduced as a smart farming technology to Korean dairy farmers, produced statistically significant improvements in calf production and milk production. Our findings suggest that the robotic milking systems successfully improve calf production and milk production, and therefore, that the Korean government continue promoting smart farming technologies such as the robotic milking system to increase the adoption rate. These findings can also provide useful information about target markets of this technology, which can be used to increase the adoption rate and ultimately enhance the sustainability and competitiveness of the Korean dairy industry.

Our findings provide useful information for future research in evaluating benefits of new smart farming technologies such as robotic milking systems. Unlike other studies in the literature, our study uses the PSM procedure to adjust for confounding effects, particularly for the situation where adopters are systematically different from non-adopters. The PSM balances the adopter group and the non-adopter group to make an “apples to apples” comparison. As suggested by earlier studies, the PSM method should provide correct estimations even when the sample size is relatively small with a low adoption rate. Our application of the PSM method to technology evaluation and the findings from our analysis certainly should be of interest to policy makers, dairy farmers, and particularly researchers who would like to evaluate benefits of smart farming technologies.

The market size for the robotic milking system is expected to grow as the demand for dairy products expands, and the need for automation in milking technology increases. The global robotic milking machine market was valued at USD 2.86 billion in 2023 but is estimated to expand to USD 7.2 billion in ten years (SNS Insider). The SNS Insider report attributes this rapid market expansion to labor shortages and the efficiency increase in milk production due to the implementation of the new technology [52].

One limitation of our study is that the current study could not provide an analysis of the profitability of adopting the new technology with price and cost effects because our dataset does not include price and cost data, including the installation cost of the robotic milking systems. A direction for extending this study may be to estimate the profitability of adopting this new technology by incorporating price and cost effects. This analysis should require an extensive farm survey about input and output prices as well as capital cost, including the investment cost of the new milking systems. As stated earlier, the adoption of robotic milking systems is still in the early stage in the Korean dairy industry, and therefore, there are not many dairy farms currently using the new systems. As a result, our sample includes only 7% of dairy farms as adopters. One of the reasons we could not find a statistically significant improvement in labor hour reduction from the full sample might be due to the small sample size and/or the small number of adopters. A future study could reexamine this issue with a larger sample size, including more adopters.