Introduction

Unmanned aerial vehicles (UAVs) equipped with sensors or tanks can be used in a range of agricultural activities such as pesticide application, fertilizer spraying, irrigation, and field monitoring (Michels et al., 2021). China has been using UAVs in agriculture since 2010 (Zheng et al., 2019). After 10 years of rapid development, 70,344 UAVs were being used in China for plant protection purposes and by 2020 they were treating14.48 million hectares of cropland (China Agricultural Machinery Industry Association, 2021). Pesticide application is one of the most common uses of UAVs in Chinese agriculture (Yang et al., 2018). UAVs equipped with tanks can fly over fields at a low altitude to ensure uniform rate of pesticide application and can treat 20–33 hectares per day, which is 30–60 times faster than manual spraying (Zheng et al., 2019). UAVs’ intelligent spraying system makes pesticide application more accurate and reduces pesticide residue (Chen et al., 2020). In addition, UAVs can overcome topography barriers and can be used in hilly or mountainous regions.

Although pesticide application with UAVs has many advantages compared with backpack sprayers, such as high efficiency and flexibility, low labor requirements, and reduced pesticide exposure, most Chinese farmers still apply pesticides manually using backpack sprayers (Wachenheim et al., 2021). Backpack sprayers can be operated in almost any situation with low operating costs. However, their inferior efficiency is one of the main sources of pesticide overuse, and it is estimated that China’s average pesticide use per hectare is more than three times above the world average (Wu et al., 2018). Nevertheless, UAV spraying also has disadvantages. For example, spray drift is more severe with UAVs than with aircraft spraying or ground based application due to the vortex airflow generated by the UAV rotors (Carvalho et al., 2020; Wang et al., 2021). This is one reason why UAV spraying is prohibited in many European countries as exposure to this drift is considered to be a hazard both for the environment and human health (Remáč, 2018). China’s regulations on UAV spraying are less stringent than in Europe and it is permitted as long as the operation is carried out in accordance with national operational specifications (The Civil Aviation Administration of China, 2019). Specifically, UAV pilots must be trained and licensed, and UAV spraying must adhere to national operational specifications and avoid harming people on the ground.

In 2017, China launched nationwide agricultural purchase subsidies in six provinces to promote the use of UAVs in agricultural production. Agricultural cooperatives and plant protection organizations are eligible to apply for these subsidies and can be granted a subsidy amounting to up to 30% of the purchase price for UAVs, whereby the maximum sum of the subsidy does not exceed 4,370 $ per UAV (Ministry of Agriculture & Rural Affairs of People’s Republic of China, 2017). The agricultural UAV purchase subsidies have had a great impact on the use of UAVs, and their numbers have increased significantly since 2017. Given the small size of farms in China, it is not economically viable for individual farmers to own UAVs. Thus, some agricultural UAV companies and plant protection organizations provide on-farm UAV spraying services which have become popular leading to a huge market demand (Lan et al., 2019). Prices of UAV spraying services (labor and machinery costs) range from $15 to $30 per hectare depending on crop types and topography (Chung, 2019; Wang et al., 2022). In addition to UAV purchase subsidies, some provinces also partially subsidize UAV spraying services based on the cumulative area of UAV operations. For example, in the Guangdong province subsidies for UAV pesticide spraying services range from $32 to $43 per hectare (Li et al., 2022).

Given the benefits of UAVs, some studies have investigated the adoption of this technology for pesticide application. Zheng et al. (2019) used a probit model involving 897 farmers in Jilin province of China to estimate the factors influencing the adoption of UAVs for plant protection. Their results suggest that perceived usefulness, perceived ease-of-use, UAV-related knowledge level, and agricultural income ratio have a positive influence on UAV adoption. Likewise, Wachenheim et al. (2021) used a probit model to estimate the effects of social networks, resource endowment, and perceptions on Chinese farmers’ intention to adopt UAVs for pesticide application. The results indicate that arable land area, agricultural income share, within-family village leadership, perceived usefulness, and credit availability have positive effects on UAV adoption. Chen et al. (2020) employed logit models to investigate the factors that influence Chinese farmers’ willingness to adopt UAVs for pesticide application. The results show that arable land area and cooperative membership are positively correlated with farmers’ adoption intention, and the land threshold for UAV adoption is estimated to be 2 hectares. Han et al. (2022) used a technology acceptance model and found that perceived usefulness, perceived ease-of-use, and external environment (e.g., government subsidies, extension services, and training) have positive effects on UAV adoption for pesticide application among Chinese farmers in Shaanxi province. To sum up, these studies reveal that farmers’ characteristics, farm household characteristics, and external environment are the main factors influencing UAV adoption.

The abovementioned studies focus on identifying factors that facilitate or constrain the adoption of UAVs for pesticide application. However, the economic effects of adoption still remain unclear. To address this research gap, propensity score matching (PSM) based on a dataset of over 2,000 Chinese grain farmers was used to identify the factors that influence their adoption of UAVs and to analyze the economic effects of adoption on outcome variables, including farmers’ revenue, pesticide costs, time spent on pesticide application, and pesticide application frequency. Furthermore, generalized propensity score matching was used to estimate the heterogeneity of outcome variables arising from differing UAV adoption intensities. Finally, the conclusions present some policy suggestions that are apt to promote the use of UAVs in China.

Materials and methods

Data source

The data used in this study is based on the “National Scientific Fertilizer Application Research Project 2019” headed by the Ministry of Agriculture and Rural Affairs of China. This national survey focused mainly on evaluating the farm-level impact of a scientific fertilizer application project. The survey was carried out in 2019 by the National Academy of Agriculture Green Development, China Agricultural University and was based on face-to-face interviews with farmers from 11 of the country’s main grain producing provinces: Heilongjiang, Jilin, Hebei, Henan, Shandong, Shaanxi, Gansu, Anhui, Jiangsu, Hunan, and Guangxi. This survey applied stratified multi-stage sampling and random sampling. Firstly, within each province, counties were classified according to the cultivated area, and 4 counties were randomly selected. Secondly, within the selected counties, townships were classified according to per capita income, and 3 townships were randomly selected. Thirdly, within the selected townships, villages were classified according to per capita income, and 2 villages were randomly selected. Finally, within the selected villages, farmers were classified according to their cultivated area and were randomly selected. The interview questions covered characteristics of farm households, aspects of farm management, agricultural production expenditure and revenues, pesticide application, and farmers’ knowledge about fertilizer application, etc.

This survey was assisted by the local government, and all the farmers selected participated in the survey, i.e., 100% response rate. The sample consisted of 3,061 farmers: 1,123 maize farmers, 817 rice farmers, and 1,121 wheat farmers. Given the research purpose and variables of this study, missing values and invalid observations were excluded, leaving a final sample consisting of 1,078 maize farmers, 763 rice farmers, and 1,045 wheat farmers.

Variable definitions and descriptive statistics

Table 1 presents definitions of the variables in this study and their descriptive statistics. The descriptive analysis shows significant differences between UAV adopters and non-adopters in many variables. UAV adopters are more likely to be male, young, fulltime farmers, and better educated than non-adopters and they also seem to have higher net incomes, bigger farm size, and fewer land parcels than non-adopters. Concerning the outcome variables, adopters show higher revenues and less time spent on pesticide application than non-adopters.

Table 1 Description of variables
Full size table

Empirical model

In this study, propensity score matching (PSM) was used to estimate the impacts of UAV adoption on outcome variables. PSM applied a set of observed covariates to construct a counterfactual comparison group to match against the treatment group based on the probability of UAV adoption (Khandker et al., 2009). The probability or propensity score, then served as the basis for matching UAV adopters with non-adopters using three different matching algorithms. The average treatment effects of UAV adoption are the mean difference of outcome variables between the treatment group and the comparison group.

Assuming farmers are risk neutral and rational, farmer i will only adopt UAVs for pesticide spraying if the expected utility of adoption (\({D}_{1}^{*}\)) is greater than non-adoption (\({D}_{0}^{*}\)): \({D}_{i}^{*}= {D}_{1}^{*}- {D}_{0}^{*}>0\), where \({D}_{i}^{*}\) is the latent variable which captures the utility difference between adoption and non-adoption. \({D}_{i}^{*}\) is unobserved, but it can be denoted as a function of observed covariates. Thus, a latent variable model is given as follows (El-Shater et al., 2016; Zheng et al., 2021):

$$D_{i}^{*} = \beta {\varvec{z}}_{{\varvec{i}}} + { }\varepsilon_{i} \;{\text{with}}\;{\text{Di = }}\left\{ {{ }\begin{array}{*{20}c} 1 & {{\text{if D}}_{{\text{i}}}^{*} { > 0}} \\ 0 & {{\text{otherwise}}} \\ \end{array} } \right.$$
(1)

where Di is a binary variable and equals 1 if farmer i adopts UAVs and 0 otherwise; zi is a vector of observed covariates that affect UAV adoption; β is a vector of parameters to be estimated; εi is the error term.

Firstly, a probit model was employed to estimate the probability of farm households adopting UAVs for pesticide spraying. Secondly, UAV adopters and non-adopters were matched based on the probability or propensity score derived from the probit model. The robustness of the results was checked using three different matching algorithms, including kernel matching, nearest-neighbor matching, and radius matching to compare UAV adopters with non-adopters (Caliendo & Kopeinig, 2008). Finally, the average treatment effects on the treated (ATT) for the outcome variables were estimated according to (Khandker et al., 2009):

$${\text{ATT}} = {\text{E }}({\text{Y }}\left( 1 \right) - {\text{Y }}\left( 0 \right)|D_{i} = 1) = {\text{E }}({\text{Y}}\left( 1 \right)|D_{i} = 1) - {\text{E }}({\text{Y}}\left( 0 \right)|D_{i} = 1)$$
(2)

where \(\mathrm{E }(\mathrm{Y }(1) | {D}_{i}=1)\) is the potential outcome of adopters in the treatment group and \(\mathrm{E }(\mathrm{Y }(0) | {D}_{i}=1)\) is the potential outcome of adopters had they decided not to adopt and become part of the counterfactual comparison group.

It is important to note that the effectiveness of PSM relies on two fundamental assumptions: conditional independence and common support (Khonje et al., 2015). Conditional independence assumes that, given a set of observable covariates (Xi) which are not affected by UAV adoption (Di), outcome variables are independent of the UAV adoption status. If Yi (1) is the outcome of UAV adopters and Yi (0) is the outcome of non-adopters, the conditional independence can be expressed as: \(({Y}_{i} (1), {Y}_{i} (0))\perp {D}_{i} | {X}_{i}\) (Caliendo & Kopeinig, 2008). Common support assumes that the number of UAV adopters is approximately equal to the number of non-adopters with whom they are matched. PSM attempts to estimate the difference between outcome variables of UAV adopters and non-adopters with similar characteristics, but bias cannot be avoided if unobservable covariates affect UAV adoption (Khandker et al., 2009). Although the UAV adoption rate is low, at 3.8% of the full sample, PSM can still generate unbiased estimates of treatment effects if the appropriate variables are selected into the model (Gitonga et al., 2013; Pirracchio et al., 2012).

Empirical results and discussion

Estimates of probit model

In the first stage of PSM, the probit model was used to analyze the determinants of UAV adoption and to calculate the propensity score of adoption for each farmer. The results of the probit analysis are reported in Table 2. The Wald χ2 test (254.29) indicates the joint significance of independent variables (p = 0.000). Farmers who lease land from others, and therefore hold a larger area of arable land, are more likely to adopt UAVs than their non-adopter counterparts. This finding is in line with Skevas et al. (2021) who pointed out that farmers who rent out land to others are less likely to be willing to use UAVs. As expected, family net income and farm size are positively correlated with the probability of adoption. Young farmers are more likely to adopt UAVs, suggesting that they are more open-minded and skilled with digital agricultural technology than older farmers. These results are consistent with Michels et al. (2020) who found that farm size has a positive effect and a farmer’s age has a negative effect on the UAV adoption process in German agriculture. Similarly, Skevas and Kalaitzandonakes (2020) observed that household income has a positive impact on American farmers’ intention to adopt UAVs and Skevas and Kalaitzandonakes (2020) and Chen et al. (2020) found that cooperative members are more likely to adopt UAVs than non-members. On the other hand, in this study, membership in an agricultural cooperative has a negative effect on UAV adoption, indicating that most of the UAV adopters did not participate in agricultural cooperatives and that agricultural cooperatives did not play a significant role in UAV extension. Soil fertility has a positive relationship with UAV adoption, implying that the probability of adoption is higher on a farm with fertile soil. Fulltime farmers are also more inclined to adopt UAVs. This could be due to the fact that farmers whose main source of income is earned in agriculture are more willing to try UAVs to enhance productivity.

Table 2 Probit regression model for UAV adoption
Full size table

Balancing tests

In the second stage of PSM, UAV adopters and non-adopters were matched on the basis of their propensity scores. The results of balancing tests before and after matching are shown in Table 3. Regardless of which matching algorithm is used, Pseudo R2, which reveals how well the independent variables explain UAV adoption (Caliendo & Kopeinig, 2008), fell from 0.282 before matching to a range of 0.019–0.025 after matching. The likelihood ratio test of the joint significance of covariates was not rejected before matching but it was rejected after matching. The mean standardized bias was below 8% after matching. The total percentage of bias reduction ranges from 4.6% to 13.7%. Thus, the PSM has significantly reduced the biases of covariates in the treatment group and the control group, suggesting a good matching quality.

Table 3 Balancing tests
Full size table

Common support

Figure 1 shows the density distributions of propensity scores for UAV adopters and non-adopters before and after matching. It reveals substantial overlaps in the density distributions of the propensity scores of UAV adopters and non-adopters after matching. Obviously, the common support assumption is satisfied after matching.

Fig. 1
figure 1

Density distributions of propensity scores for UAV adopters and non-adopters before and after matching

Full size image

The economic effects of UAVs in pesticide application

Table 4 reports the impacts of UAV adoption on revenue, pesticide costs, time spent on pesticide application, and pesticide application frequency. The results derived from kernel matching, nearest neighbor matching, and radius matching are very close, suggesting that the results are robust. The use of UAVs in pesticide application did have some positive economic effects. Firstly, UAV adoption significantly improved grain farmers’ revenue by approximately 434–488 dollars per hectare, possibly because the grain yield per unit area increased thanks to the effective pest and disease control it offers. Subtracting the cost of UAV spraying (labor and machinery), the net revenue of UAVs in pesticide application is 404–473 dollars per hectare. Secondly, the adoption of UAVs noticeably reduced the time spent on pesticide application in the range of 14.4–15.8 h per hectare, and thus it indirectly reduced labor costs for this task. This is mainly due to the fast, accurate, and efficient pesticide spraying performed by UAVs compared to traditional approaches. Finally, although UAV adoption reduced the costs and application frequency of pesticides, the impacts were not statistically significant. The fact that UAV spraying did not noticeably reduce pesticide costs could be due to the fact that a lot of Chinese farmers use UAVs for uniform rate pesticide application instead of site-specific spraying (Yang et al., 2018).

Table 4 The impacts of UAV adoption on revenue, pesticide costs, time spent on pesticide application, and pesticide application frequency
Full size table

However, it is quite likely that some important causal factors are missing (e.g., management ability and entrepreneurial capability) and these factors both boost profits and increase the likelihood of adoption. That is, while UAVs may have little to do with increased profits for the average farmer, profits and UAV use are nevertheless correlated. Further tests are needed to check if the possible unobserved covariates (e.g., farmer’s ability, risk preferences, and motivation) influence UAV adoption and, at the same time, outcome variables.

Sensitivity analysis

PSM assumes that UAV adoption and outcome variables are solely affected by the observable covariates. However, hidden bias may be a problem if the possible unobserved covariates (e.g., farmer’s ability, risk preferences, and motivation) influence UAV adoption and outcome variables simultaneously (Chagwiza et al., 2016; Gitonga et al., 2013; Mishra et al., 2016; Schreinemachers et al., 2016). Thus, the Rosenbaum bounds test was performed to check the robustness of results to hidden bias (Rosenbaum, 2002). The PSM results show that UAV adoption only has statistically significant effects on grain farmers’ revenue and the time spent on pesticide application. Thus, the Rosenbaum bounds test was performed to assess the sensitivity of these two outcome variables to unobserved variables. Since the impact of UAV adoption on grain farmers’ revenue is positive, focus should concentrate on the upper bound (sig +) in this case (Caliendo & Kopeinig, 2008). Likewise, the lower bound (sig −) should be the focus for the time spent on pesticide application. Gamma represents log odds of differential assignment arising from unobserved factors (Becker & Caliendo, 2007). When Gamma is up to 2, the sig + of revenue is still significant at 5% level and the sig − of time spent on pesticide application remains significant at 1% level, indicating that grain farmers’ revenues and time spent on pesticide application are insensitive to hidden bias (Table 5).

Table 5 Sensitivity analysis of the outcome variables to hidden bias, p-values
Full size table

Continuous treatment effects

PSM uses a binary treatment variable (UAV adoption) in the model and can only estimate the average treatment impact of UAV adoption on outcome variables. However, the heterogeneous treatment impact of UAV adoption is unclear (Shiferaw et al., 2014). The generalized propensity score (GPS) matching (Hirano & Imbens, 2004) serves as an extension to PSM. It uses a continuous treatment variable in the model and thus allows the heterogeneous treatment impact of UAV adoption on outcome variables to be explored. In this study, a continuous treatment variable, UAV adoption intensity (natural logarithm of area using UAVs for pesticide spraying) was used in the GPS matching to study the heterogeneous treatment impact of UAV adoption. Other covariates were the same as those previously used in the PSM.

Figure 2 shows that revenue and UAV adoption intensity have an inverted U-shaped relationship. The revenue increases from 2,500 dollars/hectare at adoption intensity of 1 to 2,750 dollars/hectare at adoption intensity of 3. After that, revenue drops from 2,750 dollars/hectare to 1,500 dollars/hectare as the adoption intensity increases. The marginal revenue decreases as the adoption intensity increases, declining from 250 dollars/hectare at adoption intensity of 2 to − 750 dollars/hectare at adoption intensity of 7. Figure 3 shows that the time spent on pesticide application decreases as the adoption intensity increases. The time spent on pesticide application decreases from 38 h/hectare at adoption intensity of 1 to 2 h/hectare at adoption intensity of 2. Once the adoption intensity of 2 has been reached, the time spent on pesticide application does not show any visible fluctuations, remaining at 2 h/hectare. The marginal time spent on pesticide application increases gradually and reaches a diminishing return at adoption intensity of 3.

Fig. 2
figure 2

Dose response (average treatment effects) function and marginal treatment effects function for revenue. Standard errors and confidence intervals were estimated by 100 bootstrap replications. Adoption intensity: natural logarithm of area with use of UAVs for pesticide spraying

Full size image
Fig. 3
figure 3

Dose response (average treatment effects) function and marginal treatment effects function for time spent on pesticide application. Standard errors and confidence intervals were estimated by 100 bootstrap replications. Adoption intensity: natural logarithm of area with use of UAVs for pesticide spraying

Full size image

These empirical results reveal that UAV adoption intensity has heterogeneous impacts on grain farmers’ revenue and the time spent on pesticide application. This may be due to the fact that increased farm size leads to resource misallocation and management inefficiency, and finally to a decline in the impacts of UAV adoption (Sheng et al., 2019). Likewise, Abdul Mumin and Abdulai (2022), Mohammed and Abdulai (2022), Shahzad and Abdulai (2021), and Wu (2022) also reported heterogeneous returns for adoption of agricultural technologies due to differences in resource endowments, such as farm size, financial resources, and social networks. The results of this study suggest that, in terms of marginal revenue and marginal time spent on pesticide application, the optimal UAV adoption intensity for Chinese grain farmers is estimated to be 3, referring to 20 hectares of arable land.

Conclusions

This article uses the PSM method to identify the factors that influence Chinese grain farmers’ adoption of UAVs and to analyze the impacts of UAV adoption on farmers’ revenue, pesticide costs, time spent on pesticide application, and pesticide application frequency. The UAV adoption rate among the grain farmers in this survey was relatively low at only 3.8%. The empirical results show that UAV adoption is significantly and positively correlated with arable land area, family annual net income, soil fertility, rented land, fulltime farmer status, and young farmers. Policy makers aiming to increase UAV adoption should appreciate that older farmers, small-scale farmers, part-time farmers, and low-income farmers face more barriers in UAV adoption. In the future, more UAV extension services and education programs should target at these groups. However, membership in an agricultural cooperative has a significant negative impact on UAV adoption, indicating that most of the UAV adopters in this study did not participate in agricultural cooperatives. Traditional agricultural cooperatives did not play a significant role in UAV extension. Thus, benefit-risk sharing UAV agricultural cooperatives should be established to promote UAV adoption. Farmers who are interested in using UAVs for pesticide application can set up an agricultural cooperative and buy UAVs together, whereby members jointly cover the purchase and maintenance costs and share the use of their UAVs. This reduces the cost and risk for each famer and makes UAVs more affordable for the majority.

The use of UAVs can increase revenue and reduce the time spent on pesticide application, and these results are insensitive to hidden bias arising from unobserved variables. On average, the adoption of UAVs increases revenue by approximately 434–488 dollars per hectare and reduces time spent on pesticide application by about 14.4–15.8 h per hectare. UAV adoption should therefore be encouraged. The GPS matching indicates that UAV adoption has heterogeneous effects on revenue and time spent on pesticide application. In terms of marginal revenue and marginal time spent on pesticide application, the optimal adoption intensity of UAV in Chinese grain farming is estimated to be 20 hectares, suggesting that small and medium-scale farmers are the main beneficiaries of UAV adoption. Large farms are advised to improve resource allocation and management efficiency to increase the returns of UAV adoption.