The frequency and severity of crop damage by wildlife in rural Beijing, China
Item Type Article
Authors Fang, Liang;Hong, Yiping;Zhou, Zaiying;Chen, Wenhui
Citation Fang, L., Hong, Y., Zhou, Z., & Chen, W. (2021). The frequency and severity of crop damage by wildlife in rural Beijing, China.
Forest Policy and Economics, 124, 102379. doi:10.1016/
j.forpol.2020.102379 Eprint version Post-print
DOI 10.1016/j.forpol.2020.102379
Publisher Elsevier BV
Journal Forest Policy and Economics
Rights NOTICE: this is the author’s version of a work that was accepted for publication in Forest Policy and Economics. Changes resulting from the publishing process, such as peer review, editing,
corrections, structural formatting, and other quality control mechanisms may not be reflected in this document. Changes may have been made to this work since it was submitted for publication. A definitive version was subsequently published in Forest Policy and Economics, [124, , (2020-12-31)] DOI:
10.1016/j.forpol.2020.102379 . © 2020. This manuscript version is made available under the CC-BY-NC-ND 4.0 license http://
creativecommons.org/licenses/by-nc-nd/4.0/
Download date 2024-01-14 23:30:19
Link to Item http://hdl.handle.net/10754/666886
The Frequency and Severity of Crop Damage by Wildlife in Rural Beijing, China
Liang Fang1, Yiping Hong2, Zaiying Zhou3, Wenhui Chen1,∗
1. School of Economics and Management, Beijing Forestry University, Beijing 100083, China 2. CEMSE Division, King Abdullah University of Science and Technology, Thuwal 23955-6900, Saudi Arabia
3. Center for Statistical Science, Tsinghua University, Beijing 100084, China
[email protected], [email protected], [email protected], [email protected]
*Corresponding author
Abstract
With the continuous expansion of the number of wild animals, the human-wildlife conflict becomes more prominent.
There are frequent incidents of crop damage caused by wildlife. In this paper, we investigate the crop losses caused by wildlife in rural districts of the Municipality of Beijing, using a unique dataset of 31,573 observations from 2009 to 2017. We fit Poisson and negative binomial generalized regression models to explore the possible factors influencing the crop damage. Through statistical tests on the individual coefficients and the overall fitness, we find that the negative binomial generalized regression model with the population’s covariate more accurately describes the pattern of crop loss events. The frequency of crop loss events is positively related to the village’s distance from the river system but negatively associated with the distance from woodland, the population density, and protective measures. Based on these results, we propose solutions for the effective reduction of future crop losses and the prediction of lost crops’
compensation. The results can be used to develop sound policy recommendations to reduce crop loss events.
Keywords: Wildlife, Crop damage events/costs, Poisson generalized regression model, Negative binomial generalized regression model, Rural Beijing.
1. Introduction
Induced by the successful implementation of ecological conservation and restoration programs, the populations of wildlife and the scopes of their activities have seen gradual but significant expansion across China (Huang et al., 2018). For instance, in Beijing’s rural areas, wildlife habitats and areas of human activity have gradually become more intertwined, leading to increased conflicts between humans and wildlife and significant crop damages (Shen and Cui, 2015; Zhou et al., 2010). Thus, how to mitigate these conflicts and damages has emerged as a major policy issue.
Indeed, cases of crop loss caused by wildlife also happen in many parts of the world, and scholars have investigated the relevant issues. For example, Dickman et al. (2011) observed that some endangered species brought substantial economic costs to local areas, usually in low rural regions; as a result, the economic toll was almost unbearable for families, which hindered people’s efforts to escape poverty. (Dickman et al., 2011). Pettigrew et al. (2012) reviewed the current conflict management situation between human and carnivores in China, drew lessons from the experience of other countries, and put forward suggestions for improving animal protection (Pettigrew et al., 2012). Nyhus (2016) synthesized recent advances in exploring how to mitigate the human-wildlife conflicts to achieve positive conservation outcomes. (Nyhus, 2016). Furthermore, some scholars warn that if negative economic consequences to rural livelihoods are ignored, people living near the habitats tended to resist the idea of wildlife conservation, resulting in even greater loss to both human life and crop yields (Chen et al., 2016).
To mitigate the disturbance and economic loss to local communities caused by wildlife, policymakers usually
adopt one of the following two strategies: (1) organizing villagers to carry out patrolling activities (Goodrich, 2010;
Marchal and Hill, 2009; Treves and Karanth, 2003); or (2) compensating for damages to personal properties, including farm crops (Chen et al., 2016; Li et al., 2013; Mishra et al., 2003). But these strategies require a clear understanding of the wildlife distributions and their patterns of variation over time and space. This paper thus seeks to inform the policy design and implementation by improving our knowledge about the wildlife distribution and patterns of variation in rural Beijing. Of course, undertaking more effective conservation and protection in China is well in line with the government’s call for achieving a harmonious coexistence between humans and nature (Huang et al., 2018).
These strategies require a clear understanding of the wildlife distributions and their variation patterns over time and space to undertake more effective efforts toward conservation and protection. Indeed, doing so is well in line with the Chinese governments call for achieving a harmonious coexistence between humans and nature (Huang et al., 2018). This paper seeks to inform the policy process by improving knowledge about wildlife distribution and patterns of variation in rural parts of Beijing Municipality, lacking in the current scientific literature.
To our knowledge, the early research in this topical area is related to population changes in evaluating wildlife losses. For example, Sekhar (1998) did did a semi-structured survey of villagers near the Sariska Tiger Reserve (STR) in India in order to describe wild animal types causing crop losses and the degree of harm, the protective measures in response, and villagers attitudes toward wildlife protection (Sekhar, 1998). Balmford et al. (2001) analyzed the rela- tionship between the population density and the species richness in sub-Saharan Africa using correlation coefficients and fitted linear models (Balmford et al., 2001). Osborne et al. (2001) estimated the impact of human activities on the bustard habitat in central Spain using autocorrelation logistic regression models (Osborne et al., 2001).
Scholars also examined the factors causing the damages by wildlife (Herrero et al., 2006; Honda, 2007; Naughton- Treves, 1998; Potter, 1977; Rao et al., 2002; Sokal and Thomson, 1987; Wagner et al., 1997). By fitting logistic models, for instance, Sitati et al. (2005) considered the probabilities of elephants attacking farms in estimating potential losses in the Transmara District, Kenya (Sitati et al., 2005). Jing et al. (2008) used logistic regression models to identify the factors influencing crop losses caused by wild boars (Jing et al., 2008). Chen et al. (2013) and Yang et al. (2009) studied the multi-level compensations for rubber plantation losses caused by elephants in Xishuangbanna of southwest China and proposed policy recommendations for mitigating human-elephant conflicts (Chen et al., 2013; Yang et al., 2009). Similarly, Chen et al. (2016) estimated the frequencies of elephant and human conflicts in Xishuangbanna by fitting a negative binomial model and measuring its performance based on the Akaike information criterion (AIC) and Morans I tests (Chen et al., 2016). Karanth et al. (2013), working in India’s Western Ghats protected areas, employed logistic regressions to quantify crop losses caused by wildlife and the corresponding compensation schemes. Huang et al. (2018) applied a Poisson model to data on crop losses from the Daxueshan Nature Reserve in China to discover the correlation between the extent of losses with possible spatial factors.
In general, the previous studies feature remote research areas and specific wildlife species. In comparison, our study covers all the wildlife species found in the rural areas of Beijing, and we analyze the association between the number of wildlife loss events and their likely determinants using a large panel dataset covering 2009-2017. Moreover, we also correlate the damage amount of each village with the number of wildlife damage events in response to the need for compensating the economic losses. To determine the factors influencing crop damages and their economic costs, our team has collaborated with the Beijing Municipal Bureau of Landscaping and Forestry over almost a decade, tracking and compiling instances and costs of crop losses caused by wildlife. The study area consists of seven districts (or counties) with high concentrations of crop damage in Beijing: Yanqing, Mentougou, Changping, Miyun, Huairou, Pinggu, and Fangshan. Included in the potential factors influencing the frequency of damage are the type of wild animal(s), type and amount of crop losses, location (latitude and longitude) of the occurrence, preventive and protective measures, distances from villages to forested land, distances from water systems, and human population density.
After overviewing the descriptive statistics and visual representation of the damage events, we will first charac- terize the spatial distribution of crop damages and the historical trends in losses over time in section 2. In section
3, we will consider different versions of the generalized linear model to explain the correlation between the number of events and the adopted protective measures and check our fitted models suitability through residual plotting. In addition, we will investigate the relationship between the loss amount and the number of wildlife loss events at the village level. Based on our findings from these empirical steps, we will finally propose policy suggestions to mitigate and compensate crop losses caused by wildlife in section 4.
2. Exploratory data analysis
Here, we first describe the research area and the dataset we have compiled, followed by an illustration of frequen- cies of crop damage events, which will pave the way for the subsequent modeling efforts in section 3.
2.1. Research area and data
Located in the northern edge of the North Plains, Beijing is the capital of China and organized as a municipality.
The municipality has 16 districts within the geographical coordinates E115◦420−117◦240 and N39◦240−41◦360, with its center at E116◦200and N39◦560. Beijings terrain features a gradually declining elevation from northwest to southeast, with mountains and hills surrounding the municipality in the west, north, northeast, and plain angling down toward the Bohai Sea in the southeast. Beijings principal rivers include the Yongding, Chaobai, Beiyunhe, and Juma river. Out of the mountains and hills, these rivers flow through the plains and before emptying into the Bohai Sea.
Beijing has a sub-humid warm temperate continental monsoon climate, with hot and rainy summers, cold and dry winters, and short springs and autumns. Included in its staple crops are corn, peanuts, potatoes, and tubers, which are all sought after by wild animals. National ecological restoration and wildlife protection initiatives have improved the conditions for wildlife in Beijings rural areas (Lu, 2009). As a result, wildlife habitat and human activity are increasingly intertwined, raising both conflicts between humans and wildlife and crop damages caused by wildlife (Beijing Municipal Bureau of Landscape and Forestry, 2018). One of the major thrusts for this study is to investigate the spatial and temporal dynamics of these events.
From the perspective of land-use, woodland is mainly concentrated in the northwest, southwest, and southeast of Beijing; cultivated land is primarily located in the central and southeast parts, with some in the northwest corner;
and built-up land for residency, industry, and commerce is mostly concentrated in the central and southeast regions.
Our study area is mostly located in or around the forested area, which provides favorable habitat for wildlife. But, as noted, wildlife can cause damages to farm crops. Because Yanqing and Miyun have a higher portion of farmland than in other districts/counties, they have suffered disproportionately severe crop damages by wildlife. In contrast, the central and southeastern parts of Beijing are areas with intense human activities and high levels of urbanization; as such, crop losses caused by wild animals are rare. Clearly, it is crucial to understand the dynamics of wildlife activities and the resultant crop damages in order to design effective response strategies in wildlife management and prevent and mitigate crop losses. But developing a robust model that can identify the determinants of the frequency of crop damages caused by wildlife is only viable with the availability of a large, longitudinal dataset collected by the research team. Our dataset contains 31,574 records of crop damages by wildlife (boar, badger, ocelot cat, raptor, monkey, hare, etc.) from 2009 to 2017. In addition to the geographical (longitude and latitude) and administrative (district/county, township, and village) location of each crop-loss event, these records contain information on the category and amount of crop loss, and the status of financial compensation for each event. They contain the protective measures adopted to avoid any loss as well. Additionally, we collected the demographic and topographic data as possible explanatory variables for the crop-loss events. They include the distance between each events location and the nearby woodland or water system, and the human populations density at that location. The distance and population-density statistics were obtained from the Resource and Environment Data Cloud Platform1 operated by the Chinese Academy of Sciences (Xu, 2015). Landsat 8 imagery at a 30-meter resolution was selected manually for the study area and imported into
ArcGIS for an integrated analysis. Because of the relatively short duration of our study, the overall makeup of land use in Beijing did not change substantially.
2.2. Visualization of the crop loss events
Figure 1 illustrates the frequencies of crop damages by wild animals. First, we display the event frequencies in each district of Beijing by a distribution map (panel a). Darker colors denote higher frequencies of crop losses caused by wildlife. With 13,376 events, Yanqing had the highest frequency of crop damages. Miyun (7,666 events) had the second-highest frequency, followed by Mentougou (5,378 events), Huairou (3,216 events), and Fangshan (339 events). Then, we present the distribution of the event locations using a heat map (panel b) to identify the key villages of concern in each district. It can be seen that crop losses are more likely to occur in the eastern and southern parts of Yanqing, the northwestern part of Miyun, the southeastern part of Huairou, the southwestern part of Mentougou, and the northeast of Fangshan. These preliminary findings are also reflected in the density plot (panel c).
Finally, a histogram of the numbers of events (logarithm transformed) in villages (panel d) shows that the number of villages with more crop-loss events is much smaller than that with fewer events. It is this observation, coupled with the shape of the histogram, that has shed light for us to adopt a Poisson generalized linear model or a negative binomial generalized linear model as a potentially appropriate statistical means to represent the data. Additional graphic information derived from the observations is summarized in Appendix A.
3. Modeling the damage events
Building on the exploratory data analysis in the last section, we set out to seek for a well-fitted generalized linear model to represent the crop-loss events. We took the frequency of crop-loss events caused by wild animals in each village as the response variable. Included in the explanatory variables were the distance from the woodland, the distance from the river system, the village population density, and the ratio of protective measures as captured by the percentage of the number of events with documented protective measures over the gross number of events in the village, including fencing, signing, etc.
We further modified the response variable as the frequency minus one. This is because the response variable for a Poisson or a negative binomial model should range from zero to infinity, whereas the sample frequency of events is never smaller than one because the dataset contains only the villages where they happened. As detailed below, rather than the Poisson model adopted by Huang et al. (2018), we have found that a negative binomial model would be more appropriate for our dataset.
3.1. A Poisson generalized linear model
We began our modeling by fitting a Poisson generalized linear model, an often-used and handy tool for predicting a count outcome variable (λ) from a series of continuous or categorical determinants. Denoting the count of crop loss events in villageibyYi(withi=1, . . . ,nandn=635 being the number of villages considered), this model is defined as follows:
Yi∼Poisson(λi),
log(λi)=β0+β1Woodi+β2Riveri+β3Protecti+β4Popi,
whereWoodi,Riveri,Protecti, andPopiare covariates of villagei, representing, respectively, the distance from the nearest woodland in meters, the distance from the nearest woodland in meters, the ratio of protective measures (the number of events with documented protective measures over the whole number of events), and the population density (inhabitants per square kilometer);β0is the intercept term,β1,· · ·, β4are the coefficients to be estimated.
Distribution map by district: Darker color means higher frequency.
(a)
Tongzhou Shunyi
Daxing Huairou
Pinggu Miyun Yanqing
39.5 40.0 40.5 41.0
115.5 116.0 116.5 117.0 117.5
long
lat
2 4 6 8
level
Heat map by district: Darker color means greater concentration of crop damages.
(b)
39.5 40.0 40.5 41.0
115.5 116.0 116.5 117.0 117.5
long
lat
100 200 300 400
num
Dot - density map: Blue, violet and red dots represent low, medium and high frequency of events, respectively.
(c)
0 25 50 75
0 2 4 6
Counts(logarithmic transformation)
Frequency
Histogram of the number of crop losses caused by wildlife by village.
(d)
Figure 1: Visualization of the distribution and frequencies of crop damages by wild animals in each district of Beijing from 2009 to 2017.
Note that the three independent variables in level were logarithmic transformed. Figure 2 below shows the his- togram of the number of crop loss events caused by wild animals in each village before and after the log transformation.
Before the transformation, the distribution was skewed to the left; but afterwards, it improved significantly. As such, we used log-transformed data in fitting our model. The estimated coefficients are listed in Table 1, and the corre-
0 100 200
0 100 200 300 400 500 Counts(level)
Frequency
0 25 50 75
0 2 4 6
Counts(logarithmic transformation)
Frequency
Figure 2: The counts in each village before and after the logarithmic transformation.
sponding residual plots are given in Figure 3. From the coefficient estimates, it appears that the Poisson generalized linear regression model is suitable to characterizing the number of crop loss events caused by wild animals in each village. But plotting the residuals against the predicted values shows a cone shape, suggesting a poor fit. Likewise, the normal Q-Q plotting takes a parabolic shape, indicating a poor normality; and plotting the residuals versus leverage plots reveals that certain points have excessive leverages, further implying an inappropriate model selection.
Table 1: Estimated results of the Poisson generalized linear models Coefficient Estimate Std. Error zvalue Pr(>|z|) Significance
Interceptβ0 4.90 0.11 43.04 <2×10−16 ***
Woodlandβ1 −0.10 0.00 −28.34 <2×10−16 ***
River systemβ2 0.29 0.01 46.98 <2×10−16 ***
Ratio of protective measuresβ3 −0.29 0.02 −18.93 <2×10−16 ***
Population densityβ4 −0.53 0.02 −28.48 <2×10−16 ***
Significance: ‘.’ 10% level (p≤0.1) ‘*’ 5% level (p≤0.05) ‘**’ 1% level (p≤0.01) ‘***’ 0.1% level (p≤0.001)
3.2. A negative binomial generalized linear model
Given that the Poisson generalized linear regression model was not the most appropriate statistical model for fitting our data, we decided to try a negative binomial linear model as it is a generalized Poisson regression, thereby loosening the restrictive assumption that the variance is equal to the mean of the fitted Poisson model. Here, a negative binomial linear model means that the observed count follows a negative binomial distribution. Traditionally, a negative
Figure 3: Residual plots of the Poisson generalized linear model.
binomial regression is based on the Poisson-gamma mixture distribution, because it allows the modeling of Poisson heterogeneity using a gamma distribution.
Again, letYibe the number of events in villagei. Suppose that there is an observed variableEiwith a distribution of gamma(θ)/θ, where gamma(θ) denotes a gamma distribution with a shape parameterθand a rate parameter equal to 1. A negative binomial linear model can be specified as:
Yi∼Poisson(µiEi),
log(µi)=β0+β1Woodi+β2Riveri+β3Protecti+β4Popi, θEi∼Gamma(θ),
(3.1)
where (Yi,Ei),i=1,2,· · ·,nare independent variables,Woodi,Riveri,Protecti, andPopiare the same covariates as introduced earlier. The estimation coefficients forβ0, . . . , β4are given in Table 2, and the corresponding residual plots are provided in Figure 4.
Table 2: Estimated results of the negative binomial regression model
Coefficient Estimate Std. Error zvalue Pr(>|z|) Significance
Interceptβ0 6.16 1.01 6.09 <1×10−9 ***
Woodlandβ1 −0.13 0.04 −3.74 <2×10−4 ***
River systemβ2 0.24 0.05 4.65 3×10−6 ***
Population densityβ4 −0.65 0.17 −3.87 <1×10−4 ***
Ratio of protective measuresβ3 −0.53 0.15 −3.49 5×10−4 ***
θ 0.42 0.02 - - -
Significance: ‘.’ 10% level (p≤0.1) ‘*’ 5% level (p≤0.05) ‘**’ 1% level (p≤0.01) ‘***’ 0.1% level (p≤0.001)
Table 2 indicates that the p values of all explanatory variables are very small, suggesting significant coefficient estimates. The residual plots in Figure 4 further validate the choice of a negative binomial model. The AIC value
Figure 4: Residual plots of the negative binomial generalized regression model.
for this model is 5664.2, whereas the AIC value of the Poisson linear model fitting above is 49213. A model with a smaller AIC value should be selected. Together, these indicators lend our preference to the negative binomial regression model. For comparison, we also fit a negative binomial model in which the covariates in level were not logarithmic transformed. The results, reported in Table B-1 and Figure B-1 of Appendix B, indicate that the model does not fit the data well. This is reflected in the coefficient for population density, which is no longer significant, and the residual plots showing more dispersion.
3.3. A prediction of the compensation for crop losses
Furthermore, we took an additional step to explore the relationship between the number of events in each village and the villages total economic loss. The total loss has an enormous skewness itself, too; but the skewness decreases significantly after its logarithm transformation. So, we took the logarithms total loss,Zi, as the dependent variable in response to the number of events,Yi, in a village. The fitted linear model is reported in Table 3. It looks that the number of crop-loss events is a good predictor of the total loss.
Table 3: Estimated results of the compensation for crops losses Coefficient Estimate Std. Error zvalue Pr(>|z|) Significance
Intercept β0 5.57 0.08 72.35 <2×10−16 ***
log(Yi) β1 0.96 0.02 39.60 <2×10−16 ***
Significance: ‘.’ 10% level (p≤0.1) ‘*’ 5% level (p≤0.05) ‘**’ 1% level (p≤0.01) ‘***’ 0.1% level (p≤0.001)
It can be seen from Figure 5 that the regression residuals are randomly distributed on both sides of zero, with a good normality. This indicates that the model was well fitted.
As a result of our model in Table 3, the loss amount can be predicted for considering a compensation scheme for the villagers losses. Of course, the number of loss events should be first predicted with the model in Table 2. The mean
6 7 8 9 10 11
−40246
Fitted values
Residuals
Residuals vs Fitted
237 12 236
−3 −2 −1 0 1 2 3
−2024
Theoretical Quantiles
Standardized residuals
Normal Q−Q
237 12236
6 7 8 9 10 11
0.00.51.01.52.0
Fitted values
Standardized residuals
Scale−Location
237 12 236
0.000 0.002 0.004 0.006 0.008
−4−2024
Leverage
Standardized residuals
Residuals vs Leverage
1215 13
Figure 5: Residual plots of the compensation for crops losses
of the predicted loss events is 48.18 (the actual value is 47.17), and the mean of the predicted loss amounts’ logarithm is 8.07 (the real value is 9.12). These values are fairly close to their observed counterparts. To further illustrate our fitted models’ performance, we also predicted loss events and amounts of the six villages with the highest frequencies of crop losses, as shown in Table 4, and those of the six villages with the lowest frequencies of crop losses shown in Table 5.
Table 4: The top six villages with high-risk loss.
High−risk villages Popi Woodi Riveri Cost yi Predicted number of occurrences Predicted compensations
Mangniu 53.784 12.44 8144.31 19180.00 70 212.73 45724.12
Shuikui 135.81 12.39 12301.19 21883.75 113 128.56 28153.69
Five mountain 136.26 7.54 7958.72 7340.00 6 123.7283 27133.84
Da Guantou 135.29 12.75 10246.00 20652.00 153 122.97 26972.94
Wang Jiabao 135.32 13.31 10467.17 1905.00 21 122.87 26951.96
Pian Poyu 136.31 11.50 9279.78 45459.00 114 121.22 26603.72
Table 4 suggests that the predicted numbers of loss events and the losses tend to be larger than the actual number of loss events and the loss amount for villages with high frequencies of loss occurrence. Note that Mangniou of Qianjiadian Town, Yanqing District, has the lowest population density (median density of all villages is 206.19); the distance to the river system of Shuikui of Yongning Town, Yanqing District, is very large (the median distance is 3097.44, and the quantile of 95% is 10183.47); the distance to the woodland of Five Mountain of Yongning Town, Yanqing District, is very small (the median distance is 79.37, and the quantile of 5% is 10.02). According to the fitted model, these factors significantly increase the number of wildlife incidents and thus the loss amount.
On the other hand, the six villages listed in Table 5 have relatively low predicted occurrence times. The reason for White Zhuangzi to have the smallest predicted number of loss events among all the villages is the high value of population density (the median density is 206.19). Similarly, as East Huangliang is close to the water system (the median distance to the water system of all villages is 3097.44), its predicted number of loss events is 6.73, much lower than the actual number of loss events of 18. Accordingly, in these villages with the lowest frequencies of crop losses, the predicted amounts of losses tend to be lower than the actual amounts.
In short, our models perform better in predicting the frequencies and amounts of crop damage for those villages with characteristics closer to the average, rather than the extremes.
Table 5: The last six villages with low-risk loss.
Low−risk villages Popi Woodi Riveri Cost yi Predicted number of occurrences Predicted compensations
White Zhuangzi 8268.03 1562.40 276.42 280.00 3 1.46 378.09
Du Xinzhuang 1123.73 1654.10 934.88 87.52 4 5.41 1333.09
South Nianfeng 965.43 4401.11 1331.92 350 1 5.70 1400.32
Old Shiyu 827.19 1729.55 906.61 720.00 1 6.52 1594.10
East Huangliang 269.10 219.60 14.66 3532.80 18 6.73 1645.17
Big shuiyu 271.47 726.70 45.10 2362.5 1 7.43 1808.95
4. Discussion and Conclusions
In this paper, we have modeled the frequencies and amounts of crop damages caused by a growing wildlife popu- lation, using a large sample from the rural villages of Beijing for the period of 2009-2017. We first characterized the spatial distribution of crop losses and the temporal trends through descriptive statistics and visual representations, and it was quickly realized that the locations of crop loss events were concentrated in the east and south of Yanqing, the northwest of the Miyun district, the southeast of the Huairou district, the southwest of the Mentougou district, and the northeast of the Fangshan district. Then, we considered both a Poisson linear model and a negative binomial model to represent our data. Ultimately, our regression efforts led us to conclude that the negative binomial generalized linear model gives a better fit. Notably, while there have been efforts of characterizing and reducing wildlife damage to properties, including crops, in other countries (e.g., Karanth et al., 2013; Goodrich, 2010), our work is one of the few that have been done in China.
It is found that the number of crop damage caused by wildlife are likely small if the village is far away from the woodland, close to the water system, has a high population density, or has adopted protective measures. These and other findings can allow us to make practical recommendations for mitigating and even preventing crop losses by wildlife. To reduce the frequency of crop damages in Beijing (or elsewhere), we suggest improving the villages protection measures. The cropland of these villages can be protected through fencing (net, rope, etc.), hanging (cloth, flag, etc.) or covering (plastic film) (Chen et al., 2006). Those villages either near the woodland or far from the water system should be the priority area of protection. Further, the cropland that suffer more frequent and severe damage may be converted to wildlife habitat, ecological corridors, or important wildlife food sources through long-term leases and purchase agreements. On the other hand, farming on lands with more intense human activities and less frequent wildlife activities can be continued.
Similarly, our work can provide the necessary information for formulating a compensation policy in the future.
As an example, in areas with high loss incidents, it is more critical to consider compensating the losses, and the compensation should be made with a clear understanding of frequency of damages. Our work also provides insight for transforming compensation from post- to pre-event to promote the development of the insurance business concerning the crop losses and to incentivize rural residents to protect crops from wildlife damage by using practical preventive methods. Certainly, any compensation or insurance policy must be predicated on knowing such variables as the distance from the village to the woodland or the water system and the proportion of protective measures.
Future work should consider introducing spatial autocorrelation into the model, in which the number of loss events in a village can be correlated with the number of events in neighboring villages. It is also interesting to consider introducing the relationship between the number of events and the time of occurrence into the modeling process, so as to establish a spatiotemporal autocorrelation model. Moreover, if possible, we would analyze the factors that influence how different types of wildlife cause variable categories and amounts of crop damages (He and Wu, 2010). Finally, including more independent variables, such as the location, size, and time of a damage event, can definitely improve a model’s predictive power. To obtain more and better data related to crop losses caused by wildlife, though, requires
necessary financial and personnel support.
In closing, while predicting the frequency and severity of crop damage remain challenging, we have made sig- nificant progress along that direction. We are confident that as additional data with more relevant variables are accu- mulated, we will be able to make more accurate predictions, which should go a long way in protecting wildlife and resolving its conflict with humans, thereby realizing a harmonious coexistence of humans and wildlife in China and beyond.
Acknowledgments
The authors are grateful for the experts’ and scholars’ constructive comments and suggestions during the first international forum on forest policy and economics held at Beijing Forestry University. This work was supported in part by the Fundamental Research Funds for the Central Universities under Grant BLX2019446, the National Natural Science Foundation of China under Grant 71573018, and the Beijing Forestry University Education and Teaching Research Project under Grant BJFU2020JY023. Our deepest gratitude goes to the editors and the journal reviewers for their insightful comments and suggestions that substantially improved this paper.
Appendix A: Maps and graphs of the frequencies of crop damages
The distribution maps of the annual crop loss events are shown in Figure A-1. We selected the years from 2011 to 2016 because the dataset’s sample size in those years is relatively large. The pattern of the crop loss frequencies had apparent changes over the years. For instance, the Miyun district had the largest frequencies in 2015 and 2016, but the frequencies of the Yanqing district were the largest for the other years.
We also used the line chart to show the yearly change in crop damage frequencies in different districts in Figure A- 2. It shows that the crop loss events in the Yanqing district increased in 2012, decreased till 2015, and then increased again in 2016. The events in Miyun and Huairou showed an upward trend. For Mentougou, the events fluctuated before 2014 and then declined. The events in Fangshan slightly increased till 2013, decreased in 2014, peaked in 2015, and finally decreased in 2016. The events in Pinggu remained stable throughout the period. Heat map for the frequencies of crop loss events is shown in Figures A-3, generated with the annual observations. It confirms that the location range of crop damages by wild animals remains roughly the same, but the frequencies had an apparent variation over the years. The number of occurrences went up in the northern region for the whole period, but they decreased in Mentougou from 2014 to 2016.
Chaoyang Fengtai
Haidian Mentougou
Fangshan Tongzhou
Shunyi Changping
Daxing Huairou
Pinggu Miyun Yanqing
0 400 800 1200 1600 freq
2011
Chaoyang Fengtai
Haidian Mentougou
Fangshan Tongzhou
Shunyi Changping
Daxing Huairou
Pinggu Miyun Yanqing
0 500 1000 1500 2000 freq
2012
Chaoyang Fengtai Haidian Mentougou
Fangshan Tongzhou
Shunyi Changping
Daxing Huairou
Pinggu Miyun Yanqing
0 500 1000 1500 2000 freq
2013
Chaoyang Fengtai
Haidian Mentougou
Fangshan Tongzhou
Shunyi Changping
Daxing Huairou
Pinggu Miyun Yanqing
0 500 1000 1500 freq
2014
Chaoyang Fengtai
Haidian Mentougou
Fangshan Tongzhou
Shunyi Changping
Daxing Huairou
Pinggu Miyun Yanqing
0 400 800 1200 freq
2015
Chaoyang Fengtai Haidian Mentougou
Fangshan Tongzhou
Shunyi Changping
Daxing Huairou
Pinggu Miyun Yanqing
0 500 1000 1500 freq
2016
Figure A-1: Distribution maps of the frequencies of crop damages by wild animals within one year (from 2011 to 2016) for each district. The corresponding year for the damage is located on the top in each plot.
Figure A-2: Line chart of the frequencies of crop damages by wild animals in each district from 2011 to 2016.
Figure A-3: Heat maps of the annual frequencies of crop damages by wild animals, from 2011 to 2016. The corre- sponding year for the damage is located on the top in each plot.
Appendix B: Estimated results of the negative binomial model in which the covariates do no logarithmic trans- formation
Here is a summary of the estimated negative binomial generalized linear model in which the covariates do no logarithmic transformation. The results are shown in Table B-1, and the residuals are plotted in Figure B-1.
Table B-1: Estimated results of the negative binomial generalized linear model (no logarithmic transformation) Coefficient Estimate Std. Error zvalue Pr(>|z|) Significance
Interceptβ0 3.6720 0.1741 21.092 <2×10−16 ***
Distance from woodlandβ1 −0.0004 0.0001 −3.332 9×10−4 ***
Distance from river systemβ2 0.0001 0.0000 5.940 3×10−9 ***
Ratio of protective measuresβ3 −0.6244 0.1538 −4.061 5×10−5 ***
Population densityβ4 −0.0002 0.0005 −0.465 0.6418
θ 0.4134 0.0217 - - -
Significance: ‘.’ 10% level (p≤0.1) ‘*’ 5% level (p≤0.05) ‘**’ 1% level (p≤0.01) ‘***’ 0.1% level (p≤0.001)
Figure B-1: Residual plots of the negative binomial generalized linear model.
The coefficient for population density becomes insignificant, and the multi-faceted residual plots do not perform well. The Wald test indicates that we cannot reject the null hypothesis that the population density coefficient is equal to zero. Preliminary analysis demonstrated the challenge that using the whole dataset in the above model would not work well because of the population-density variable’s strong influence.
References
Balmford, A., Moore, J.L., Brooks, T., Burgess, N., Hansen, L.A., Williams, P., Rahbek, C., 2001. Conservation conflicts across Africa. Science 291, 2616–2619.
Chen, L., Yang, S., Shen, Y., 2006. Harm of wild animals to crops in xishuangbanna and preventive measures. Forest Inventory and Planning 31, 33–35.
Chen, S., Yi, Z.F., Campos-Arceiz, A., Chen, M.Y., Webb, E.L., 2013. Developing a spatially - explicit, sustainable and risk - based insurance scheme to mitigate human - wildlife conflict. Biological Conservation 168, 31–39.
Chen, Y., Marino, J., Chen, Y., Tao, Q., Sullivan, C.D., Shi, K., Macdonald, D.W., 2016. Predicting hotspots of human-elephant conflict to inform mitigation strategies in Xishuangbanna, Southwest China. PLOS ONE 11, e0162035. doi:10.1371/journal.pone.0162035.
Dickman, A.J., Macdonald, E.A., Macdonald, D.W., 2011. A review of financial instruments to pay for predator conservation and encourage human - carnivore coexistence. Proceedings of the National Academy of Sciences of the United States of America 108, 19836–19836.
Goodrich, J.M., 2010. Human - tiger conflict: A review and call for comprehensive plans. Integrative Zoology 5, 300–312.
He, X., Wu, Z., 2010. The present situation and management of wildlife accidents in china. Sichuan Journal of Zoology 29(1), 141–143.
Herrero, J., Garc´ıa-Serrano, A., Couto, S., Ortu˜no, V.M., Garc´ıa-Gonz´alez, R., 2006. Diet of wild boarSus scrofa L.and crop damage in an intensive agroecosystem. European Journal of Wildlife Research 52, 245–250.
Honda, T., 2007. Environmental factors affecting damage by wild boars (Sus scrofa) to rice fields in Yamanashi Prefecture, central Japan. Mammal Study 32, 173–176.
Huang, C., Li, X.Y., Shi, L.J., Jiang, X.L., 2018. Patterns of human-wildlife conflict and compensation practices around Daxueshan Nature Reserve, China. Zoological Research 39(6), 406 – 412.
Jing, C., Jiang, Z., Zeng, Y., Li, C., Bravery, B.D., 2008. Factors affecting crop damage by wild boar and methods of mitigation in a giant panda reserve. European Journal of Wildlife Research 54, 723–728.
Li, X., Buzzard, P., Chen, Y., Jiang, X., 2013. Patterns of livestock predation by carnivores: Human - wildlife conflict in northwest Yunnan, China.
Environmental Management 52, 1334–1340.
Lu, J., 2009. Analysis of the effect of returning farmland to the forest in beijing and strategies for consolidating the achievements. Forest Resources Management 5, 44–49.
Marchal, V., Hill, C., 2009. Primate crop-raiding: A study of local perceptions in four villages in North sumatra, Indonesia. Primate Conservation 24, 107–116.
Mishra, C., Allen, P., Mccarthy, T., Madhusudan, M., Bayarjargal, A., Prins, H., 2003. The role of incentive programs in conserving the snow leopard. Conservation Biology 17, 15121520. doi:10.1111/j.1523-1739.2003.00092.x.
Naughton-Treves, L., 1998. Predicting patterns of crop damage by wildlife around Kibale National Park, Uganda. Conservation Biology 12( 1), 156–168.
Nyhus, P.J., 2016. Human - wildlife conflict and coexistence. Annual Review of Environment and Resource 41, 143–71.
Osborne, P., Alonso, J., Bryant, R., 2001. Modelling landscape-scale habitat use using GIS and remote sensing: a case study with great bustards.
Journal of Applied Ecology 38, 458–471. doi:10.1046/j.1365-2664.2001.00604.x.
Pettigrew, M., Xie, Y., Kang, A., Rao, M., Goodrich, J., Liu, T., Berger, J., 2012. Human - carnivore conflict in China: A review of current approaches with recommendations for improved management. Integrative Zoology 7, 210–226.
Potter, R.B., 1977. Locational analysis in human geography: Vol. I: Locational models. Vol. II: Locational methods. Geography 63, 245.
Rao, K.S., R.K.Maikhuri, S.Nautiyal, K.G.Saxena, 2002. Crop damage and livestock depredation by wildlife: A case study from Nanda Devi Biosphere Reserve, India. Journal of Environmental Management 66, 317–327.
Sekhar, N.U., 1998. Crop and livestock depredation caused by wild animals in protected areas: The case of Sariska Tiger Reserve, Rajasthan, India.
Environmental Conservation 25, 160–171.
Shen, J., Cui, G., 2015. Current situation and prevention and control measures of wildlife accidents at home and abroad. World Forestry Research 28(1), 43–49.
Sitati, N.W., Walpole, M.J., Leader-Williams, N., 2005. Factors affecting susceptibility of farms to crop raiding by African elephants: Using a predictive model to mitigate conflict. Journal of Applied Ecology 42, 1175–1182.
Sokal, R.R., Thomson, J.D., 1987. Applications of spatial autocorrelation in ecology. Develoments in Numerical Ecology 14, 431–466.
Treves, A., Karanth, K.U., 2003. Human - carnivore conflict and perspectives on carnivore management worldwide. Conservation Biology 17, 1491–1499.
Wagner, K.K., Conover, M.R., Schmidt, R.H., 1997. Compensation programs for wildlife damage in north America. Wildlife Society Bulletin (1973-2006) 25, 312–319.
Xu, X., 2015. Remote sensing monitoring data of land use status in china in 2015.http://www.resdc.cn/data.aspx?DATAID=184.
Yang, W., Shuguang, H.E., Shen, Y., 2009. Spatio - temporal patterns of wildlife damages in southern part of Mt.Baima - Xueshan. Journal of Mountain Science 3, 52–56.
Zhou, H., Tang, J., Guo, B., 2010. China’s key wildlife protection animals involved in the relevant legal and economic compensation issues. Journal of Beijing Forestry University 9(1), 36–40.
